University of Ghana http://ugspace.ug.edu.gh UNIVERSITY OF GHANA COLLEGE OF HEALTH SCIENCES SCHOOL OF PUBLIC HEALTH SMALL AREA ANALYSIS OF POPULATION HEALTH: AN EXAMINATION OF THE VALUE OF FACILITY-BASED DATA IN GHANA BY SETH KWAKU AFAGBEDZI (10291421) A THESIS SUBMITTED TO THE SCHOOL OF GRADUATE STUDIES IN PARTIAL FULFILLMENT OF THE AWARD OF DOCTOR OF PHILOSOPHY IN PUBLIC HEALTH DECEMBER 2017 University of Ghana http://ugspace.ug.edu.gh DECLARATION I hereby declare that except for the references cited to other people’s work which has been duly acknowledged, this thesis is the result of my original independent research done under supervision and has neither in part nor in whole been presented elsewhere for another degree. PhD Candidate Date June 10, 2018 Seth Kwaku Afagbedzi Academic Supervisors: Date June 11, 2018 Doctor Alex Barimah Owusu June 11, 2018 Doctor James Wright June 11, 2018 Professor Allan G. Hill June 11, 2018 Doctor Alfred E. Yawson i University of Ghana http://ugspace.ug.edu.gh DEDICATION I dedicate this work to my dear wife Phidelia and my children Keli, Elorm, Kafui and Elikem whose understanding and patience provided the magic bullet. ii University of Ghana http://ugspace.ug.edu.gh ACKNOWLEDGEMENT My sincere thanks go to my primary supervisor Dr. Alex Barimah Owusu, and Dr. Alfred Yawson – co-supervisor, both of University of Ghana, for their technical support and useful suggestions which helped to improve this piece of work. I am very grateful to Professor Allan Hill of University of Southampton, a member of my supervision team, who facilitated my one semester study trip to University of Southampton and helped chose my topic. I am also highly indebted to Dr. James Wright of University of Southampton, a member of my supervision team who began this work with me and provided enoumous technical support and advice at every stage of this work. I also thank University of Southampton - Faculty of Social, Human and Mathematical Sciences for sponsoring my 6 months study trip with them. My sincere appreciation goes to Professor Richard Adanu, the Dean of School of Public Health, for his immense financial support and for his roles for my study trip in University of Southampton. My warmest appreciation to Dr Ernest Kenu for reading through my first draft and making very significant contributions. I thank you Doctor Mawuli Dzordzomenyo for your encouragement and your interventions when things got tough. To Prof. Moses Aikins, Prof. Edwin Afari, Dr.Samuel Sackey, Dr Richard Kofie, Dr Patricia Akwengo and Dr. Francis Anto, I say thank you for your encouragement and your eagerness to see me finish this project. My warmest appreciation goes to Mr Kofi Agyabeng for helping me with data management and analysis. Thanks to Mr Peter Ofori-Atta ( RS/GIS Lab Geography Dept) and George Owusu ( CERGIS) for giving me GIS support. To Team Biostatistics – Samuel Dery, Chris Guure, Duah Dwomoh, Baaba da-costa Vroom, Samuel Bosomprah, Tony Godi and Edith Teteh I say Ayeeko, I indeed appreciated your never flinching support to me throughout the course of my PhD work. To all my friends Seth, Peace, Mac, Busy and Favour I say thank you for your support in diverse ways. iii University of Ghana http://ugspace.ug.edu.gh Table of Contents DECLARATION....................................................................................................................... i DEDICATION.......................................................................................................................... ii ACKNOWLEDGEMENT ..................................................................................................... iii LIST OF TABLES .................................................................................................................. ix LIST OF FIGURES AND MAPS .......................................................................................... xi LIST OF ABBREVIATIONS ............................................................................................. xiii ABSTRACT ........................................................................................................................... xvi CHAPTER ONE ...................................................................................................................... 1 INTRODUCTION.................................................................................................................... 1 1.1 Background ................................................................................................................. 1 1.2 Problem Statement ...................................................................................................... 7 1.3 Justification ............................................................................................................... 10 1.4 Objectives .................................................................................................................. 11 1.4.1. General Objective ................................................................................................... 11 1.4.2. Specific Objectives ................................................................................................. 11 1.5 Conceptual Framework ............................................................................................. 12 1.6. Justification for the choice of malaria and diarrhoea for this study .............................. 15 1.7. Structure of the Thesis............................................................................................... 16 CHAPTER TWO ................................................................................................................... 19 LITERATURE REVIEW ..................................................................................................... 19 2.1. Institutional arrangement of the Ghana’s Health Sector ............................................... 19 2.2. Disease Burden in Ghana .............................................................................................. 21 2.2.1 Malaria ..................................................................................................................... 23 2.2.2 Diarrhoea ................................................................................................................. 25 2.3. Small Area Statistics ................................................................................................. 27 2.3.1. Definitions and delineation of small areas ......................................................... 27 2.3.2. Small area analysis in health .............................................................................. 28 2.4. Sources of small area data ......................................................................................... 29 2.4.1. Population and Housing Censuses ..................................................................... 29 2.4.2. Routine administrative records .......................................................................... 30 2.4.3. Sample surveys .................................................................................................. 31 2.4.4. Routine health information system data ............................................................ 31 2.4.5. Ghana’s routine health information systems...................................................... 33 2.4.6. Quality of DHIMS2 data .................................................................................... 34 2.4.7. Geographic cordinates of health facilities.......................................................... 35 iv University of Ghana http://ugspace.ug.edu.gh 2.5. GIS and Health .......................................................................................................... 36 2.6. Comparability between DHIMS2 and GDHS data ................................................... 39 2.7. Spatial relationship between health facility visits & environmental/ household risk factors ................................................................................................................................... 40 2.8. Spatial heterogeneities of relationship between health facility visits & environmental / household risk factors ........................................................................................................ 42 2.9. Approaches to small area analysis ............................................................................ 43 2.9.1. Statistical approaches to small area analysis ..................................................... 44 2.9.2. Geospatial approaches to small area analysis .................................................... 48 2.9.2.1. Disease Mapping ........................................................................................ 48 2.9.2.2. Diseases Clustering..................................................................................... 50 2.10. Areal data analysis ................................................................................................. 51 2.10.1. Spatial Autocorrelation and Moran’s I ........................................................... 52 2.10.2. Spatial Data Modelling................................................................................... 53 2.11. Health facility catchment area ............................................................................... 55 2.12. Population denominators for health facility catchment area ................................. 57 CHAPTER THREE ............................................................................................................... 59 METHODS ............................................................................................................................. 59 3.1. Study Area ................................................................................................................. 59 3.1.1. Profiles of Districts in GAR ............................................................................... 60 3.2. Study design .............................................................................................................. 72 3.3. Study population ....................................................................................................... 73 3.4. Inclusion and exclusion Criteria ................................................................................ 73 3.5. Operational definition of main study outcome .......................................................... 73 3.6. Data types and sources .............................................................................................. 74 3.6.1. Routine health facility-based data ...................................................................... 74 3.6.2. Population and housing census data .................................................................. 75 3.6.3. Household Survey data ...................................................................................... 76 3.6.4. Meteorological Data........................................................................................... 77 3.6.5. Administrative boundary Data ........................................................................... 78 3.7. Data Processing ......................................................................................................... 78 3.7.1. Health facility visits data ................................................................................... 78 3.7.2. Population census data ....................................................................................... 80 3.7.3. Ghana Demographic and Health Survey Data ................................................... 81 3.7.4. Linking of HFV to district -level population census data. ................................. 82 3.7.5. Creating Thiessen polygon catchment areas ...................................................... 82 v University of Ghana http://ugspace.ug.edu.gh 3.7.6. Linking of population census data to health facility catchment areas. .............. 83 3.7.7. Data flow diagram.............................................................................................. 84 3.8. Data Analysis ............................................................................................................ 86 3.8.1. Determining the level of completeness in reporting into DHIMS2 ................... 86 3.8.2. Comparison of facility-based DHIMS2 and GDHS data ................................... 88 3.8.3. Measuring health differentials using routine health facility-based data ............ 91 3.8.4. Visualising spatial variations in routine facility-based data (HFV due to malaria and diarrhoea .................................................................................................................... 92 3.8.4.1. Graduated symbol map ............................................................................... 93 3.8.4.2. Choropleth map .......................................................................................... 93 3.8.4.3. Hot Spot Analysis (Getis-Ord Gi*) ............................................................ 94 3.8.5. Analysing spatial relationship between HFVR for malaria or diarrhoea and household factors .............................................................................................................. 95 3.8.5.1. Study variables ........................................................................................... 96 3.8.5.2. Checking Model Assumptions .................................................................. 100 3.8.5.3. Fitting of Models ...................................................................................... 103 3.8.6. Visualising GWR output for spatial heterogeneities ...................................... 104 3.9. Ethical considerations ............................................................................................. 105 CHAPTER FOUR ................................................................................................................ 106 RESULTS AND FINDINGS ............................................................................................... 106 4.1. Health facility distribution in GAR ............................................................................. 106 4.2. Determining the level of completeness in the coverage of DHIMS2 dataset ............. 108 4.3. Comparing DHIMS2 data to GDHS data .................................................................... 113 4.3.1. Expected number of children <5 years who were treated for malaria and diarrhoea in medical facilities (public and private) during the two weeks preceding 2014 GDHS113 4.3.2. Comparing DHIMS2 and GDHS data .................................................................. 118 4.4. Measuring variations in HFV due to malaria and diarrhoea among population of GAR ............................................................................................................................................ 123 4.4.1. Overview of HFVR due to malaria and diarrhoea in GAR in 2014 ..................... 123 4.4.2. Age by sex distribution of HFV due to malaria and diarrhoea for GAR in 2014 123 4.4.3. Percentage distribution of HFV due to malaria and diarrhoea at the district levels in GAR, 2014 ................................................................................................................. 125 4.5. A-spatial analysis of HFV and HFVR for malaria and diarrhoea and environmental factors ................................................................................................................................. 129 4.5.1. Relationships between climatic conditions and HFV due to malaria and diarrhoea ........................................................................................................................................ 129 4.5.2. Relationship between HFV due to malaria and diarrhoea and population density by districts............................................................................................................................ 132 vi University of Ghana http://ugspace.ug.edu.gh 4.6. Spatial distribution of HFVR for malaria and diarrhoea by districts .......................... 134 4.7. Spatial distribution of HFVR for malaria and diarrhoea by catchments ..................... 135 4.8. Cluster detection using spatial autocorrelation (Global Moran’s I) tool ..................... 137 4.9. Getis-Ord Gi* (hotspot) analysis of HFVR for malaria and diarrhoea ....................... 138 4.10. Distribution of HFVR for malaria and diarrhoea by health facility types ................ 140 4.11. Model variables and relationships ............................................................................. 142 4.11.1. Summary Statistics of malaria and diarrhoea risk factors ................................. 142 4.11.2. Bivariate analysis of the dependent variables and the corresponding risk factors ........................................................................................................................................ 143 4.11.3. Checking assumptions underlying OLS model for HFVR for malaria and diarrhoea ......................................................................................................................... 149 4.12. Ordinary Least Square (OLS) Regression Analysis .................................................. 151 4.12.1. HFVR for malaria OLS regression models ........................................................ 151 4.12.2. Comparing HFVR for malaria primary/overall and secondary models ............. 155 4.12.3. HFVR for diarrhoea OLS regression models ..................................................... 158 4.12.4. Comparing HFVR for diarrhoea primary and secondary models ....................... 162 4.13. Testing for clustering of residuals using spatial autocorrelation (Global Moran’ I) . 165 4.14. Geographically Weighted Regression (GWR) Model .............................................. 167 4.14.1. Spatial distribution of dependent (HFVR for malaria) and independent variables (Pwalpma and Plowweaq) .............................................................................................. 167 4.14.2. GWR model for HFVR for malaria .................................................................... 169 4.14.3. Visualising the GWR output............................................................................... 170 4.15. GWR model result and spatial variation in HFVR for malaria ................................. 172 4.16. Key findings .............................................................................................................. 177 CHAPTER FIVE ................................................................................................................. 181 DISCUSSION ....................................................................................................................... 181 5.1. What can DHIMS2 tell us about outpatient attendance? ........................................ 181 5.1.1. Examining completeness of DHIMS2 data...................................................... 182 5.1.2. Determining the external consistency of facility-based (DHIMS2) dataset .... 186 5.2. Do DHIMS2-derived pseudo-incidence rates show plausible patterns? ................. 191 5.2.1. Climate variables and heath facility visits due to malaria and diarrhoea ........ 191 5.2.2. Using routine facility-based data to measure health differentials .................... 193 5.2.3. Plausibility of association between HFVR for malaria and diarrhoea and environmental /household risk factors ............................................................................ 198 5.2.4. Spatial heterogeneities in the relationship between HFVR for malaria and household risk factors ..................................................................................................... 201 vii University of Ghana http://ugspace.ug.edu.gh 5.3. Data system and methodological challenges associated with analysis of outpatient data ................................................................................................................................. 205 5.3.1. Overview .......................................................................................................... 205 5.3.2. Data system challenges .................................................................................... 206 5.3.2.1. Completeness of facility reporting challenges.......................................... 206 5.3.2.2. Completeness of indicator data (zero / missing value) ............................ 206 5.3.2.3. Missing and ambiguous geographic coordinates of health facilities ........ 207 5.3.2.4. Effect of attractiveness among facility types ............................................ 207 5.3.2.5. Challenges with the use of census data in outpatient data analysis .......... 208 5.3.2.6. Challenges with small area boundaries .................................................... 209 5.3.3. Methodological Challenges ............................................................................. 211 5.4. Limitations of the study........................................................................................... 213 5.5. Review of study objectives .................................................................................... 215 CHAPTER SIX .................................................................................................................... 220 SUMMARY, CONCLUSION AND RECOMMENDATION .......................................... 220 6.1. SUMMARY OF KEY FINDINGS ............................................................................. 220 6.2. CONCLUSION ........................................................................................................... 221 6.3. RECOMMENDATIONS ........................................................................................... 223 6.3.1. For policy .............................................................................................................. 223 6.3.2. For future research ................................................................................................ 224 6.4. CONTRIBUTION TO KNOWLEDGE ...................................................................... 224 APPENDICES ...................................................................................................................... 225 REFERENCES ..................................................................................................................... 229 viii University of Ghana http://ugspace.ug.edu.gh LIST OF TABLES Table 3.1: Independent variables of HFVR for malaria and diarrhoea.................................... 97 Table 3.2: Malaria and diarrhoea predictors and household factors ........................................ 98 Table 4.1. Distribution of health facilities in GAR by districts, NHIS status, ownership and type, 2014 ............................................................................................................................... 106 Table 4.2: Facility reporting rate within districts in GAR ..................................................... 108 Table 4.3. Sample distribution and treatment status of children <5year old with fever during 2014 GDHS in GAR .............................................................................................................. 114 Table 4.4. Determination of expected number of children <5year old from the population who received anti-malaria treatment at a medical facility(public and private) for the two weeks preceding the GDHS .............................................................................................................. 115 Table 4.5. Sample distribution and treatment status of children <5year old who visited medical facility due to diarrhoea during two weeks preceding 2014 GDHS in GAR ........... 116 Table 4.6. Determination of expected number of children from the population who received diarrhoea treatment at a medical facility(public or private) for the two weeks preceding the GDHS ..................................................................................................................................... 117 Table 4.7: Ratio of DHIMS2 to GDHS HFV due to malaria of <5 years old in GAR for a two-week average between September to December in 2014 ............................................... 119 Table 4.8: Ratio of DHIMS2 to GDHS HFV due to diarrhoea of <5 year old reported for GAR for an average of two weeks between September to December in 2014 ...................... 121 Table 4.9: Summary of HFVR hotspots for malaria and diarrhoea in GAR, 2014 .............. 140 Table 4.10: Summary statistics of study variables ................................................................. 143 Table 4.11: Correlation Coefficients among study variables of HFVR for malaria. ............. 147 Table 4.12: Correlation Coefficients among study variables for HFVR for diarrhoea. ........ 148 Table 4.13: Variance inflation factor (VIF) values for malaria and diarrhoea predictors. .... 150 Table 4.14: OLS result for the overall HFVR for malaria ( observed + imputed rates) model ................................................................................................................................................ 152 Table 4.15: OLS model result for observed HFVR for malaria ............................................ 154 Table 4.16: OLS model result for imputed HFVR for malaria. ............................................ 155 Table 4.17: Comparison OLS diagnostics of HFVR for malaria primary and secondary models .................................................................................................................................... 156 Table 4.18: OLS model result for primary HFVR for diarrhoea ( observed + imputed visit rates) ....................................................................................................................................... 159 Table 4.19: OLS model result for observed HFVR for diarrhoea. ........................................ 161 Table 4.20: OLS model result for imputed HFVR for diarrhoea. .......................................... 162 Table 4.21: Comparison of OLS diagnostics of HFVR for diarrhoea primary and secondary models. ................................................................................................................................... 163 Table 4.22: Comparison between OLS and GWR model for HFVR for malaria. ................. 170 ix University of Ghana http://ugspace.ug.edu.gh Table 5.1: Data system and methodological challenges and actions taken ........................... 205 x University of Ghana http://ugspace.ug.edu.gh LIST OF FIGURES AND MAPS Figure 2.1: Structure of the health sector of Ghana ................................................................. 19 Figure 2.2. John Snow’s cholera outbreak map of London ..................................................... 37 Figure 2.3. Choropleth map showing measles incidence per district – wave 8 ....................... 50 Figure 2.4. Creating Thiessen polygon ................................................................................... 56 Figure 3.1. Map of study area – GAR ...................................................................................... 59 Figure 4.1. Geographical distribution of health facilities in GAR in 2014............................ 107 Figure 4.2. Percentage comparison of observed and imputed HFV due to (a) malaria and (b) diarrhoea by districts in GAR in 2014 ................................................................................... 110 Figure 4.3. Distribution of observed and imputed HFV due to (a) malaria and (b)diarrhoea by facility ownership, in GAR,2014 ........................................................................................... 111 Figure 4.4. Distribution of observed and imputed HFV due to (a) malaria and (b) diarrhoea by facility type in GAR, 2014 ..................................................................................................... 113 Figure 4.5: Column chart showing age by sex comparison of DHIMS2 and GDHS data of <5 year old HFV due to malaria in GAR for two-week average between September to December in 2014 ................................................................................................................................... 120 Figure 4.6: Column chart showing age by sex comparison between DHIMS2 and GDHS HFV due to diarrhoea for <5 years old reported for GAR for an average of two weeks between September to December in 2014 ............................................................................................ 122 Figure 4.7: Bar chart showing distribution of HFV due to (a) malaria) and (b) diarrhoea reported for GAR in 2014 expressed as a proportion of 2014 population projected from 2010 census ..................................................................................................................................... 124 Figure 4.8: Bar chart showing percentage distribution of annual cumulative HFV due to (a) malaria and (b) diarrhoea by sex and districts in GAR in 2014. ............................................ 126 Figure 4.9: Bar chart showing distribution of Age-adjusted annual cumulative HFVR for (a) malaria and (b) diarrhoea by districts in GAR in 2014. ......................................................... 128 Figure 4.10: Mean monthly rainfall and HFV due to (a) malaria and (b) diarrhoea reported for GAR in 2014 .......................................................................................................................... 129 Figure 4.11: Monthly distribution of maximum and minimum temperatures and HFV due to (a) malaria and (b) diarrhoea reported for GAR in 2014 ....................................................... 131 Figure 4.12: Graduated symbol maps of HFVR of malaria and diarrhoea superimposed on population density maps by districts in GAR in 2014. .......................................................... 133 Figure 4.13: Choropleth map showing distribution of age-adjusted annual cumulative HFVR for (a) malaria and diarrhoea per 1,000 population by GAR districts in 2014 ...................... 134 Figure 4.14: Choropleth map showing distribution of HFVR for (a) malaria and (b) diarrhoea by catchment areas in GAR, 2014. ........................................................................................ 136 Figure 4.15: Moran’s, I spatial autocorrelation test results for the distribution of HFVR for (a) malaria and (b) diarrhoea ....................................................................................................... 137 xi University of Ghana http://ugspace.ug.edu.gh Figure 4.16: Getis-Ord Gi*’s maps showing hotspots of HFVR for(a) malaria and (b) diarrhoea in GAR, 2014 ......................................................................................................... 139 Figure 4.17: Side-by-side boxplot comparing the distribution of HFVR for (a) malaria and (b) diarrhoea and facility types in GAR in 2014 ......................................................................... 141 Figure 4.18: Scatter plot vs fitted graph showing relationship between annual cumulative HFVR for malaria and environment /household factors in GAR, 2014. ............................... 144 Figure 4.19: Scatter plot vs. fitted graph showing relationship between HFVR for diarrhoea and environmental/ household risk factors in GAR, 2014 .................................................... 145 Figure 4.20: Histograms showing square root transformed HFVR for (a) malaria and (b) diarrhoea residuals ................................................................................................................. 149 superimposed with normal density functions ........................................................................ 149 Figure 4.21: Comparison of OLS StdResid maps of HFVR for malaria:(a) overall (b) observed and (c) imputed ....................................................................................................... 157 Figure 4.22: Comparison of OLS Residual maps for HFVR for diarrhoea primary and secondary models ................................................................................................................... 164 Figure 4.23: Spatial autocorrelation on OLS regression residuals of HFVR for malaria and diarrhoea. ............................................................................................................................... 166 Figure 4.24: Spatial distribution of (a) HFVR for malaria; (b) Proportion of persons living in walls prone to malaria (Pwalpma); and (c) Proportion of population within low wealth quintile level (Plowweaq). ..................................................................................................... 168 Figure 4.25: GWR standard residual map for HFVR for malaria .......................................... 170 Figure 4.26: Spatial autocorrelation result for HFVR for malaria GWR model residual ...... 171 Figure 4.27: Spatial mapping of (a) locally weighted coefficient of determination (R2) and (b) condition number by GWR modelling. .................................................................................. 172 Figure 4.28: Spatial mapping of pseudo-t values of (a) intercept (b) wall prone to malaria (Pwalpma) and (c) Low wealth quintile (Plowweaq) for each catchment by GWR modelling ................................................................................................................................................ 174 xii University of Ghana http://ugspace.ug.edu.gh LIST OF ABBREVIATIONS AMA – Accra Metropolitan Assembly BCS – Behaviour Change Support CHPS – Community-Based Health Planning and Services DHIMS – District Health Information Management System ESRI - Environmental Systems Research Institute FHD – Family Health Division FRHP – Ghana Focus Region Health Project GAR – Greater Accra Region GDHS – Ghana Demographic and Health Survey GHS – Ghana Health Service GIS – Geographic Information Systems GSMF – Ghana Social Marketing Foundation GWR – Geographically Weighted Regression HC – Health Centre HEFRA – Health Facilities Regulatory Agency HF – Health facility HFV – Health facility visits HFVR – Health facility visits rate xiii University of Ghana http://ugspace.ug.edu.gh IRS – Indoor Residual Spray LANKMA – La-Nkwantanang-Madina Municipal Assembly LEKMA – Ledzokuku-Krowor Municipal Assembly LLIN – Long Lasting Insecticide Net MAUP – Modifiable Area Unit Problem MSIG – Marie Stopes International Ghana NACP – National AIDS Control Programme NCDCP – Non Communicable Diseases Control Programme NGO – Non-Governmental Organization NHIA – National Health Insurance Authority NHIS – National Health Insurance Scheme NMCP – National Malaria control Programme NTP – National Tuberculosis Control Programme OCP – Obesity Control Program OLS – Odinary Least Squares Regression P&S – Procurement and Supplies PCA – Principal Component Analysis PHC – Population and Housing Census PHRL – Public Health Reference Library PPAG – Planned Parenthood Association of Ghana xiv University of Ghana http://ugspace.ug.edu.gh PRISM - Performance of Routine Information System Management ProMPT – Promoting Malaria Prevention and Treatment R2 – Coefficient of Determination SARS - Severe Acute Respiratory Syndrome SSDM – Supply, Stores and Drug Management TMA – Tema Metropolitan Assembly xv University of Ghana http://ugspace.ug.edu.gh ABSTRACT Background There is a growing demand for reliable small area health statistics by governments, researchers, health managers and programmes due to continuous surge in communicable and non-communicable diseases in developing countries. Countries’ routine health information systems data (facility-based data) have been identified as a key system building block of primary health care which produces information in greater detail to meet the high data demand. However, the huge amount of facility-based data collected daily, aggregated and reported monthly into the national health information systems have rarely been used for evidence-based decision-making. Instead, household surveys have become the primary source of data for local area public health decision-making. Household surveys have been designed to measure indicators at the regional and national levels only, and therefore are unable to produce reliable estimates for small areas due to the small sample sizes. Furthermore, household surveys are mostly conducted once every 5 years or more, therefore health managers and policy-makers must wait up to five years or above for Demographic and Health Surveys or Multiple Indicator Cluster Surveys to be conducted to enable them to assess what progress has been made or be able to monitor implementation of programmes and interventions which must be conducted on monthly, quarterly and yearly basis. Poor health facility reporting, inadequacy of methods, inadequacy of population denominators for calculating population-based rates and lack of external consistency checks were among the key constraints identified to be responsible for the limited use of health facility-based data for public health decision-making. Objective The main objective of this study was to explore the usage of geospatial and statistical techniques to address issues associated with limited use of facility-based data and to evaluate its value for small area evidence-based decision-making. xvi University of Ghana http://ugspace.ug.edu.gh Method This study is both a descriptive cross-sectional and an exploratory one. It made use of a number of datasets including: health facility visits data for malaria and diarrhoea recorded by all health facilities offering outpatient services in Greater Accra Region (GAR) in 2014, population and housing census data, Ghana demographic and health survey data (GDHS), climate data, health facility locations and base map of GAR. Completeness of facility reporting rate of the region and the districts were determined. Multiple Imputation by Chain Equation was used to fill the data gap created by non reporting facilities. Health facilities locations, public and private, have been geo-coded and linked to the enumeration boundaries of 2010 population census data to provide the denominators for small area disease rates calculations. Annual cumulative health facility visit rates for malaria and diarrhoea, calculated as health facility visits per 1,000 of the population for districts and catchment levels, was the main dependent variable for the spatial regression analysis. This study used geospatial tools such as choropleth map, hotspot analysis, and spatial autocorrelation (Moran’s I) to visualize and explore patterns in the facility-based data. Ordinary Least Squares (OLS) fit, and Geographically Weighted Regression models (GWR) were used to analyse spatial relationships between independent variables (household predictors of malaria and diarrhoea extracted from 2010 PHC data) and health facility visit rates for malaria and diarrhoea. The statistical analyses were performed using Microsoft Excel version 16 and Stata software package, version 14.4. The geospatial analyses were performed using GWR4 (Newcastle University, UK) and ArcGIS ( version 10.4.1, ESRI Inc., Redlands, USA). Results The study has found out that about 49% of the health facilities in Greater Accra Region in 2014 did not report outpatient data to District Health Information Management System 2 (DHIMS2), and the majority of these facilities were privately owned or a quasi-government. xvii University of Ghana http://ugspace.ug.edu.gh Completeness of facility reporting rate for the region was 53.1%. Half of the districts in GAR recorded completeness of facility reporting rate below the 80% limit set by WHO, suggesting poor facility reporting for the region as a whole. The overall annual cumulative health facility visits due to malaria and diarrhoea for Greater Accra Region in 2014 were 880,735 and 233,003 respectively. These included 44% and 43% imputed health facilityvisits due to malaria and diarrhoea respectively due to non-reporting facilities. A data consistency check has revealed a statiatically significant difference between 2014 DHIMS2 and GDHS under-5 year old health facility visits due to malaria and diarrhoea. There were far fewer health facility visits in total featured in DHIMS2 than one would estimate based on the GDHS, and the difference (for under-5s as a whole) was highly unlikely to be due to sampling error. Significant spatial variations were observed in the distribution of health facility visits due to malaria and diarrhoea at district and catchment levels. The rate ranges between 577 visits per 1,000 population to 90 visits per 1,000 populations for malaria and that of diarrhoea ranges between 118 visits per 1,000 population to 16 visits per 1,000 populations. For all age groups above 14years, more females visited the medical facilities (public and private) with malaria episodes than their male counterparts. Similar trend was observed for HFV due to diarrhoea. The local GWR model for health facility visit rates for malaria improved the Ordinary Least Square model fit considerably. Locally weighted 𝑅2 value (goodness of fit statistics) indicate how well the GWR model replicates the HFVR for malaria around proportion of persons living within house walls prone to malaria and proportion of persons within the lowest wealth quintile. 𝑅2 value by catchment area varies between 10% to 46% with a mean value of 30%. Health facility visit rates for malaria was found to vary spatially over Greater Accra Region. GWR results indicated that proportion of persons living in walls prone to malaria and xviii University of Ghana http://ugspace.ug.edu.gh proportion of persons within the lowest wealth quintile were significantly spatially non- stationary – evidence that spatial heterogeneity exists in these relationships. Conclusions The result of the study suggested that Greater Accra Region has poor facility reporting. Nonetheless, with the use of geospatial and statistical techniques this data was enhanced to produce plausible population-based indicators comparable to those found in literature for public health decision-making. The study used health facility visits due to malaria and diarrhoea to demonstrate the various methodological stages involved in enhancing the value of facility-based data. The study showed variations in health facility visits due to malaria and diarrhoea by age and sex groups. Spatial variations were also observed at district and catchment levels in health facility visit rates for malaria and diarrhoea. Disaggregation of the data to catchment levels has shown further variations in health facility visits which were obscured at district levels. The study revealed that modelled risk factors for malaria vary geographically over short distances and can be used to focus interventions at the population most at need. The majority of the findings in this study agree with literature and is a confirmation that DHIMS2 data can generate plausible indicators that can be used for evidence-based decision making. This study also demonstrates that facility-based data with its current challenges can be manipulated with geostatistical techniques to generate plausible population-based indicators for small area evidence-based decision. It is hoped that the findings of this study will encourage public health practitioners and other researchers to have more confidence in using routine facility-based data (DHIMS2) for effective health sector planning and decision-making xix University of Ghana http://ugspace.ug.edu.gh CHAPTER ONE INTRODUCTION 1.1 Background Disease dynamics in developing countries, have changed over the years due to changes in social, cultural and economic factors. Most countries in Africa now suffer from what is known as “triple burden of disease”- the onset of significant burden from non- communicable diseases (NCDs), while the burden from communicable diseases (CDs) remains high plus increased accidents, suicides and homicides (Frenk & Gómez-Dantés, 2011). Due to the continuous elevated level of disease burden of countries, key stakeholders such as governments, the research community, health practitioners and the general public have recognised the failures in use of regional and national level health indicators for decision-making for conditions at local or community levels, necessitating increasing demand for small area statistics (Manninen et al., 2016; Alho, 2001). Public health decision-making requires a frequent, comprehensive, accurate and more detailed data that shows trends and variation in the disease outcomes among different communities or geographical areas (Haggard et al., 1998). Over the years, the need for small area analysis has been of great importance, formulating local level policy, allocating of government resources, strategic planning and policy decision- making in the wake of decentralized systems of administration (Burgert, 2014; Whitworth, 2013). Health system researchers and policy makers have acknowledged that small area statistics are critical for providing answers to questions on diseases burden, possible risk factors, care-seeking behaviour, healthcare access and utilization, and quality of healthcare (Burgert, 2014). Due to data availability issues, quality of data, population denominators and 1 University of Ghana http://ugspace.ug.edu.gh suitable methodologies, small area analysis face challenges not commonly associated with large geographical scale analysis (Alho, 2001). Interestingly, there is no universally accepted definition for what constitutes a small area. Rao (2003) refers to small area as “any subpopulation or domain for which direct estimates of adequate precision cannot be produced”. Burgert (2014) provides a more comprehensive attempt at a definition, She refers to a small geographical area such as counties, wards, local government area, post codes, school districts census tracts and census blocks where if a population based nationwide survey were carried out, the resulting sample size may be too small to generate a direct estimation of adequate precision. With respect to this study, small area refers to health facility catchment areas. The benefits of small area analysis are critical to public health professionals, for identification of new threats, determining variations in health outcomes and contributing factors to disease conditions. By using small area analysis, one can comprehend specific local problems, understand where the actual needs are, implement interventions at the areas where they are really needed, and distribute the scarce resources so that they will be as effective as possible. If one does not have information identifying where exactly the risks are, addressing them can be very challenging. Despite the enormous benefits of small area statistics, its estimation is constrained by several factors such as: (1) non-availability of data, especially those covering more detailed population characteristics. The existing data are not tabulated for smaller areas, hence necessitating the use of proxy variables; (2) lack of suitable and easy to use methodologies for small area analysis; (3) non-availability of methods to determine location-specific factors such as population sizes, seasonal populations, facility closing and changes in zonings which have huge impact on calculating population-based rates; (4) confidentiality issues and; (4) 2 University of Ghana http://ugspace.ug.edu.gh non-existing small area boundaries and changing boundaries of small areas over time making time series analyses challenging (Manninen et al., 2016; Alho, 2001). The primary sources of data used in small area health analysis include censuses and routine health data (Smith & Morrison, 2005). Censuses are the largest, most elaborate data collection activity undertaken by any country (Leete, 2001). Censuses are conducted by countries mainly every ten years with the population structure and corresponding geographical distribution being primary purpose. Additional information collected may include housing, living condition, education, migration, industry, families and households, selected disease conditions etc. Many developing countries, because of very weak economies, struggle to get adequate resources to conduct a smooth census exercise (UNFPA, 2009). Nevertheless, countries endeavour to collect the census data even if the intervals are not regular due to its usefulness. Ghana has been quite regular in conducting censuses. Since 1921, the then Gold Coast, now Ghana, has conducted eight (8) successful censuses with the latest being the 2010 Population and Housing Census (PHC) (GSS, 2012). Routine Health Information Systems (RHIS) are created by most developing countries and are expected to provide routine data covering health services delivery at all levels of the health system. According to some authors (Belay & Lippeveld, 2013; Lippeveld, 2013) routine facility-based data is the most appropriate data source for local health managers for their daily decision-making. In 2004, USAID funded MEASURE Evaluation and other collaborators initiated a worldwide project for strengthening RHIS in less developed countries. Lippeveld (2013) stated that enhancing the efficiency in health service delivery require daily decision making on the services such as community outreach, monitoring quality care and progress towards health plan targets and equity. Since the 1990s, awareness among health managers and researchers about the importance of routine facility-based health 3 University of Ghana http://ugspace.ug.edu.gh information - for crucial public health decision-making, resource allocation, strategy development - and design of suitable health interventions has increased (Belay & Lippeveld, 2013; Lippeveld, 2013; WHO, 2008a; AbouZahr & Boerma, 2005; Kintu et al., 2004). However, these rich sources of data are seldom used for planning and decision-making at both local and national levels as a result of issues such as: completeness of facility reporting (Wagenaar et al., 2016; WHO, 2014; Mate et al., 2009), lack of external consistency of coverage rates (WHO, 2014; Skarbinski et al., 2008), methodological issues (Rao, 2003; Pfeffermann, 2002), and lack of population denominators and risk factor information (Purcell et al., 2016; Stefan, 2015; Alexandrescu et al., 2008). Previous studies have documented the challenges of facility-based data resulting in its limited use (WHO, 2008a; Heeks, 2006; Godlee et al., 2004). In Kenya and Mozambique for example, assessments of routine RHIS in 2003 showed poor data quality (Mavimbe et al., 2005). Limited RHIS data use for health planning and decision-making was found in South Korea and Brazil (Da Silva & Laprega, 2004; Chae et al., 1994). Currently, household surveys are being used as the primary source of data for small area evidence-based decision-making but with serious limitations – they are designed to measure health indicators at national or regional level but not for small geographic areas, they have small sample representation for small areas therefore any direct parameter estimate lack precision, they are collected every 5 to 10 years so, one has to wait for such a long period for any update (WHO, 2008b). The situation in Ghana, in terms of facility-based data production and use, is not different from other developing countries around the world. In 2005, the Ministries of Health (MOH) of Ghana, Mexico, Thailand, National Statistics Offices of South Africa and Uganda benefited from the first global partnership with Health Metrics Network (HMN). This network was dedicated to strengthening nation’s health information system. The goal of the 4 University of Ghana http://ugspace.ug.edu.gh partnership was to increase the availability and use of timely and accurate health information by facilitating joint funding and development of core country health information systems. The above initiative laid the foundation for the development of the first ever national health information management system for the country called District Health Information Management System (DHIMS). Later, the Ghana Health Service (GHS) through Centre for Health Information Management (CHIM) and in collaboration with University of Oslo, developed a more robust software christened DHIMS2. The DHIMS2 system enables health facilities to enter their summary reports directly into an electronic database. It is a comprehensive Health Management Information System (HMIS) for the management of routine health facility data at every level of the GHS (Nyonator et al., 2013). There has been massive investment made in the management and maintenance of DHIMS2 over the years in terms of ICT infrastructure (computers, servers, network, antivirus software and internet connectivity) and human resources. DHIMS2 database currently stores data from all the 216 districts in Ghana and has a very large mortality, morbidity and other health information from both private and public healthcare facilities. Despite the successful development and deployment of District Health Planning and Reporting Tool (DiHPART) by GHS for district health management teams to align their budgets with the evidence based interventions (Awoonor-Williams et al., 2016), very few districts however, are making efforts to use DHIMS2 data for planning and budget cycles. Consequently, a chunk of this data is left unused. Concerning availability of methods to determine location-specific factors such as population sizes, seasonal populations, facility closing and changes in zonings, and non-existing small area boundaries, small area statistics offer a means of geographically analysing and presenting health statistics at the local level. Many approaches have been adopted by researchers and statistical authorities to fill this data gap. One of such approaches is - to 5 University of Ghana http://ugspace.ug.edu.gh estimate small area statistics through the use of indirect methods and complex statistical techniques (Pfeffermann, 2002; Ghosh & Rao, 1994). The use of such complex statistical techniques was very unpopular. The next approach involved increasing the sample sizes to get a representative sample for the desired small area so that direct methods can be applied. Automatic increase in the sample size eventually increases the survey cost and survey time which render the approach unsustainable for most resource limited countries. The most recent and yet to be fully exploited approach by many developing countries, is the use of geospatial analysis techniques (Burgert, 2014; Lippeveld, 2013). GIS techniques are increasingly being used in small area analysis. GIS at the heart of small area statistics provides spatial data infrastructure, geospatially enabled statistics, interoperable multisource data and integration of statistical and geographic data (Manninen et al., 2016). GIS can meet new needs by making use of new data and better use of already existing data.` GIS has tools to provide accurate denominators for small areas from census data and investigate potential associations between small area health indicators and, health system factors (Cromley & McLafferty, 2012). Geospatial approaches have proven to be very effective tools for analysing disease conditions across space and time. It helps to identify gaps in healthcare equity, access and utilization. Spatial approaches are able to determine where the individuals can be exposed to risk or resources (Cromley & McLafferty, 2002), which makes it easier for evaluation of interventions. GIS has been considered an innovative and appropriate technology for the implementation of many community-based public health needs activities which have been hindered by lack of data in many developing countries. Routine health facility-based data and census data have been identified as ideal for small area analysis. Not much attempt however, has been made to use GIS technique on these data for better informed decisions and evidence-based policies. This study aims at exploring the use of geospatial and statistical techniques to provide solution to issues associated with the 6 University of Ghana http://ugspace.ug.edu.gh limited use of facility-based data which will eventually enhance the value of the data for small area public health decision-making. 1.2 Problem Statement A country’s RHIS data ( facility-based data) has been identified as a key building block of primary health care which produces information to enable public health decision -makers at all levels of the health system identify problems and needs, and make evidence-based decisions on optimal resource allocation and health policies (WHO, 2008a). Effective monitoring and supervision of health care programmes (Peersman et al., 2009) and, how risk factors and health resources vary between communities (Kogan et al., 2015), depend on complete and accurate facility-based data. Many developing countries have made huge investment in the management and maintenance of RHIS over the years in terms of ICT infrastructure (computers, servers, network, antivirus software and internet connectivity) and human resources. However, due to data quality and other methodological issues, huge amounts of routine facility-based data collected at numerous services points on daily basis are seldom used for decision-making (Belay & Lippeveld, 2013; WHO, 2008a; Rao, 2003; Smith, 2003). There is a growing demand for reliable small area health statistics which is the consequence of the fact that many institutions need more detailed information for evidence-based decision- making for specific geographic areas (Manninen et al., 2016; Alho, 2001). According to Voight (1967) there is a high demand for more data and in greater detail now (small area data) than in the past. Programs and interventions such as National Malaria Control Programme (NMCP), National Aids Control Programme (NACP), National Tuerculosis Control Programme (NTCP) and Non Communicable Diseases Control Programme 7 University of Ghana http://ugspace.ug.edu.gh (NCDCP), need small area statistics to improve health conditions in response to the continuous surge in communicable and non-communicable infections (Marshall, 2004; Dean et al., 1999). Local health managers therefore need small area statistics to guide decision- making and programme implementation. Lippeveld (2015) noted that RHIS data is the most suitable for planning and monitoring of health services and the time is now to improve this data for evidence-based decision-making at all levels of health delivery. Coupled with the known general poor quality of routine facility-based data, there are other major concerns militating against its use for evidence-based planning and policy decision making at the local and national levels (Hotchkiss et al., 2012; WHO, 2008a; Heeks, 2006; Godlee et al., 2004). These concerns include the following: (1) poor facility reporting (WHO, 2014; Mate et al., 2009); (2) lack of external consistency of coverage measures (WHO, 2014; Skarbinski et al., 2008); (3) inadequate population denominators and risk factor information (Purcell et al., 2016; Stefan, 2015; Alexandrescu et al., 2008); and (4) Lack of capacity and easy to use methods for estimating small area statistics (Rao, 2003; Pfeffermann, 2002). As a result, government, public health managers, researchers, local communities and general public do not usually use facility-based data for decision-making, rather they rely on household surveys. Household surveys have become the primary source of data for small area public health decision-making in developing countries at the expense of facility-based statistics (WHO, 2008b). In household surveys such as Demographic and Health Survey (DHS) and Multiple Indicator Cluster Surveys (MICS), sample sizes are not large enough to provide reliable estimates for small areas (Buescher, 1997). The statistics generated do not tell what one needs to know on daily or monthly or yearly basis about small areas such as 8 University of Ghana http://ugspace.ug.edu.gh districts level and smaller geographic areas because of the nature of their design which is tailored to measure regional or national indicators. In some cases, there will be no observation measured in some communities; therefore, it is impossible that any indicator using survey data could be calculated with 100% certainty in small geographic area. Sometimes, health managers and policy-makers have to wait for five years or more to conduct of the next round of DHS or MICS in order to assess what progress had been made for example in maternal mortality rate or burden of malaria. The use of national or regional health indicators to estimate disease risk in small areas such as districts, towns and communities is misleading and very problematic since these measures are characteristically subject to huge chances of disparities (Nutley & Reynolds, 2013). The resultant effect of not using the appropriate information/evidence for local public health decision-making is that, there is lack of sufficient knowledge about diseases conditions and population at risk by public health planners and policy makers on a continuous basis; healthcare interventions are not targeted at key risk factors nor population most at risk, and hence targets set for disease prevention and eradication are usually not achieved. This eventually leads to waste of scarce resources and further weakening trust of the health system. The GHS through its DiHPART and other programmes have started using DHIMS2 data for evidence-based planning for the regions, districts and even sub-districts but faced challenges in its implementation including systems design dysfunction, training problems, participant computer literacy limitations, staff turnover problems and staff resistance to change among others (Awoonor-Williams et al., 2016). As a result, the service is unable to take full advantage of the enormous investment made in DHIMS2 database, which collects majority of the routine data generated by healthcare facilities across the country. So, this study adopts geospatial approaches to integrate facility-based data with other data sources 9 University of Ghana http://ugspace.ug.edu.gh such as census, climate data and DHS to enhance its value. Consequently, one will be able to produce plausible population-based indicators and relationships for evidence-based public health decision-making in local areas. 1.3 Justification Using small area data to improve public healthcare delivery has become increasingly unavoidable due to its potential in providing population-based evidence for local area decision-making. Small area data provides evidence to help identify inequalities in healthcare delivery in terms of access, underserved areas and populations that suffer more than others from various adverse health conditions. By comparing indicators of several small areas, one may be able to see why a disease condition exists in one area and not in another. This may lead to identifying causes or contributing factors to a disease condition. It also unearths any associations between factors one would not otherwise observe. For a region that appears to have no disease conditions, small area analysis can identify geographic or population clusters within the larger area where disease conditions are more pronounced. It helps in prioritizing where to allocate resources. Scarce resources are judiciously used if they are targeted at areas where disease conditions are most serious. It is expected that the outcome of this study will contribute to knowledge and methodology by providing GIS-based methodologies for determining population denominators for small area disease rate calculations and demonstrate how crucial evidence locked up in routine facility-based data can be exploited for small area public health decision-making. Policy makers, information producers, health managers, other health professionals can use the procedures employed in this study to better understand where the real health needs of 10 University of Ghana http://ugspace.ug.edu.gh communities are, so that they can prioritize and tailor interventions to where they are really needed, and apportion the scarce resources more effectively. Linking routine facility-based data to other traditional data sources such as census and DHS for public health decision-making, is a novelty that will inure to the benefit of health researchers, health managers and health policy decision makers. Users of health data will be able to validate their information and cross-check their evidence for optimal output. This linkage will extend the base of data to provide evidence on broad range of health issues and at all levels of public health decision making. 1.4 Objectives 1.4.1. General Objective To explore the usage of geospatial and statistical techniques to address issues associated with limited use of facility-based data and to evaluate its value for small area evidence-based decision-making. 1.4.2. Specific Objectives To achieve the general objective, the following specific objectives are addressed: 1. To determine completeness of facility reporting rate for Greater Accra Region into DHIMS2. 2. To determine the comparability between <5 years old health facility visits (outpatient cases) in DHIMS2 and GDHS for Greater Accra Region. 3. To determine variability in health facility visit rates (outpatient rates) and how this variability is compared between district and community levels. 11 University of Ghana http://ugspace.ug.edu.gh 4. To determine plausible spatial relationships between health facility visit rates (outpatient rates) and environmental / household risk factors extracted from census. 5. To establish geographical heterogeneities of the relationship between health facility visit rates (outpatient rates) and environmental / household risk factor 1.5 Conceptual Framework Figure 1.1 shows a conceptual framework for geospatial techniques applied to facility-based data by addressing the challenges that limit its use for public health decision-making at all levels and most especially in small areas. The framework indicates the key challenges that limit the use of facility-based data by government agencies, public health managers and researchers, and highlights the a priori solutions to each of the challenges which can ultimately lead to transformation of the data to be used for evidence-based decision-making. In this framework, limited use of facility-based data for decision-making is influenced by several factors including: completeness of facility reporting, lack of external consistency of coverage measures, inadequate population denominators and risk factor information, lack of easy to use methods and unobserved factors. Completeness of facility reporting and lack of external consistency of coverage measures are quality assessment needs of the facility-based data so that users can have confidence in the health data presented and results of the analysis. WHO (2014) noted that, it is vital to know the reliability of coverage estimates derived from RHIS data that are generated, as these often form the basis for evidence-based decision- making. Completeness of facility reporting is a measure of missing outpatient data (referred to in this study as health facility visits) of all health facilities (public and private) that do not report data to Ghana’s RHIS (DHIMS2). 12 University of Ghana http://ugspace.ug.edu.gh Figure 1.1 A framework for geospatial techniques applied to facility-based data Arrows represent the flow of information. The outputs measure the value of facility-based data for small area evidence-based decision-making. Source: Adapted from (Clements et al., 2013). These are a common occurrence with RHIS data and have a significant effect on the use of RHIS data for policy decisions as well as deductions. Missing data, if not dealt with in the most appropriate and desirable way, can affect the result and consequently the conclusions that can be drawn from such data. Since the missing values cannot be ignored, a statistical technique - imputation is used to fill in the missing values (Van Ginkel et al., 2007; Tsikriktsis, 2005; Rubin, 1976). External consistency of coverage measures refers to examining the external consistency or validation of the indicators generated from the health facility data. This involves comparing 13 University of Ghana http://ugspace.ug.edu.gh the outcome of facility-based data (DHIMS2) with that from a survey data, for example GDHS (WHO, 2014). GDHS is a very useful source of socio-economic information on households. It provides information such as, infant and child mortality, information on malaria and diarrhoea treatment, prevention, and prevalence, anaemia among women and children etc. In this framework, facility-based data (under-5 years old DHIMS2 HFV due to malaria and diarrhoea) is the basis of comparison with that of GDHS for external consistency check. Small area population denominators are of particular importance to local public health mangers for calculation of a range of rates and indicators, and for planning and monitoring services (ONS, 2016). PHC data provides information on population sizes, structure and distribution for all levels (i.e. from national, regional, district to EA levels) and on household characteristics such as household sizes, housing, utilities, etc. Similarly, the environmental data from the community provides information about the geographical locations of healthcare facilities and some drivers of diseases such as rainfall and temperature. It also includes geographically referenced coordinates of administrative areas and boundaries. In this framework, facility-based system data gives information about features of the health system and disease information that can be linked to environmental, socio-economic and demographic information in the same location. In this context, facility-based data serves as the major input which provides the anchor for small area analysis. Determining population-based rates for major diseases of small areas poses methodological challenges. The lack of consistent and easy method to enumerate population denominators for small areas outside the political administrative well-defined geographical areas has made it difficult for public health managers to calculate population-based rates for most local areas. Small area population estimation methods that utilize GIS techniques are used in this framework (Wang & Wu, 2010; Wu et al., 2005). Since the method was not tied to 14 University of Ghana http://ugspace.ug.edu.gh administrative boundaries associated with census, it allows population estimates to be made for user defined areas such as catchment areas in this study. Visualisation tools are used to create proportional circles and choropleth maps to show disease distribution in the study area. Showing data by the above-mentioned methods can be very useful for observing general patterns, detection of clusters and determining health-seeking behaviour of the people. Observed patterns are analysed using explorative tools such as Spatial Autocorrelation (Moran’s I) tool and Getis-Ord Gi* statistic to determine significant clusters and possible hotspots respectively. Geospatial models tools such as Geographically Weighted Regression are used to examine relationships between dependent and independent variables. All other factors that may be correlated to the limited use of facility-based data but could not be measured by this study were categorised as unobserved factors. It is believed that if users become aware of the availability and importance of facility-based data, there will be political will to employ additional qualified data officers with the right logistics and deliberate attempt to drive demand for facility-based data; its value will be enhanced for decision making. Geostatistical process and approaches adopted in this framework will enhance the value of facility-based data by generating plausible population-based rates, predictive maps and regression diagnostics for small area evidence-based public health decision-making. 1.6. Justification for the choice of malaria and diarrhoea for this study The following reasons were considered for the choice of malaria and diarrhoea as case studies for this study. Firstly, malaria and diarrhoea diseases remain major public health threats, and they are among the first five causes of death in Ghana, especially among children under 14 years old and therefore place a huge financial burden on the local governments and the nation 15 University of Ghana http://ugspace.ug.edu.gh at large for their prevention, control and possible eradication. Secondly, malaria and diarrhoea have a focal spatial distribution which is critical for this study. For example, malaria, in pre-elimination and elimination phases, has hotspots of transmission in which the risk of the disease (including asymptomatic parasitaemia) and number of cases are higher than in surrounding areas (Bousema et al., 2012; Ernst et al., 2006). Spatial heterogeneity of malaria occurs over a wide range of scales (Clements et al., 2013), and includes within- community variation, between-community variation, and variation among communities within administrative or other areas at subnational, national, regional, and global scales. Thirdly, malaria and diarrhoea diseases were selected to enable comparison and checking external consistency of routine health facility-based data against GDHS data - which equally collect data on same disease conditions but for children under-5 years old. Finally, these diseases are most suitable for any small area analysis because they are widespread conditions and therefore are not subject to problems with small numbers of cases when data has been disaggregated to small areas. 1.7. Structure of the Thesis This thesis is organised into six (6) chapters. The first chapter provides the general idea about the research, starting with a background of the study. This was followed by the research problem, justification and objectives (main and specific objectives).The conceptual framework of the thesis was described followed by justification for the choice of malaria and diarrhoea for this study. The chapter concludes with the structure of the thesis. Chapter two presents institutional arrangements of the Ghana’s Health Sector, disease burden in Ghana, small area statistics and sources of small area data. A GIS and Health, comparability between DHIMS2 and GDHS data, spatial relationship between health facility 16 University of Ghana http://ugspace.ug.edu.gh visits & environmental/ household risk factors, spatial heterogeneities of relationship between health facility visits & environmental / household risk factors, statistical approaches to small area analysis, approaches to small area analysis and areal data analysis were discused. The chapter concludes with a discussion of health facility catchment area and population denominators for health facility catchment area. Chapter three gives a general description of the study area. It includes physical, demographic and socioeconomic conditions in GAR and profile of 16 districts of the region. This is followed by study design, study population, inclusion and exclusion criteria and operational definition of the main outcome variable. The data types, sources and data process have also been captured. The chapter concludes with data analysis and ethical considerations. Chapter four has been organised into 16 major sections. The first section discussed health facility distribution in GAR. The next two sections discussed determining the level of completeness in the coverage of DHIMS2 dataset and comparing DHIMS2 data to GDHS data. This is followed by measuring variations in HFV due to malaria and diarrhoea among the population of GAR. A-spatial analysis of HFV and HFVR for malaria and environmental determinants and spatial distribution of HFV due to malaria and diarrhoea by districts and spatial distribution of HFV due to malaria and diarrhoea by catchment. Also discussed in this chapter include cluster detection using spatial autocorrelation (Global Moran’s I) tool, Getis- Ord Gi* (hotspot) analysis of HFVR for malaria and diarrhoea, and distribution of HFVR for malaria and diarrhoea by health facility types. These are followed by discussions of model variables and relationships, Ordinary Least Square (OLS) Regression Analysis, testing for clustering of residuals using spatial autocorrelation (Global Moran’ I) and Geographically Weighted Regression Model The final two sections present GWR model result and spatial variation in HFVR for malaria and some key findings. 17 University of Ghana http://ugspace.ug.edu.gh In Chapter five the main findings of the study based on areas captured in the main objective are presented. The first two sections are captured in a question form as what can DHIMS2 tell us about outpatient attendance? and Do DHIMS2-derived pseudo-incidence rates show plausible patterns? The following sections include data system and methodological challenges associated with analysis of outpatient data, limitations of the study and review of study objectives. Chapter six presents summary of key findings, conclusions reached in this study together with recommendations and contribution to knowledge. 18 University of Ghana http://ugspace.ug.edu.gh CHAPTER TWO LITERATURE REVIEW 2.1. Institutional arrangement of the Ghana’s Health Sector Figure 2.1: Structure of the health sector of Ghana Source: Second Five Year Programme of Work (2002 – 2006, p.43) As shown in Figure 2.1, Ghana’s Ministry of Health (MOH) is at the apex of healthcare delivery and is answerable for all health related issues in the country. It is responsible for direct public health service delivery or provision in the country. However, with the enactment of an ACT 525 of parliament, the functions of promotion, preventive, curative and rehabilitative care has been delegated to the Ghana Health Service (GHS) and Teaching 19 University of Ghana http://ugspace.ug.edu.gh hospitals. Hence, the ministry is now responsible for policy formulation, monitoring and evaluation, resource mobilization and regulation of the health service delivery in the country (MOH, 2016). The ministry provides a framework for human resource development and management for health, regulation of health service delivery and practice; procurement, distribution, and use of health sector commodities, drugs, services and works. These functions of the MOH are executed directly by the following seven (7) directorates: Policy, Planning, Monitoring & Evaluation (PPME), Human Resource for Health Development (HRHD), Research, Statistics & Information Management (RSIM), Procurement & Supply, General Administration, Finance and Internal Audit. The MOH has oversight responsibility over GHS and other institutions such as the Teaching Hospitals (Komfo Anokye Teaching Hospital, Korle-Bu Teaching Hospital, Cape Coast Teaching Hospital and Tamale Teaching Hospital); Christian Health Association of Ghana (CHAG); Ghana Medical and Dental Council; The Pharmacy Council; Ghana Ambulance Service; Ghana Traditional and Alternative Medicine Council. The ministry monitors and evaluates the health service delivery of the above-mentioned organizations. The entire health services delivery responsibilities have previously been performed by MOH until Ghana Health Service ACT 525 (GHS and Teaching Hospitals Act) was passed in 1996 when it was relinquished to GHS. ACT 525 led to the establishment of GHS and granting of autonomy to Teaching hospitals in Ghana (MOH, 1996). The Minister of Health through the GHS governing Council is responsible for the implementation of national policies. Administratively, GHS is organised at three levels: National, Regional and District levels. In terms of functions GHS is organised at five (5) levels which include national, regional, district, sub-district and community levels. The main objectives of GHS includes implementation of approved national policies for health delivery in Ghana, prudent management of resources available for the provision of the healthcare services and increased access to quality healthcare delivery (MOH, 2016). 20 University of Ghana http://ugspace.ug.edu.gh The core functions of GHS are the provision of comprehensive healthcare delivery at all levels in Ghana directly and by outsourcing out to other health agencies in the country such as Quasi-Government institutions - an institution in the private sector that began as government agencies but have since become separate entity, Faith-Based or Mission Hospitals, Private Health Institutions and Traditional Health Practitioners. The GHS performs a number of functions in order to achieve its objetives which include: 1. Development of appropriate strategies and set technical guidelines to achieve national policy goals/objectives. 2. Management and administration of the overall health resources within the service. 3. Promote healthy mode of living and good health habits of the people. 4. Establish very efficient and effective mechanism for disease surveillance, prevention and control. 5. Determine charges and fees for GHS with the approval of the Minister of Health. 6. Provide work-related training and continuing education for workers. 7. Perform any other functions relevant to the promotion, protection and restoration of health in Ghana (MOH, 2016). 2.2. Disease Burden in Ghana In Ghana, an even bigger problem exists; the problem of the triple burden of disease. The burden from CDs remains high, the burden from NCDs surges on and there is an increasing number of injuries - road traffic accidents, falls, self-inflicted injuries and violence. The burden of communicable diseases, non-communicable diseases and injuries together forms the triple burden of disease. Changes in social, cultural and economic factors over the years have given rise to a change in disease pattern from largely communicable to non- communicable diseases (NCD) and injuries (Adams et al., 2004). Hypertension, diabetics and 21 University of Ghana http://ugspace.ug.edu.gh cardiovascular diseases have become widespread problems in Ghana with increasing prevalence of risk factors. Under the millennium development goals (MDGs) efforts have been made to improve disease conditions in the country. Despite these improvements, the ten leading causes of outpatient visits in Ghana have changed very little over time. Communicable diseases continue to be a major concern in Ghana with a significant volume of resources allotted for their control, elimination and eradication (Adams et al., 2004). The top ten causes of HFV and death include malaria, lower respiratory infection, diarrhoea, rheumatism and other joint pains, upper respiratory tract infections, skin diseases, sepsis, hypertension, acute urinary tract infection (WHO, 2015b; CDC, 2013). Malaria accounts for 38% of health facility visits, 35% of total hospital admissions and 19% of all causes of deaths recorded in 2013 (MOH, 2014). The CIA World Fact book (2016) rates the degree of risk of major infectious diseases in Ghana such as malaria, typhoid fever, acute respiratory infection, hepatitis, dengue fever and other infectious diseases as very high. The Centre for Disease Control (CDC, 2013) gives the following figures for mortality in Ghana: Malaria (13%), Lower respiratory infection (9%), HIV(9%), Stroke (8%) and Cancer(6%). The challenge threatening the Ghanaian health sector is the co-existence of a high burden of infectious diseases with an increasing prevalence of Non Communicable Diseases (NCDs). These require fundamental changes in the capacity and structure of demand for health care, with a more complex case mix and a more detailed service utilization patterns. Addressing these health concerns in a planned and timely manner would require evidence-based decision- making at local levels and as well as targeted interventions. 22 University of Ghana http://ugspace.ug.edu.gh 2.2.1 Malaria About 100 countries worldwide are at risk of malaria (WHO, 2015c). In 2015 alone, an estimated 214 million cases of malaria were recorded which resulted in approximately 429, 000 deaths. More than 60% of these deaths occurred in under five children living in Africa (WHO, 2016b, 2015c). The WHO African Region bears a sizeable proportion of the global malaria burden and accounted for 90% of morbidity and 92% of mortality in 2015 (WHO, 2016c). Like most WHO African Regions, malaria is endemic in Ghana. The nation’s entire projected population of 27 million in 2014 were at risk of malaria disease with children under five years old and pregnant women judged to be at a higher risk of the disease due to lowered immunity. According to the MICS 2012 report, malaria accounted for 3.1 to 3.5 million cases annually of the disease burden in Ghana. In 2012, malaria accounted for 34% of under-five mortality and 18,000 deaths (WHO, 2015b), although a life-threatening malaria disease is preventable and treatable. Malaria is caused by a parasite known as Plasmodium and the disease transmission is a life cycle. The transmission begins when people are infected through the bites of infected anopheles - female mosquito known as the malaria vector. In humans five types of malaria parasite species exist and 2 of these species are P. falciparum and P. vivax – which pose the greatest threat (WHO, 2016a). In Africa, the most prevalent malaria parasite is P. falciparum, which is the most perilous of the five human malaria parasites. The main malaria parasite in most countries outside sub-Saharan Africa is P. vivax. The mosquito parasites when injected into the bloodstream through a bite travels to the liver producing millions of harmful organisms called merozoites. The merozoites later takes over the blood cells, attacking them and further infecting them with parasites. This action leads to tiredness, elevated temperature flunctuations of the body, sweating, loss of appetite and feverish conditions. Sickness and complications from the disease could last for months if left 23 University of Ghana http://ugspace.ug.edu.gh untreated. At the later stage the merozoites then start to change into sexual forms of the parasite which mature to reproduce and circulate in the bloodstream. The cycle starts all over again when the next time a mosquito bites an infected person and the disease is transferred through their bloodstream. Several studies have investigated risk factors of malaria and available literature has shown several factors found to influence rapid spread of the malaria disease. These factors have been broadly categorised as natural vector breeding sites and environmental factors (Antonio- Nkondjio et al., 2011; Matthys et al., 2010; Dongus et al., 2009), artificial vector breeding sites (Castro et al., 2010; Chaki et al., 2009), type of housing construction (Snyman et al., 2015; Wanzirah et al., 2015), household risk factors (Ghebreyesus et al., 2000), socio- economic factors (Imbahale et al., 2010; Mensah & Kumaranayake, 2004), household behavioural indicators (Ernst et al., 2009), migration or travel factors (Martens & Hall, 2000) and socio-demographic factors (Sharma et al., 2015). It is widely agreed that natural breeding and environmental factors mostly sustain the vector population and are commonly found in the rural areas (Sharma et al., 2015). These natural breeding grounds include coastal environments where most of the places are characterised by shallow salt waters including lagoons which are known to collect stagnant waters serving as excellent breeding reservoirs for vectors. Rivers and floodplains are also among the natural breeding grounds providing good aquatic environment for the mosquitoes. Altitude also plays a key role in malaria spread. Altitude influences the distribution and transmission of malaria indirectly, through its effect on temperature. As altitude increases, temperature decreases, so highlands are colder and lowlands are warmer. The increased temperature allows the development of parasites to occur in the mosquitoes, and the mosquito population also increases as the temperature rises (Bødker et al., 2003). 24 University of Ghana http://ugspace.ug.edu.gh The artificial vector breeding sites identified by literature include urban agriculture, drains, ditches, and gutters and swimming pools. These sites are mostly found in urban and peri- urban areas provide abundant sources of mosquito larvae near African urban dwellings. Housing characteristics and their association with malaria have been widely investigated. The housing construction features identified to be associated with malaria infection include type of dwelling, type of wall, type of roof and type of floor. Household behavioural risk factors such as water sources, availability of toilet facilities and type of fuel use for cooking, animals co-residing inside household, sleeping place at night and means to prevent mosquito bites are known to have significant impact on spread of malaria. 2.2.2 Diarrhoea Diarrhoea is defined as the passage of three or more liquid or loose stools per day, or more frequently than is normal for the individual (WHO, 2013). According to Ghana’s Ministry of Health’s Standard Treatment Guidelines (MOH, 2010), diarrhoea is passing frequent, loose, watery stools three or more times in a day. Diarrhoea can be caused by viral and parasitic organisms. It is a symptom of an infection in the intestinal tract which is usually accompanied by vomiting. Diarrhoea occurs in all age groups however, it is very common in children. WHO (2013) publication has shown that globally, there are 1.7 billion cases of diarrhoea disease each year and it is the second leading cause of death among under five children. The disease is responsible for the death of nearly 760, 000 children every year. Over the years, Africa and South Asia account for more than 80 per cent of child deaths due to diarrhoea (Unicef, 2010). In Ghana, diarrhoea is the third leading cause of death in children under five years and 5,100 children die every year (GNA, 2015a). 25 University of Ghana http://ugspace.ug.edu.gh There are three types of clinical diarrhoea which include acute watery diarrhoea, acute bloody diarrhoea and persistent diarrhoea. Accute watery diarrhoea includes cholera which can lasts several hours or days. Acute bloody diarrhoea is also known as dysentery and persistent diarrhoea lasts 14 days or longer. The episode of diarrhoea usually is accompanied by dehydration – loss of water. Most people who die from diarrhoea die from severe dehydration and fluid loss. Essential body nutrients such as sodium, chloride, potassium, bicarbonate and water are lost through liquid stools, vomit, sweat, urine and breathing. The main modes of transmission for most diarrhoea diseases are by absorption of contaminated food and water. Most of the diarrhoea causing pathogens are transmitted mainly by the faeco – oral route which may be food borne, water borne or direct transmission (WHO, 2013; Black et al., 2003). Direct transmission may imply collection of other faecal oral routes such as via fingers, fomites or dirt which may be ingested by young children. Diarrhoea infection is also spread from person-to-person for example, while caring for a sick person, via contaminated water, food or because of poor personal hygiene. Exposure to diarrhoea pathogens in developing countries has been widely associated with factors such as quality of safe drinking water, availability of toilet facilities, place of residence, housing conditions, level of education, poor socio economic status, feeding practices, and the general sanitary condition of household which include both personal or domestic hygiene (Anteneh & Kumie, 2010; Eshete, 2009; Green et al., 2009; Genser et al., 2006). Drinking unsafe water is a major cause of diarrhoea and cause of death for children below the age of five in developing countries. In a study to explore the association between environmental conditions, drinking water quality, and feeding practices with diarrhoea, Peletz et al. (2011) found the disease to be higher in children from households with poor water quality. Additional evidence suggests that continued exposure to excreta-related pathogens in 26 University of Ghana http://ugspace.ug.edu.gh early life limits cognitive development and lowers immunity from diarrhoea diseases caused by loss of vitamins and mineral deficiencies (Millward & Bell, 1998). Significant variation was observed in cases of diarrhoea between children in households with an improved toilet facility and those with non-improved toilet facility (Obila, 2012). A study by Gachogu (2003) in Kenya on factors causing diarrhoea and its treatment found out that location of residence significantly affects the disease. The study revealed that children born to educated mothers were more likely to receive treatment for diarrhoea compared to children born to less educated mothers. A study that explored the impact of diarrhoea in babies on the quality of life of low-income household by Murray et al. (2008) concluded that poor housing conditions worsened the effect of diarrhoea in infants on the primary caregiver as well as the other household members. 2.3. Small Area Statistics 2.3.1. Definitions and delineation of small areas Small area statistics have long been used (Rao & Choudhry, 1995). However, in recent years the demand for small area statistics has grown tremendously as they are required by health managers and policy-makers who wish to make a disparity-focused assessment (Song et al., 2016) or to adopt evidence-based public health policies (Whitworth, 2013). They demand more detailed geographic information to better target interventions or distribute resources more efficiently (Portnoy et al., 2014; Whitworth, 2013). Small area analysis in health has been described by most researchers as a procedure or method of analysing the variation in rates of health outcome or in utilization of care in small geographic areas in order to improve the performance of local health system and to promote health equity (Portnoy et al., 2014; Parchman, 1995; Cohen et al., 1992). It is important in revealing local-level disparities often 27 University of Ghana http://ugspace.ug.edu.gh concealed by health indicators for large areas such as regions and country. Small area analysis procedure focuses on the evaluation of small well-defined geographic areas or populations to examine the differences among them and between them and compare with the pattern in the larger geographic area. According to some authors (Smith & Morrison, 2005; Smith, 2003), the definition of small area is relative as demographers might treat a county as small area and likewise the scientist might treat a metropolitan region as small area and so on. The Office for National Statistics (ONS)-UK, define small area as the national population broken down into smaller geographic areas such as super output areas (SOA), health geographies, wards, parliamentary constituencies, zip codes and national parks (www.ons.gov.uk). The term small area typically refers to a small geographical area or a spatial population unit for which reliable statistics of interest cannot be produced due small sample representation from a nationwide survey. A recent research conducted by Asiimwe et al. (2011) used districts in Uganda as small areas to examine mortality of children under 5 years of age using small area analysis techniques. The term “small area” is synonymous to other terms such as “sub-national”, ‘small domain’, ‘local area’, ‘minor domain’ ‘small sub-domain’(Rao, 2003). In the present work small area has been defined as districts and catchment areas of health facilities in GAR. 2.3.2. Small area analysis in health Small-area analysis in health is crucial in revealing local-level disparities often masked by health estimates for large areas. The role of GIS technology in this analysis cannot be over emphasised. Health managers, providers and public health practitioners are increasingly concerned about how health outcomes, risk factors, and health resources vary over space and time. Understanding how these factors vary across small segments of the population, such as 28 University of Ghana http://ugspace.ug.edu.gh across dissimilar districts or communities, can aid health professionals design interventions for the populations who are most at need. Application of advanced and more rigorous small area analyses are needed to help inform public health decisions that can improve healthcare delivery. 2.4. Sources of small area data Prioritizing communities for targeted intervention is essential for effective public healthcare delivery especially in resource constraint countries. National and local health authorities and policy makers need valid and specific local data for health system evaluation and resource allocation. Main sources of small area data (i.e. most widely collected) for public health decision making include: Population and Housing Census Data, Routine administrative data, Targeted sample surveys and RHIS data (e.g. facility-based data from DHIMS2), (Smith, 2003). 2.4.1. Population and Housing Censuses The PHC has been defined as the total process of collecting, compiling, evaluating, analysing and publishing or otherwise disseminating demographic, economic and social data pertaining, at a specified time, to all persons in a country or a well-defined part of the country (UN, 2010). Most of the developed countries conduct censuses at regular intervals (e.g., every 5 or 10 years). Developing countries, due to scarcity of resources also conduct censuses but at lower frequencies (UN, 2010; Smith, 2003). These censuses constitute the countries’, most comprehensive source of small-area data covering a range of crucial data characteristics such as: population - age, sex, race, marital status, income and education; housing – number of housing units, number of rooms and type of material used. 29 University of Ghana http://ugspace.ug.edu.gh Censuses are a very intensive source of data which are more accurate and suitable for heterogeneous unit analysis but are also very expensive to conduct, limited in data collected and time consuming. The ten year interdental period creates gap in the data which is usually filled by projections which sometimes turn out to be too high or too low once the actual census results are released (Raymondo, 1992). This study used malaria and diarrhoea risk factors extracted from the 2010 PHC data at the EA level. The malaria risk factors used include person’s educational level, economic status, housing characteristics and type of agriculture activities. The type of housing characteristics contained in the census data are type of walls, dwellings, roofing materials and floor materials. The diarrhoea risk factors were people’s toilet facility sharing, source of water for domestic use, method of rubbish disposal, source of drinking water and type of toilet facility and method of liquid waste disposal. 2.4.2. Routine administrative records Administrative records are gathered routinely and primarily for administrative purposes and not research. This type of data is harvested routinely by government departments and allied institutions for the purposes of transaction, registration, and record keeping, usually during the delivery of a service. These administrative records provide small-area data for years in between censuses (Smith, 2003). Information captured by administrative records include: health facility-based data, school enrolment, social insurance, births, deaths, building permits, drivers licenses, tax returns and voter registration. It is reported that even some European countries can produce census statistics based solely on administrative records (Longva et al., 1998). On the contrary, the administrative records of many developing countries, are inundated by inaccurate and incomplete coverage (Cleland, 1996). This poses serious challenge for data use for small area analysis. 30 University of Ghana http://ugspace.ug.edu.gh 2.4.3. Sample surveys A sample survey is a process for collecting data on a sample of observations which are selected from the population of interest using a probability-based sample design (Carlson et al., 1998). Sample surveys are another source of demographic and socio-economic data for small area analysis. This type of survey requires a smaller scale of operation, reduces time and cost needed to collect and to process the data. Most survey data do not provide health statistics at district, sub-districts and community levels as they are designed to collect national and regional statistics. This has become a major limitation, as samples become too small to provide reliable or meaningful estimates for decision-making (Rao, 1999). Thus, policy-makers most of the times had to rely on nation-wide datasets to understand the health needs of their communities which is by itself problematic. 2.4.4. Routine health information system data The foundation of a good decision-making across all health systems (policy development and implementation, governance and regulation, health research, human resources development, health education and training, service delivery and financing) relies on sound and reliable information (WHO, 2008c). According to Ashton (2015), health system performance depends on the use of quality patient information. RHIS collects health data from the health sector and other relevant agencies and ensures their overall quality. Hotchkiss et al. (2012), defines RHIS as “a system that provide information at regular intervals of a year or less to meet predictable information needs”. Another definition by Tegabu (2011) is “an ongoing data collection of health status, health interventions, and health resources”. In both cases routine and ongoing data collection is important and it serves as the main information source for both the population and individual interventions (Wagenaar et al., 2016). Examples of RHIS include routine health facility-based statistics; vital events registration and special program 31 University of Ghana http://ugspace.ug.edu.gh reporting; administrative data - drugs, workforce, training and research, revenue and costs and documentation; epidemiological and surveillance data; data on community-based health activities. There is a growing demand for the use of RHIS data for decision making to reach majority of people to improve the quality of healthcare delivery (WHO, 2011b). The global community agrees that a well-functioning RHIS is a backbone of a strong health system and hence pivotal to achieving better health outcomes (Belay & Lippeveld, 2013; Gimbel et al., 2011; WHO, 2008a). A well-structured and robust RHIS provides essential information for all levels and facets of health system services delivery, including health system policy development and implementation, governance and regulation, human resources development, health research, health education and training, financing and service delivery. The role of RHIS in healthcare delivery cannot be over emphasized. RHIS is very useful for planning and management of the health services at district level and below, and can potentially play a key role in program improvement and reporting at all levels. Most experts agree that much needed to be done by most developing countries in the collection of RHIS to provide the integrated management support between clinical care and public health care. Limited use of RHIS information for evidence-based decision making is perhaps the main causes of the current neglect of this very important resource. What is worrying is that, no extra resources are needed for the generation of this data, on a daily basis the healthcare sector is overflowing with this information, but the sector lacks the capacity to collect, communicate or use it effectively. 32 University of Ghana http://ugspace.ug.edu.gh 2.4.5. Ghana’s routine health information systems The Ministry of Health in Ghana has mandated the GHS to collect, analyse and report on all routine health services including health service data from missions, private and quasi- government health facilities throughout the country. Over the years efforts at strengthening HMIS have been ad hoc and focused on evaluation of reporting systems. Since 2003, the health care information system in Ghana has undergone organizational and structural changes as the need for timely and reliable information for policy development and health planning has become almost indispensable (Adjei, 2003). Of concern for the health sector was the need for the development of an information system with the focus on aggregation of information for review and monitoring of service performance at all levels in the healthcare delivery system. The health services information existed almost entirely in paper form, thus, the need to capture data digitally. The CHIM and PPME divisions of the MOH have designed a health management information system, which was started on a pilot basis. In 2012, the GHS in collaboration with the University of Oslo, developed a web-based open- source information system DHIMS2 for data collection, management, transmission and analysis. It provides baseline data for district planning implementation and monitoring on major indicators of disease patterns, preventive services and physical resources. It gives information feedback to all the levels of health care facilities. DHIMS2 has advanced features for data visualization, such as GIS, charts, reports, pivot tables and dashboards which give meaning to the data. In 2013, Ghana won the African Development Bank e-Health Competition held in Tunisia with DHIMS2 implementation. This was in recognition of Ghana delivering a record-breaking DHISM2 rollout covering the whole country in 6 months. Currently, the system is receiving data from all the 216 districts of the 10 regions in Ghana. 33 University of Ghana http://ugspace.ug.edu.gh 2.4.6. Quality of DHIMS2 data The quality of data is critical, not only for planning and delivery of services, but also for monitoring the performance of the health service and employees. Data collected and presented must be accurate, reliable, complete, legible, timely, and accessible to authorized users if they are to meet the requirements of the researchers, policy makers and health authorities (Askham et al., 2013). With the implementation of DHIMS2, all public and private health service providers are required to enter health data into the database directly. Managers at all levels have direct access to this information. As a result, processes that previously delayed reporting have been eliminated. Information consistency and reliability have improved as a result of the unified reporting system compared to previously. DHIMS2 implementation came with a number of external and internal mechanisms to ensure data meets standards and is accurate (Nyonator et al., 2013; Osei et al., 2013). With a consolidation module at the regional and national levels, DHIMS2 allows data entry at multiple entry points in each district. The regional information officers are trained to provide further training for district information officers. A simplified information capturing system was adopted which focus on critical information for district level planning, thereby reducing the reporting burden in primary care settings (Mutale et al., 2013). There are opportunities for quality improvement at different levels through feedback mechanisms. Health services data are visible to health services managers for their prompt attention and action on poor performing districts, programmes and activities. Data coverage has increased through the Community-based Health Planning and Services (CHPS) initiative. The CHPS innovation enabled trained frontline workers including data management officers to be integrated into communities to provide door-to-door health services. Data from these services are captured into the national database which previously was not the case and hence increasing data 34 University of Ghana http://ugspace.ug.edu.gh coverage and completeness. All the above mechanisms increase the quality and relevance of year-on-year comparisons. Though the current study has not instituted any quality check on its own, available literature has shown that Ghana’s health sector has observed tremendous improvements in data quality and reliability over the past years (Mutale et al., 2013; Osei et al., 2013). Amoakoh-Coleman et al. (2015) have concluded that routine maternal health services data in the GAR, available at the district level through DHIMS2 system is complete when compared to facility level - primary source data and is reliable for use. However, incompleteness of facility reporting to DHIMS2 remains one the key challenges affecting the quality of DHIMS2 data badly (JSI, 2013). All facilities (public and private) are expected to submit reports on key service outputs monthly to DHIMS2. However, mainly the public health facilities and a few private facilities report their data into DHIMS2. Most of the private health facilities located in the major cities in Ghana and majority of the quasi- government health institutions fail to report data to DHIMS2 and therefore compromise the quality of the data. Non-reporting of such magnitude is bound to affect any indicator estimated from the data. WHO (2014) noted that it is critical to know the facility reporting completeness rate to make informed interpretation on key indicators estimated from such data. 2.4.7. Geographic Corodinates of health facilities Geographic coordinates of the health facilities were obtained from two sources namely DHIMS2 database and USAID|Deliver Project, Ghana. USAID|Deliver Project is one of the JSI Research and Training Incorporation’s flagship international health projects funded by U.S. Agency for International Development (USAID). The USAID | Deliver Project provides 35 University of Ghana http://ugspace.ug.edu.gh support for global goals to reduce malaria morbidity and mortality around the world. The project supply anti malaria commodities to over 28 countries, contributing enormously to the worldwide decline of malaria prevalence (JSI, 2017). In Ghana, the project works with country collaborators such as GHS and MOH directorates and programs (P&S, SSDM, NTP, OCP, NACP, NMCP, FHD, and PHRL), several International & local NGOs (PPAG, GSMF, MSIG.) and USAID implementing partners (ProMPT, BCS, FRHP, EXP-SM) to improve outcomes in country product availability through the procurement and delivery of high quality health commodities and strengthening the in-country systems that manage and deliver them (USAID|DeliverProject, 2016). This data was collected by the officers of the above- mentioned institutions who physically visited the health facilities and recorded their coordinates with Geographic Positioning System (GPS) devices. 2.5. GIS and Health GIS is a computer system used to store, manipulate, analyse, model, and visualise spatial and non-spatial data (Chang, 2015; Granell et al., 2009). This emerging technology and new research approach (though not really well-established now) is simply an information system that can be used to efficiently capture, organize, store, manipulate and analyse spatial data. GIS for health applications remain untapped most especially in developing countries and offers significant potential for exploitation. The power of GIS as a tool is through (i) its ability to link geographic data and attribute databases and (ii) provide query ability to these databases to identify patterns of health outcomes. Its ability to link geographical features on a map with attribute data is proving more and more useful in the analysis of health data and planning of healthcare services. GIS plays a crucial role in determining increasing accessibility of health care services and improving the quality of care. GIS has been used for 36 University of Ghana http://ugspace.ug.edu.gh several decades in developed countries to examine health care systems and the extent of its contribution has grown very rapidly and widely (McLafferty, 2003; Briggs & Elliott, 1994). The remarkable potential of GIS to benefit the healthcare industry is just beginning to be realized in developing countries. One of the earliest application known to students and practitioners of public health is when Dr John Snow in 1854 during an outbreak of cholera in the Soho district of London which killed 127 residents over 3 days, created maps of these cholera deaths (see Figure 2.2). Snow mapped the 13 public wells and all the known cholera deaths around Soho, and noted the spatial clustering of cases around one particular water pump on the southwest corner of the intersection of Broad (now Broadwick) Street and Cambridge (now Lexington) Street. Figure 2.2. John Snow’s cholera outbreak map of London The letter X indicates the locations of the water pumps. Source: https://clareverse.files.wordpress.com/2006/12/613px-snow-cholera-map.jpg 37 University of Ghana http://ugspace.ug.edu.gh He examined water samples from various wells under a microscope, and confirmed the presence of an unknown bacterium in the Broad Street samples. Despite strong scepticism from the local authorities, he had the pump handle removed from the Broad Street pump and the outbreak quickly subsided (Brody et al., 2000; Snow, 1855). Most developed countries use the resource integration capabilities of GIS to create analytical and descriptive solutions. In a research conducted by Yerramilli and Fonseca (2014) in ten counties with urban-rural settings in Mississippi, USA, GIS Network Based Methods were used to identify hotspots of vulnerable populations burdened with diseases due to a lack of geographical accessibility to right kind of health services. In another research census tract (CT) mapping and kernel density estimation methods in GIS were used to present geographical distribution of late-stage breast cancer diagnoses in Los Angeles County (Agustin et al., n.a). Yet in another study, Geraghty et al. (2010), used GIS to assess outcome disparities in patients with type 2 diabetes and hyperlipidaemia. GIS experts in developing countries are also making use of this emerging technology. Tanser and Wilkinson (1999) for example used GIS / GPS technology to document and quantify improved access to tuberculosis treatment through a community-based programme in Hlabisa, South Africa. Through the MARA /ARMA project, African experts were able to seize the opportunities offered by GIS technology to develop a product which will be of prime importance in the epidemiological surveillance of malaria in Africa (Adjuik et al., 1998). In Hong Kong, China, during the 2003 outbreak of Severe Acute Respiratory Syndrome (SARS) disease, Environmental Systems Research Institute (ESRI) used SARS Mapping Web site to provide clear and early notification of outbreaks and dissemination of information (ESRI, 2009). During the 2015 series of the Data Speak conference organised by Health Resources and Services Administration (HRSA) - Maternal and Child Health, Kogan et al. (2015) explaining GIS techniques in small area analysis stated that: 38 University of Ghana http://ugspace.ug.edu.gh “Small area analysis is a set of GIS methods to analyse phenomena at the local level, using data at county, ZIP Code, census tract, block group, block or latitude-longitude coordinate levels, provided we have access to that level of data”. GIS methods for small area analysis help examine differences in small geographic areas and find variations in health risks, resources and outcomes. Routine administrative and census data are very crucial in these analysis as long as they are available to the smallest geographic unit. Geospatial tools in small area analysis produce spatial patterns which are displayed on choropleth maps. A choropleth map is one in which the map is divided into small areas and each of the areas is categorised based on where its value falls within a range of values. The maps produced are not the end to the analysis as they are used as starting point to find out what is responsible for the pattern. 2.6. Comparability between DHIMS2 and GDHS data District Health Information Management System (DHIMS2) is a web-based database used to manage routine data collected by the health facilities. This data management system allows health facilities to collate and upload their data directly to the DHIMS2 database with instant access at the district, regional and national level. Small health facilities that lack internet facilities and manpower to upload their data to the database continue to forward data to the district office for uploading. Monthly morbidity data were collected on outpatient visits at each health facility by age and sex groups for 61 different diseases with diagnoses by the attending medical practitioner (Adaletey et al., 2013; Osei et al., 2013). Malaria and diarrhoea diseases used as case studies in this study were part of the diseases managed by DHIMS2. 39 University of Ghana http://ugspace.ug.edu.gh On the other hand, Ghana Demographic and Health Survey (GDHS), is a nationally representative survey of women age 15-49 and men age 15-59 from a number of interviewed households. The primary purpose of the GDHS was to generate recent and reliable information on fertility, family planning, infant and child mortality, maternal and child health, and nutrition. In addition, the survey collected information on malaria treatment, prevention, and prevalence among children age 6-59 months; blood pressure among adults; anaemia among women and children; and HIV prevalence among adults (GDHS, 2015). The field work for 2014 GDHS occurred between September to December 2014. The data were self-reported (by mothers of children under 5 years who were part of the sample selected) malaria and diarrhoea visits of children <5 years who had malaria or diarrhoea within the 2 weeks preceding the survey and sought advice or treatment in a health facility (both public and private). Based on this background information, it is expected that HFV due to malaria and diarrhoea extracted for an average of 2 weeks within the period September to December 2014 will be same as the number reported by 2014 GDHS for the same period. This study examines the level of agreement between facility-based DHIMS2 HFV due to malaria and diarrhoea and self-reported malaria and diarrhoea HFV obtained from GDHS dataset of 2014. 2.7. Spatial relationship between health facility visits & environmental/ household risk factors The health and well-being of humans have being gravely affected by the environment and therefore, targeting public health interventions to populations and places with greatest need is an essential and effective strategy for improving population health (Cromley & McLafferty, 2012). Knowledge of a health-care facility’s catchment area is important for assessing health service utilization, for calculating population-based rates of diseases and for performing other 40 University of Ghana http://ugspace.ug.edu.gh important analyses (Kate Zinszer et al., 2014). One of the objectives of this study was to determine plausible spatial relationships between health facility visits (outpatient cases) and environmental / household factors extracted from census data, which is key for evidence- based decision making in small areas. The impact of where people live in terms of space on their health cannot be overemphasised. In sub-Saharan Africa and other developing countries, most deaths of children under age five have been linked to the household environment (Wanzirah et al., 2015; Lwetoijera et al., 2013; Antonio-Nkondjio et al., 2011; Ghebreyesus et al., 2000). Literature has shown significant relationships between the household environment and child survival. Household environmental health hazards have been found to explain the level of differences in childhood morbidity and mortality between low and high under-five mortality countries. According to (Gollogly, 2009), providing safe drinking water and access to improved sanitation within the household environment can reduce the risk of mortality and morbidity among children under age five. Certainly, environmental household hazards are threats to the health of millions of people in the settings where they live (Bellamy, 2001; WorldBank, 2000). A variety of health hazards, including poor air quality, poor building standards, and contamination of water and food are present in the household environment (Racioppi, 2002). Both women and children are at elevated risk of exposure to the smoke emitted from burning charcoal, firewood, and other sources of fuel, due to women’s traditional role in food preparation. Rapid urban growth with its consequent crowded living conditions and lack of safe water and sanitation facilitate the spread of diseases that can affect adult and child survival (Mishra & Retherford, 2006; Rutstein, 2000). Child morbidity and mortality have been strongly associated with household environmental conditions in urban areas of Ghana, Egypt, Brazil, and Thailand (Timæus & Lush, 1995). 41 University of Ghana http://ugspace.ug.edu.gh It is widely agreed that the type of housing construction, housing characteristics including type of wall, type of roof etc., natural and artificial breeding environment and household behavioural risk factors such as water sources, availability of toilet facilities play a key role in malaria spread. Studies in some developing countries have found a significant incidence of diarrhoea diseases because of water shortage and contamination, early exposure to measles infections because of household crowding, and high risks of accidents or injury because of poor housing (Murray et al., 2007; Macassa et al., 2004; Yassin, 2000; Brockerhoff, 1995). . 2.8. Spatial heterogeneities of relationship between health facility visits & environmental / household risk factors Spatial heterogeneity refers to the uneven distribution of a trait, event, or relationship across a region (Anselin, 2010). The concept is to answer the question “Is the intensity of occurrence of an event equally distributed across the landscape?” Understanding the factors influencing these distributions is vital in designing efficient disease elimination programmes. It is important to determine the existence of spatial heterogeneity in coverage of infectious disease such as malaria and diarrhoea, so that the heterogeneity can be reduced in future through targeted effort in high endemic areas. Spatial analysis of infectious disease processes recognizes that host–pathogen interactions occur in specific locations at specific times and that often the nature, direction, intensity and outcome of these interactions depend upon the particular location and identity of both host and pathogen. Geographical landscape and spatial setting contribute to the likelihood of initial disease occurrence, direction and velocity of disease spread and susceptibility, and the design of appropriate control and management strategies (Real & Biek, 2007). 42 University of Ghana http://ugspace.ug.edu.gh Over the years, studies on the variations between urban and rural malaria transmission often made comparisons on a regional (Real & Biek, 2007) or even global scale (Tatem et al., 2013; Gething et al., 2011; Hay et al., 2005). Most of these studies compared malaria transmission on a binary scale between urban areas and a disjoint rural region, some-times separated by hundreds of kilometres. This global method does not capture malaria transmission adequately as spatial heterogeneity is already observed over short distances (Kreuels et al., 2008), even between households (Bousema et al., 2010). Heterogeneity was incorporated into the current studies to establish geographical heterogeneities of the relationship between health facility visits & environmental / household risk factors. 2.9. Approaches to small area analysis Small area analysis approaches fundamentally depend on questions of how to produce reliable estimates of the outcome of interest for small areas using complex statistical theories and practices referred to as Small-Area Estimation (Chandra, 2003; Rao, 2003). A critical review of small area analysis approaches has revealed that data source and data adequacy in a geographical area to provide reliable and stable estimates are key in the choice of a small area analysis approach. Thus, an approach that tends to use sample survey data and other auxiliary data utilizes statistical model-based methods popularly known as “Small-Area Estimation techniques” and small area analysis approach based mainly on routine administrative data uses the relatively new GIS-based methods. The two main approaches to small area analysis are the statistical approach and GIS approach 43 University of Ghana http://ugspace.ug.edu.gh 2.9.1. Statistical approaches to small area analysis Small area methodologies try to fill the gap between official statistics and local request of data. Small Area Estimation (SAE) is concerned with the development of statistical procedures for producing efficient (precise) estimates for small areas that is for domains with small or zero sample sizes (Rao, 2015; Salvati et al., 2010; Chambers & Tzavidis, 2006). The target is the estimation of a parameter (average/percentile/proportion/ rate) and the estimation of the corresponding prediction error. A sizeable number of different methods have been proposed for estimating small area estimates. The estimators concerned can be classified in many ways. One way of classifying estimators is by the three-way distinction which include: Direct estimators -derived from the data in the specific area concerned; Synthetic estimators - derived by fitting a model, often a regression model, to the available data, and reading off the predicted value; and Composite estimators: - which consist of weighted combinations of the two. Another way in which estimators can be classified is according to the type of unbiasedness properties they are meant to have. Design-unbiased estimators are unbiased in each specific area under repeated sampling. Model-unbiased estimators are unbiased in a sense that depends on the statistical model used in the estimation process (Rao, 2015; Salvati et al., 2010; Rahman, 2008; Chambers & Tzavidis, 2006; Rao, 2003; Pfeffermann, 2002; Ghosh & Rao, 1994). The model-based approaches borrow information from outside areas with similar characteristics to the area of interest. Statistical models are used to link information from respondents who are outside to those in geographical area of interest. Census, administrative data and other auxiliary data from related small areas are borrowed and incorporated through statistical linking models, increasing the small area sample size and increase the precision and reliability of the estimate. 44 University of Ghana http://ugspace.ug.edu.gh Three of the estimators – the direct estimator, the generalised regression estimator (GREG) and Empirical Best Linear Unbiased Predictors (EBLUP) were explained in this study for the benefit of the readers (Heady & Ralphs, 2005). The direct estimator is design unbiased. The GREG is a composite estimator which is approximately design-unbiased, and the EBLUP is composite estimators which is approximately model-unbiased. In order to describe the three estimator set-ups, the following standard notation were used throughout: 1. Y denotes the survey variable of interest 2. Lowercase letters refer to sample statistics and uppercase to population statistics; bold print refers to vectors and matrices; 3. Indices i and d refer to individuals and small areas (domains), respectively; 4. A bar above a variable refers to the mean – e.g. ?̅?𝑑 is the sample mean of for area d ; ?̅? is the vector of sample means for all areas. 5. A hat above a variable refers to an estimate – e.g., ?̂̅?𝑑 is an estimate of ?̅?𝑑 6. 𝑤𝑖𝑑 is the corresponding weight. 7. N is the population size 8. s refers to the sample The Direct Estimator The Direct estimator, is mainly based on survey design. Usually, sample representation in small areas are typically small within the areas of interest (e.g. region, districts and towns) leading to large sampling variability for these direct estimators. In other words, an estimate is referred to as "direct" if it is based only on responses from sample units in the area or domain of interest. 45 University of Ghana http://ugspace.ug.edu.gh The direct estimator is given by the formular: ?̂̅?𝐷𝐼𝑅𝐸𝐶𝑇 1 𝑑 = ∑𝑁 𝑖𝜖𝑠 𝑤𝑖𝑑𝑦𝑖𝑑 2a 𝑑?̂? where ?̂?𝑑 = ∑𝑖𝜖𝑠 𝑤 𝑑 𝑖𝑑 However, most sample surveys due to their designs, are unable to produce direct estimates for small areas like district and community levels, as the sample sizes are too small. The effect is that the direct estimates produced from such surveys for such small areas are not reliable or stable and therefore have low precision. The GREG Estimator Generalized Regression (GREG) is a model assisted approach or indirect approach, which is a more widely used option. The GREG is obtained by adjusting the direct estimator for an area for differences between the sample and population area means of covariates. The adjustments are calculated by using a model relating y and X. As a standard, the ordinary regression model is used. The formula for the GREG estimator is: 𝑇 ?̂̅?𝐺𝑅𝐸𝐺 1 1 𝑑 = ∑ 𝑤 𝑦 + [?̅? − ∑ 𝑤 𝑿 ] ?̂? 2b 𝑁 𝑖𝜖𝑠𝑑 𝑖𝑑 𝑖𝑑 𝑑 ?̂? 𝑖𝜖𝑠𝑑 𝑖𝑑 𝑖𝑑?̂? 𝑑 𝑇 where ?̅?𝑑 = (?̅?𝑑,1, … , ?̅?𝑑,𝑝) is a vector of p population mean covariates and ?̂?𝑑 = ∑𝑖𝜖𝑠 𝑤𝑑 𝑖𝑑 and ?̂? is the sampling-weighted regression estimator with unit-level covariates −1 ?̂?𝑆𝑊 = (∑𝑖,𝑑𝜖𝑠 𝑤𝑖𝑑 𝑿𝑖𝑑𝑿 𝑇 𝑖𝑑) ∑𝑖,𝑑𝜖𝑠 𝑊𝑖𝑑 𝑿𝑖𝑑𝑦𝑖𝑑 Another, equivalent, formula for the GREG estimator is ?̂̅?𝐺𝑅𝐸𝐺 𝑇 1 𝑑 = ?̅?𝑑?̂? 𝑆𝑊 + ∑𝑖𝜖𝑠 𝑊𝑖𝑑 (𝑦 − 𝑿 𝑇 ?̂?𝑆𝑊𝑖𝑑 𝑖𝑑 ) 2c ?̂? 𝑑𝑑 46 University of Ghana http://ugspace.ug.edu.gh When an area contains no data, the GREG reduces to a synthetic estimator, ?̅?𝑇 𝑆𝑊𝑑?̂? The main advantages of this estimator include design-based under linear regression model, asymptotic design un-biasness and design consistency. Its disadvantages include one level linear regression only with fixed effects, sensitivity to extreme values of the sampling inclusion probabilities, not robust against outliers and prediction not inclusive of spatial information. The EBLUP Estimator Model-based estimation - empirical-best linear unbiased prediction (EBLUP) estimator. This method assumes a linear mixed model that relates survey data to both unit and area level covariates, with variance components estimated using a technique which ensures strictly positive consistent estimation of the model variance. A parametric bootstrap method that incorporates all sources of uncertainty can be used to estimate variability parameters. The formula for the EBLUPS estimators are: For the model with unit-level covariates we define EBLUP-A = ?̅?𝑇 𝑢𝑛𝑖𝑡 𝑇 𝑢𝑛𝑖𝑡𝑑 ?̂? + 𝛾𝑑 (?̅?𝑑 − ?̅?𝑑 ?̂? ) 2d For the model with area level covariates we have EBLUP-B = ?̅?𝑇?̂?𝑎𝑟𝑒𝑎𝑑 + 𝛾𝑑 (?̅?𝑑 − ?̅? 𝑇 𝑑 ?̂? 𝑎𝑟𝑒𝑎) 2e The advantages of EBLUP estimator include efficiency under the assumption of Normality of the LMM and random effects at area. Its disadvantages include linearity of the relation with the auxiliary information, correlation between the random area effect, not design consistent, not robust against outliers and prediction not inclusive of spatial information. 47 University of Ghana http://ugspace.ug.edu.gh 2.9.2. Geospatial approaches to small area analysis GIS, though an emerging modern technology in healthcare industry in Africa, is growing rapidly as a new research methodology (incorporate spatial thinking and GIS technology into research design and analysis) to provide answers to small area analysis challenges (Graves, 2012). GIS approach to small area analysis is a set of tools for explaining patterns in health data displayed spatially (as maps). These output maps are often considered, erroneously, as the result of spatial analysis. Rather, the maps should be considered as the starting point for analysis more than the end point as what one is really interested in are the factors that drive that pattern. An infinite number of maps can be created for any variable measured across small areas. The goal is to understand the processes that create the observed patterns and to confirm they are patterns. Approaches to small area analysis include disease mapping (Bithell, 2000; Diggle, 2000), disease clustering (Diggle, 2000) and parameter shrinkage - Aspatial Multilevel Regression, Aspatial Empirical Bayes and Spatial Empirical Bayes (Meza, 2003; Bithell, 2000; Cressie, 1992). Other approaches to small area analysis are Fully Bayesian disease mapping (Gómez-Rubio & López-Quılez, 2006), Kernel Density smoothing (Hart & Zandbergen, 2014; Eck et al., 2008), Spatial Kriging (Berke, 2004) and Iterative weighted head-banging (Mungiole et al., 1999). 2.9.2.1.Disease Mapping Disease mapping is considered as exploratory analysis used in providing an impression of the geographical distribution of disease or the corresponding risk. The three basic types of disease maps corresponding to certain types of data are dot density maps for point (or case- event) data, choropleth maps for small domain data and isopleth maps for geostatistical data representing spatially continuous phenomena from a limited number of sampling locations. 48 University of Ghana http://ugspace.ug.edu.gh The dot density technique started in the 19th century and is today known as one of the prime techniques for representing geographic patterns. Dot density maps are used to illustrate spatial distribution of discrete objects. This type of map is designed and used to communicate variation in spatial density within the boundaries of geographic areas. Dots show the occurrence of an event thus illustrating a spatial pattern and relative density. Single dots might represent a single event, with each dot placed to denote the true location. In a situation, whereby it is not possible to map every event, dots may represent multiple events and are placed randomly within an area. To understand the dynamics of cancer cases for future strategies in the Western region of Tamil Nadu in India, Rathan et al. (2012) used dot density maps. While dot density maps are the versatile, easy-to-grasp and are appealing, they have some fundamental drawbacks. They are terrible for retrieving rates or numbers from the map. For example, few people will have the time or interest in counting hundreds (or thousands) or dots in order to know the precise number of people in Ghana. They are likely know that some places have “more” people than others, but they won’t know necessarily by how much. To help with this problem, you can add numbers directly on the map or provide a table to accompany the map. Also, although most dot density maps distribute dots randomly, map readers may potentially infer dot locations as precise locations of the phenomenon being mapped (e.g., the actual exact location of a person). To combat this, dot density maps should not be made at too large a scale. Furthermore, to avoid confusion, ideally dots should be distributed only in areas where the phenomenon actually exists (e.g., no dots in lakes for a map of population). In demonstrating how best maps can communicate geographic statistics more effectively, Pickle (2003) used breast cancer mortality ( 1988-92) rates, and choropleth maps with rates categorized into quintiles (20% of places in each category) were used. A choropleth map 49 University of Ghana http://ugspace.ug.edu.gh shows divided geographical areas that are coloured or shaded in relation to a data variable. The shading provides a way to visualise values over a geographical area and can show variation or patterns across the displayed location. Figure 2. is an example of a choropleth map, each area represents a district. Colour shows a classified number of disease cases per 1000 inhabitants. Figure 2.3. Choropleth map showing measles incidence per district – wave 8 2.9.2.2.Diseases Clustering Density is able show the locations where clusters exist in a data, but not if the clusters are statistically significant. Spatial clustering are those unusual concentrations of health events in space and time which can be detected either through surveillance or by searching patterns in routinely collected data (Neutra et al., 1992). The geographic extent or scale at which clusters are analysed is very critical and therefore affect the kind of inference drawn (Cromley & McLafferty, 2012). The scale at which a health problem is studied also reflects an understanding of the disease process and possible causative factors. Small area cluster 50 University of Ghana http://ugspace.ug.edu.gh analysis reflects localised factors such as point sources of health events. Cluster detecting methods can be divided into three groups which are global method (those that assess overall clustering in a study area), local method (those that seeks to identify cluster locations) and focused method (those that assess clustering around a point source) (Cromley & McLafferty, 2012). Flexible scan statistic for local cluster detection method was used to detect clustering in childhood leukaemia in the Republic of Belarus. The maximum spatial cluster size was set to 60 regions and the maximum size of time in a cluster was set to 5 years. The dataset was tested for two periods of time: 1986-2005 and 1990-2005. In the case of the 1986-2005 period a space-time clustering was detected over 1987- 1989 in 51 regions of the southeast Belarus with a lot of secondary significant clusters. In the case of the 1990-2005 period only one significant cluster was detected. Hotspot cluster analyses was used to detect clusters of kala-azar cases in Vaishali district (Bihar), India. Mapping kala-azar cases reveals the spatial heterogeneity in the case rate of kala-azar affected villages in Vaishali district. There was a significant positive spatial autocorrelation of kala-azar cases for five consecutive years (Bhunia et al., 2013). Hotspot analysis uses vectors to identify the locations of statistically significant hotspots and cold spots in the data. 2.10. Areal data analysis Areal data is associated with a fixed spatial region. Such data is usually aggregated to various levels of spatial units for both regular and irregular areas / zones. Levels of spatial aggregation may include townships, districts, regions or other less political boundaries such as catchment areas. Examples of such data include number of persons having malaria illness, number of health facilities, population density, average income etc. In most developing countries, routine health data at individual level is usually not available due to capacity, logistic and financial constraints. Health data is usually aggregated by facilities to the district, 51 University of Ghana http://ugspace.ug.edu.gh region and national levels. In terms of surveys, individual health data may be aggregated to protect patient’s personal information. Areal data like all other spatial data have the underlying characteristic of exerting spatial correlation to the data from neighbouring polygons. This phenomenon was expressed by Tobler (1970) in his first law of geography which states that “Everything is related to everything else, but near things are more related than distant things”. With the analysis of areal data, it is this spatial correlation which is visualised, explored and modelled. Moran’s I (Index) is used to measure the level of spatial autocorrelation. 2.10.1. Spatial Autocorrelation and Moran’s I Spatial autocorrelation analysis test measures the degree of spatial dependency (correlation) of an observed value of a variable at one locality and the values of the variable at neighbouring localities. If a dependence exists, the variable is said to exhibit spatial autocorrelation. Spatial correlation statistics may be classified as positive, negative and no spatial auto-correlation. A positive spatial autocorrelation value shows all similar values appear together or there is clustering of similar values across the geographic area, whiles negative spatial autocorrelation value is an indication that dissimilar neighbouring values cluster together in a map. Zero or no autocorrelation refers to the absence of positive and negative spatial correlation. i.e. the dataset is scattered (Wu, 2012). Measures of spatial autocorrelation include Joint Count Statistics, Moran’s I, Geary’s C and Getis-Ord G statistic. These measures are categorised as local and global measures. A global measure produces a single value which applies to the entire data set. The same pattern or process occurs over the entire geographic area (i.e an average for the entire area). In otherwise, a local measure produce values calculated for each observation unit. This means different patterns or 52 University of Ghana http://ugspace.ug.edu.gh processes may occur in different parts of the region (i.e. a unique number for each location)(Mathur, 2015) The two measures of spatial autocorrelation used in this study were Global Moran’s I and Getis-Ord G-Statistics. Moran’s I is the most common measures of Spatial Autocorrelation. It produces a correlation coefficient that measures the level of spatial autocorrelation of your entire data set. The tool computes Moran’s I Index value, a Z score and evaluates the significance of the index value. There is spatial clustering of the values associated with the geographic features in the study area when p-value is small and the absolute value of the Z score is large (or small) enough that it falls outside of the desired confidence level. If I> 0, then the set of features shows a clustered pattern, otherwise, if I<0, then the set of features shows a dispersed pattern (Mathur, 2015). Getis-Ord G-Statistic tool assess the overall pattern and trend of your data and identifies spatial concentrations. The G statistic distinguishes between hot spots and cold spots. G is relatively large if high values cluster together and relatively low if low values cluster together. The General G statistic is interpreted relative to its expected value. The value for which there is no spatial association, if G > (greater than) expected value it means potential “hot spots” and if G < (less than) expected value it means potential “cold spots”. A Z test statistic is used to test if the difference is statistically significant. Calculation of G is based on a neighbourhood distance within which cluster is expected to occur (Getis & Ord, 1992). 2.10.2. Spatial Data Modelling To understand why an event of interest occurred and what might be causing the occurrence of that event, regression analysis is used. Regression analysis allows the researcher to model, explore and examine spatial relationships, and explain the factors beneath the observed 53 University of Ghana http://ugspace.ug.edu.gh spatial patterns. Some of the best tools used in modelling spatial relationships include Ordinary Least Square (OLS) Regression and Geographically Weighted Regression (GWR). Ordinary Least Square (OLS) Regression OLS is a widely-used regression technique that serves as a good starting point for all spatial regression analyses. It provides an understanding of the global model of the variable or process and it creates a single regression equation to represent that process. It is used to explore a linear relationship between dependent variable and a set of explanatory variables. The technique may be applied to single or multiple explanatory variables. The OLS regression model can be extended to include multiple explanatory variables by adding additional variables to the equation. Numerous factors can produce residuals that are correlated with each other, such as an omitted variable or the wrong functional form. The OLS tool in ArcGIS automatically checks for redundancy (multi-collinearity )(O’brien, 2007). Each explanatory variable is given a computed Variance Inflation Factor (VIF) value. When this value is large (> 7.5, for example), it indicates existence of multi-collinearity . Significant multi-collinearity creates a problem in the multiple regression because since the inputs are all influencing each other, they are not actually independent and it is difficult to test how much the combination of the independent variables affects the dependent variable, or outcome, within the regression model. After these variables are identified they must either be removed from the model or modified by creating an interaction variable or increasing the sample size (O’brien, 2007). Geographically Weighted Regression. GWR is one of numerous spatial regression techniques, widely used in geography and other disciplines. GWR provides a local model of the variable or process one is trying to understand by fitting a regression equation to every feature in the dataset. The model assumes 54 University of Ghana http://ugspace.ug.edu.gh that the parameters are non-stationary but are functions of location. This method takes into account certain local characteristics of the regionalization that the stationary approach is unable to retrieve. It shows the ability of non-stationary approach to manage data with locally varying anisotropy. The method demonstrates how the non-stationary spatial dependence structure changes the shape from one place to another as compared to the stationary one (Fotheringham et al., 2003). Each of the regression coefficients estimated by GWR is a function of location hence, they can be mapped. 2.11. Health facility catchment area A health facility catchment area is a geographical area delineated around a healthcare facility that describes the population that utilizes its services. Usually facility catchment areas divide geographic space into neighbouring regions of healthcare delivery, which are used to assess health service utilization and calculate population-based rates of disease of the region. Several methods have been proposed to define healthcare facility catchment areas. Most of the approaches focused on defining the catchment area using a threshold value for geographic distance and population size. Phibbs and Robinson (1993), proposed one of such methods which used a circle of a specified radius around the physical location of the healthcare facility which contained from 70% to 90% of the healthcare facility’s patients. Luo and Qi (2009), when defining a catchment area used a fixed distance (e.g. 20 kilometres) or a road network travel time, whiles Alexandrescu et al. (2008) define a catchment area by selecting area that together account for 80% of a healthcare facility’s patients. A study by Baker (2001), used patient-flow method, which assigns geographical units to a healthcare facility catchment area, if the proportion of the facility’s total activity from that geographical region is above some threshold value or margin. One of the limitations of the above-mentioned approaches were that the threshold is pre-specified and not calculated from the data. 55 University of Ghana http://ugspace.ug.edu.gh Several other approaches have used data from catchment areas. For example, recent literature has shown healthcare catchment areas defined using K-means clustering (Gilmour, 2010) and cumulative case ratio (Zinszer et al., 2015) methods which the authors claimed are more superior to those existing approaches that do not use data from the catchment area. There are also SaTScan (Kulldorff (1997) and Bayesian hierarchical regression modeling (Wang & Wheeler, 2015) approaches to defining healthcare facility catchment area. This study has adopted the Thiessen polygon approach (Judge et al., 2009a) to define a catchment area. The Thiessen polygons (otherwise known as Voronoi polygons) were used to define health facility catchment areas in GAR. With Thiessen polygon, locations in space are closest to that healthcare facility than any other facility. With this approach, it is assumed that patients prefer to consult a health facility which is closest to their place of residence and in Euclidean space. Figure (2.) shows how the Thiessen polygons are constructed using ArcGIS software. In constructing the Thiessen polygons, all the health facility locations within a triangulated irregular network (TIN) that meets the Delaunay criterion. Figure 2.4. Creating Thiessen polygon Source:http://resources.esri.com/help/9.3/arcgisdesktop/com/gp_toolref/coverage_toolbox/creating_thiessen_pol ygons.htm 56 University of Ghana http://ugspace.ug.edu.gh For each triangle edge, the perpendicular bisectors are generated, which form the edges of the Thiessen polygons. The perpendicular bisectors are constructed by drawing circles with radius around the corresponding points. The vertices of the Thiessen polygon are at the location, at which the bisectors intersect. The resulting limits build the health facility catchment area and the resulting polygon is called Thiessen polygon. 2.12. Population denominators for health facility catchment area The absence of appropriate denominator data has been identified as one substantial problem associated with population–based rates of diseases for small areas, most especially in developing countries(ONS, 2016; Wang & Wu, 2010; Alexandrescu et al., 2008). The demand for good quality demographic data for small area analysis cannot be overemphasised (Smith, 2003) Small area population estimates are very important for both central government departments and local authorities for a range of purposes including planning and monitoring of services and as denominators for the calculation of various rates and indicators (ONS, 2016). In the United Kingdom for example, population estimates for small area geographies are often used for research and analysis. Currently, the two main categories of human population denominators available for small area analysis are population census data to the enumeration level and a gridded human population dataset. With the use of GIS, the enumeration level population census data can be linked to the geo-coded enumeration boundaries of population census data to provide the denominators for small area disease rates calculations as used in this research. However, unlike many developed countries, in developing countries, national census data are most of the times over a decade old, incomplete, unreliable and inconsistent (Devillea et al., 2014; Tatem et al., 2011) which is of concern of many of the users of this data. Advancement in GIS technology has led to the design and development of several spatial population models 57 University of Ghana http://ugspace.ug.edu.gh and databases which proved to be more accurate. These modelled spatial population datasets, mostly called gridded human population dataset have overcome majority of the challenges enumerated earlier about the census and survey data. Research has shown that the gridded population datasets are modelled at moderate to high resolutions that represents a more accurate distribution of human population than the existing census datasets and that makes explicit reference to urban and rural areas (Salvatore et al., 2005). The Gridded population is produced from data sets incorporating boundary or population updates for countries, additional attributes in the centroids and national identifier data sets, an updated water mask which includes more recent glacier data and local water data sources for high latitude countries, and additional format and resolution options. Separate rasters are available for population counts and population density consistent with national censuses and population registers, or alternative sources in rare cases where no census or register was available. All estimates of population counts and population density have also been nationally adjusted to population totals from the United Nation’s World Population Prospects. The process models the distribution of human population (counts and densities) on a continuous global raster surface to provide a spatially disaggregated population layer that is compatible with data sets from social, economic, and Earth science disciplines, and remote sensing. It provides globally consistent and spatially explicit data for use in research, policy-making, and communications. Most of these gridded human population datasets are free and form an essential population denominator required for many epidemiological studies (SEDAC, 2017). The following are some examples of such databases which are freely available on the internet: Worldpop, Gridded Population of the World (GPW) database, Global Rural Urban Mapping Project (GRUMP) and LandScan Global Population database. 58 University of Ghana http://ugspace.ug.edu.gh CHAPTER THREE METHODS 3.1. Study Area The study was carried out in GAR which has Accra as its capital city and which is also the administrative capital of Ghana. The GAR is bordered to the north by the Eastern Region, to the east by the Lake Volta, to the south by the Gulf of Guinea, and to the west by the Central Region (Figure 3.1). Figure 3.1. Map of study area – GAR Source: Survey Department, 2015 GAR is the smallest area of Ghana's 10 administrative regions, occupying a total land surface of 3,245 square kilometres or 1.4 per cent of the total land area of Ghana. It is made up of 16 administrative areas which include Accra Metropolitan Assembly (AMA), Ada East District, Ada West District, Adentan Municipal, Ashaiman Municipal, Ga Central Municipal, Ga East Municipal, Ga South Municipal, Ga West Municipal, Kpone-Katamanso, La Dade-Kotopon 59 University of Ghana http://ugspace.ug.edu.gh Municipal, La Nkwantanang-Madina (LANKMA) Municipal, Ledzokuku-Krowor (LEKMA), Ningo-Prampram District, Shai-Osudoku District and Tema Metropolitan Assembly (TMA). The GAR is the second most populated region, after Ashanti Region, with an estimated population of 4,010,054 according to the 2010 PHC, with 48.3% males and 51.7% females, accounting for 16.4% of Ghana’s total population (GSS, 2013). It is the most densely populated region in Ghana. The region has a total of 533 health facilities offering outpatient services in 2014 of which 394 (74%) are private, 120 (22%) are public (government) and 19(4%) are quasi-government. 3.1.1. Profiles of Districts in GAR Accra Metropolitan Assembly (AMA) AMA is mainly an urban settlement. According to the 2010 PHC: District Analytical Report the population of AMA, is 1,665,086 which is projected (GSS (2014a), to be about 1,841,091 in 2014 representing 42% of the region’s total population. Females constitute approximately 51% and males represent 49%. Among the population above 10 years, 89% are literate and 11% are non-literate. Among the economically active population, 93% are employed while 7.0% are unemployed. About 3.2% of households in the Metropolis are engage in agriculture. The housing stock of Accra Metropolis is 149,689. The main construction materials for outer walls of dwelling units are cement blocks (81.9%) and wood (11. 3%). Cement (82.5%) and wood (4.5%) are the two main materials used in the construction of floors of dwelling units. Metal sheets (47.0%) and slate/Asbestos (45.2%) are the two major roofing materials for dwelling units. The main sources of fuel for cooking by most households are charcoal (46.5%) followed by Gas (41.4%). The main sources of lighting in dwelling units in the Metropolis are electricity (93.8%) flashlight/torch (2.1%) and kerosene lamp (1.7%). The 60 University of Ghana http://ugspace.ug.edu.gh main sources of drinking water are pipe-borne inside dwelling (31.8%) bagged water in plastics popularly known as sachet water (28.3%), pipe-borne outside dwelling (28.1%) and public tap (9.1%). The most commonly used toiletry facility are public toilet units (41.6%), Water closet (31.9%) and KVIP (14.9%). About 2.3% of the households have no toilet facility. A third of households (36.8%) share separate bathrooms in the same house with 22.2% having their own bathroom for their exclusive use. The most widely used methods of solid waste disposal in the Metropolis is collection from home (57.4%) and by public dump (container) accounting for 32.9%. Liquid waste is mostly disposed into gutters (48.0 %), through a drainage system into a gutter (26.6%) and through a sewerage system (7.8 %) (GSS, 2014a). Ada East District Ada East District’s population is about 1.8% (71,671) of GAR’s population with 68.3% of them dwelling in rural areas. It has a household population of 70,470 with 15,631 households resulting in household size average of 4.6 persons. Females constitute about 52.5% and malesaccount for 47.5%. Among persons of 11 years and above, 72.8% are literate while 27.2% are not. Only 44.1% of households are into agricultural activities. Cement block and mud brick are the major building materials for house walls and floors in the district. In roofing houses, slates and metal sheets are mostly used. The district has an employment rate of 95% among its 70% economically active population. The main source of fuel for cooking by most households is charcoal (53.2%). The main sources of lighting in dwelling units in the district are electricity (60.6%) flashlight/torch (3.3%) and kerosene lamp (34.5%). Sachet water (20.8%), outside dwelling pipe-borne (20.6%), Public tab (20.3%) as well as protected well (11.7%) are the major sources of water in the district. Most (35.2%) dwellers in the district are without toilet facilities. KVIP serves about 20.8% of the residents while 1% of the 61 University of Ghana http://ugspace.ug.edu.gh district’s population uses bucket toilets. Solid waste is mostly disposed off by burning (41.3%) and open space refuse dumping (25.1%). Compounds (60.3%) and streets (24.6%) dropping happen to be the most used means of liquid waste disposal in the district (GSS, 2014b) Ada West District According to GSS (2014c), Ada West District has a population of 59,124 being 1.5% of the whole Greater Accra population. Almost 70.3% of its dwellers live in the rural places. The males’ population constitute 48.3%. The household population for the district is 57,746 with 11,642 households averaging its house household size to 5.1 persons. The literacy rate among people aged 11 years and more stands at 68.5%. The economic active employed group is 95.8% , with 42.1% employed in skilled agricultural, forestry and fishery works. Cement (68.6%) and earth (27.7%) are the main construction materials used in building house walls and floors while metal sheets (52.6%) and slates (25.0%) are used for roofing. Lighting is normally sourced from electricity (66.0%), kerosene lamp (27.2%) and dry cell batteries- torch (5.0%). Charcoal (55.2%) is the most used cooking fuel. Outside dwelling pipe-borne, public tap and sachet water are their major water sources. In the district 58.4% of households do not own toilet facilities while 18.2% uses public toilet (W.C, KVIP, pit, pan, etc.). Also, KVIP users are 11% of households in the district. Flinging waste on compounds and streets are the major media of disposing off liquid waste in the district. Solid waste is mostly dumped off at public refuse dumping sites (39.4%) while 29.7% of households burn their solid waste. About 10.5% indiscriminately dispose-off their solid waste while 9.6% of households have refuse collected from their homes by waste management companies. Most residents throw their liquid waste onto compounds (50.1%) and streets (34.2%) (GSS, 2014c). 62 University of Ghana http://ugspace.ug.edu.gh Adentan Municipal Assembly According to GSS (2014d), this assembly has 78,215 dwellers. It has 76,601 household population with 20,478 households with average household size of 3.7 persons per household. It has 50.3% (males) and 49.7 (females). Among people who are aged 11years and above, 91.9% are well-educated. The economically active population currently having jobs constitute about 91.2%. Agricultural activities are engaged in by only 7.1% of households. About 80% of dwellings are made of cement while the floorings are mostly done with cement and ceramic\tiles. Majority of the residents use electricity (71.7%), torch (12.1%) and kerosene lamp (9.3%) as lighting source. Cooking is mostly done with liquefied petroleum gas (47.2%) and charcoal (40.1%). The source of water in the municipality is from water tank supply (13.4%), sachet water (53.4%), and pipe-borne outside dwelling (15.4%). WC, public toilet and KVIP are the mostly used toilet facilities (59.5%). About 23.5% of households have no toilet facilities while others use pit latrine (26.8%). The commonly solid waste disposal method used is house-to-house collection (45. 8%) by waste contractors. Indiscriminate dumping of solid waste is found among only 4.0% of the households. Most (48.4%) of the households dispose liquid waste onto compounds and 15.4% onto streets (GSS, 2014d). Ashaiman Municipal Assembly According to GSS (2014e) report Ashaiman Municipality has a population of 190,972, which constitute 4.8% of GAR population. It has 185,804 household population with 49,936 households giving an average household size of 3.7 persons. Females constitute 50.9% of the districts population. The active economic population currently employed is 91.6%. Only 3.8% of the households are farmers. Majority of their dwelling structure walls are made of cement (76.7%) and wood (20.3%). Cement is the main floor material for dwelling structures 63 University of Ghana http://ugspace.ug.edu.gh (85.6%). Roofing of dwellings are mostly made of metal sheets (76.2%) and slates (19.9%). Lighting is normally sourced from electricity (88.5%). Dwellers of the municipality source their drinking water from pipe borne (91.9%) and sachet water (6.4%). House to house collection was identified as the most common method of solid waste disposal (45.8%). In the municipality 63.5% of households use the public toilet while only 4.0% have no toilet facility. Liquid wastes are mostly flung into gutters and through the drainage systems. Collection of solid waste is about 62.6% of household’s population while 28.5% of them dump into public refuse containers (GSS, 2014e). Ga Central Municipal Assembly The GSS (2014f) showed that Ga Central Municipality in 2010 had a population of 117,220 of which household population was 114,745 with 28936 households. i.e. an average household size of 4.0 persons. Females are 51.1% while males are 48.9%. In terms of education 92.8% of those aged 11 years and older were educated. About 4.6% of the 92.3% who are employed among the economically active group were engaged in agriculture activities. Majority (93.1%) of houses are built with cement and 2.9% with wood. About 79.8% of the households have the floors constructed with cement, while 7.5% had mud floors. Metal sheets and slate were the most common roofing materials. Sachet water (43.3%), pipe-borne within dwellings (18.9%) and outside dwellings (17.5%) serve as the major source of drinking water for households. Electricity (83.6%), kerosene lamp (6.7%) and flashlight (6.2%) were the basic source of light. Majority used charcoal (49.0%) and gas (43.8%) for cooking. The most used toilet facilities in the municipality were pit latrine (40.3%), WC (27.4%) and KVIP (20.8%). About 5% of households are without toilet facilities. In Ga Central, solid wastes are mostly disposed-off by burning (43.3%) and house to house collection (33.9%) whilst 9.6% use the public refuse container. Liquid wastes are 64 University of Ghana http://ugspace.ug.edu.gh mostly disposed by throwing onto compounds (46.5%), streets (18.6%) and into gutters (14.2%) (GSS, 2014f). Ga East Municipal Assembly With reference to GSS (2014g) report, the population of the Ga East Municipality in 2010 was 3.68% (147,742) of GAR’s population with 90% of them dwelling in rural areas. With household population of 144,863 this translate to an average household size of 3.9 persons. The female percentage stands at 51%. Among persons aged 11 years and above, 93.6% were literate while 6.1% were not. The municipality had an employment rate of 92.1% among its 70% economically active population. Only 5.5% of households are into agricultural activities. Cement blocks (83.7%) and wood (9.2%) are the major building materials for house walls but most floors are made of cement. In roofing houses, metal sheets are mostly used in the municipality. Most of the households (54.7%) uses gas for cooking and 36.3% uses charcoal. The main source of lighting was from electricity (78.4%). Sachet water (54.7%), pipe-borne water (24.9%) as well as boreholes (6.5%) accounted for the major sources of drinking water in the municipality. Most dwellers use WC (42.9%) and pit latrines (22.7%) while about 8.2% of the municipality’s population do not have toilet facilities. About a quarter (24.6%) of household’s burn off their solid wastes while the rest are carried by waste contractors. Compounds and streets are the common final destinations for liquid wastes (GSS, 2014g). Ga South Municipal Assembly Using the GSS (2014h) report- Ga South Municipality has a population of 411,377 of which 51.1% are females.Total household population was 404,130 and 100,701 households giving an average household size of 4.0 persons. For people who are aged 11years and above, 87.9% were educated with 92.0% of the economically active population are employed. Agricultural activities were undertaken by only 12.3% of households. About 84% dwelling block walls are made of cement while 7.8% are made of mud bricks. The flooring is mostly done with 65 University of Ghana http://ugspace.ug.edu.gh cement (78.2%) and mud (7.6%). Metal sheets were the major materials for roofing in the municipality. The main sources of fuel for cooking by most households is charcoal (48.6%). The main sources of lighting in dwelling units in the Metropolis are electricity (75.5%) flashlight/torch (6.6%) and kerosene lamp (13.4%). Water in the municipality is mostly sourced from pipe-borne water (65.5%), sachet water (22.1%), and boreholes (4.6%). WC, public toilet, pit latrines and KVIP are the predominantly used toilet facilities (72.6%) while about 13.5% of the population have no toilet facilities. The commonly solid waste disposal methods used are burning (37.8%) and house-to-house collection (21. 3%). Indiscriminate dumping of solid waste is found among 4.0% of the households. Liquid wastes are mostly disposed-off onto compounds (43.0%) and streets (22.4%) (GSS, 2014h). Ga West Municipal Assembly With reference to GSS (2014i), the population size of Ga West Municipality was 219,788 with most (51.0%) being females and 49.0% males in 2010. The district had a household population of 55,913 and an average household size of 3.9 persons. The census revealed that 92.8% of those aged 11 years and above were educated. Among the economically active group, 91.5% were employed with 6.9% of households into agricultural activities. Majority of house walls are built with cement (89.0%) and wood (5.6%) and flooring made with cement (76.1%), mud (5.6%) and terrazzo tiles (5.5%). Metal sheets and slate are commonly used in roofing dwelling structures in the area. Sachet water (63.2%), pipe-borne within (10.5%) and outside dwellings (8.6%) serve as the major source of drinking water for households. Electricity (85.5%), kerosene lamp (6.0%) and torch (5.0%) are the major used lighting source in the municipality. The main sources of households cooking fuel was liquefied petroleum gas (46.7%) and charcoal (44.0%). Most used toilet facilities in the municipality were pit latrine (28.9%), WC (29.7%) and KVIP (22.6%). About 6.2% of households are without toilet facilities. In Ga West, solid wastes are mostly disposed-off by house to house 66 University of Ghana http://ugspace.ug.edu.gh refuse collection (47.4%) whilst 11.5% use the public refuse container. Liquid wastes are mostly thrown onto compounds (37.7%), streets (22.5%) and into gutters (16.0%) indiscriminately (GSS, 2014i). Kpone-Katamanso District According to the 2010 Population & Housing - District Analytical Report (GSS, 2014j), Kpone-Katamanso District has a population size of 109,864 with 90.4% of the population dwelling in urban areas. Males are 48.7% and females are 51.3% of the district population. It has a household population of 106,398 and 24800 households with household size averaged 3.9 persons. Among persons of 11 years and beyond, 90.7% are literates. The district has an employment rate of 91.6% among its 75.1% economically active population. Only 8.3% of households are into agricultural activities. About 83.1% of dwelling block walls are made of cement while 12.0% are made of wood. The flooring is mostly done with cement and ceramic\tiles. Dwelling structures are mostly roofed with metal sheets and slates Charcoal is the main (45.8%) source of cooking fuel for households while they access light mainly from electricity (74.5%), kerosene lamp (13.4%) and flashlight (7.5%). Water in the municipality is mostly sourced from water tanker supply, public standpipe and pipe-borne water (15.4%). Public toilet (WC, Pit, Pan, VIP) are the mostly (27.1%) used toilet facilities. The commonly solid waste disposal method used is house-to-house collection (29.2%) and public dumping site (32.0%). Liquid wastes are mostly disposed of onto compounds (37.5%) and streets (29.3%) (GSS, 2014j). La Dade-Kotopon Municipality La-Dade Kotopon Municipality is fully an urban settlement with a population size of 183,528 (GSS, 2014k). Females constitute about 52.7%. With a household population of 179,251 and 51,154 homes, this gives an average of 3.6 persons per household. The economically active 67 University of Ghana http://ugspace.ug.edu.gh population was 91.4% with 3.1% of households being farmers. Most (83.6%) of the resident have cement structured outer walls and 11.3% have wooden outer walls. Cement and marble tiles are the main material used for residential structure floors (85.6%). Roofs of houses consist mainly of metal sheets (28.7%) and slate (63.8%). Charcoal (37.6%) and liquefied petroleum gas (45.7%) are the predominantly used sources of cooking fuel while households mostly get light from electricity (93.7%), kerosene lamp (2.1%) and flashlight (1.7%). Residents in the municipality get drinking water from pip-borne, sachet water and public taps/ boreholes. About 44.4% of residents use public toilet, 43% have access to WCs while 4.5% use KVIP. The rest (4%) have no toilet facility. House to house collection (77.1%) is the most common way of disposing solid waste among the households, with public dumping containers constituting 15.6%. For the disposal of liquid waste, the discharge of residues into gutters (39.4%), through drainage system into gutters (39.4%) and through sewerage system (11.6%) are the most common practices of households in the municipality (GSS, 2014k) La Nkwantanang-Madina Municipality (LANKMA) According to GSS (2014l) La Nkwantanang-Municipality is populated with 111,926 residents being 2.8% of the whole Greater Accra population. Most of the residents are females (51.5%). Almost 84.0% of its dwellers are in the urban areas. Of those aged 11 years and older 91.3% are educated. Of the economically active group, 92.3% are employed with only 5.3% of households being into agriculture. About 82.6% dwelling structures have walls made of cement while 12.1% are made of wood. The flooring is mostly done with cement (73.4%) and mud (6.7%). Metal sheets are the major materials for roofing in the municipality. Water in the municipality is mostly sourced from sachet water (61.9%), pipe-borne water (20.6%) and boreholes. About 52.6% of the households use liquefied petroleum gas the main source of cooking fuel, 85.8% of them have access light from electricity, 4.3% use kerosene lamp and 6.0% flashlight. Most used toilet facilities in the municipality are pit latrine (10.7%), WC 68 University of Ghana http://ugspace.ug.edu.gh (39.8%) and public toilet (18.1%), while some 5.1% of households are without toilet facilities. In La Nkwantanang-Madina Municipality, solid wastes are mostly disposed-off by house to house collection management contractors (67.6%) whilst 11.5% burn theirs solid wastes. Liquid wastes are mostly thrown onto compounds (22.3%), streets (22.5%) and into gutters (21.0%)(GSS, 2014l). Ledzokuku-Krowor Municipality (LEKMA) The GSS (2014m) reports that in 2010 the Ledzokuku-Krowor Municipality population size was 227,932 with 47.9% males and 52.1% females. The municipality has a household population of 221,757 with 60,859 households leading to an average household size of 4.0 persons. About 92.2% of those aged 11 years and older were literate. Among the economically active group, 91.1% are employed with 3.3% of households engaging in agricultural activities. Majority of houses were built with cement (91.1%) and wood (4.8%) with either cement (83.4%) or mud (5.4%) floors. Slate was commonly used (79.7%) in roofing houses. Charcoal (43.5%) and liquefied petroleum gas (47.6%) were predominantly used as sources of cooking fuel while households mostly get light from electricity (92.9%), kerosene lamp (2.1%) and flashlight (1.7%). Sachet water (23.6%), pipe-borne within (25.9%) and outside dwellings (23.6%) serve as the major source of drinking water for households in the municipality. Within the municipality there was access to public toilet (38%) and WC (25.7%) while 7.8% of households were without toilet facilities. Solid wastes are mostly disposed-off by house to house collection (33.9%) and the use of public refuse container (49.8%). Liquid wastes are mostly thrown onto compounds (43.3%), streets (18.6%) and into gutters (33.6%) (GSS, 2014m). 69 University of Ghana http://ugspace.ug.edu.gh Ningo-Prampram District According to the GSS (2014n) report, Ningo-Prampram District has a population of 70,923 dwellers about 1.8% of Greater Accra’s population. There were 68,521 people in 14,627 households i.e. an average household size of 3.7 persons. Females constitutes 52.7% and male 47.3%. About 71.2% of the people aged 11years and above are elites with 94.1% of the economically active population are currently employed. Agricultural activities were engaged in by 33.7% of households. Most of dwelling walls are made of cement and mud brick while the floorings are mostly done with cement (84.8%) and tiles (9.2%). Households are normally roofed with metal sheets and slates. Water availability is pipe-borne and public standpipe. Electricity (64.1%), kerosene lamp (28.1%) and torch (5.2%) are used as lighting source in the municipality. The main sources of households cooking fuel is charcoal (52.2%). Most (55.1%) of the residents are without toilet facility. Public toilet and KVIP are the most used toilet facilities. The commonly solid waste disposal method used is the public dumping container (30.5%) and burning (32.3%). Indiscriminate dumping of solid waste is found among 4.0% of the households. Liquid wastes are usually thrown onto compounds (58.5%) and streets (26.9%) indiscriminately (GSS, 2014n). Shai-Osudoku District The population size of Shai-Osudoku District in 2010 according to GSS (2014o) report was 51,913 about 1.3% of GAR’s population with about 76.7% residing in rural areas. Males and females constitute 48.7% and 51.3% of the population respectively. The household population stands at 50,021 with 11,862 households making its household size averaged 4.4 persons each. About 70.7% of all individuals aged 11 years and above are literates. The census revealed that 93.3% of the economically active group were employed with 85.6% of households engaged in agricultural activities in the rural areas. Most dwelling structures in the district were built with cement (59.3%) and earth bricks (32.8%). The floors are mostly 70 University of Ghana http://ugspace.ug.edu.gh made of cement (81.6%) and earth (14.1%). Metal sheets are the major materials for roofing in the municipality. Charcoal (45.7%) and wood (33.3%) are the major sources of cooking fuel for households while they access light mainly from electricity (53.7%) and kerosene lamp (32%). Drinking water in the municipality is mostly source from pipe-borne water (37.3%). A large number of dwellers in the district (31.2%) are without toilet facilities, public toilet users stand at 30.0% and pit latrines serves about 22.7% of the dwellers. Solid wastes are mostly disposed by burning (34.6%) and public dump space (31.2%). Compounds (63.1%), streets (22.2%) and drainage system (6.2%) happen to be the major destinations for liquid wastes in the district (GSS, 2014o). Tema Metropolitan Assembly (TMA) Tema Metropolitan has a population of 290,773 representing 7.3% of GAR and has an average household size of 4.1 persons per household because of having 285,139 household population with 70,797 households (GSS, 2014p). Males constitutes 47.8% of the population with 52.2% being females. Of the people aged 11years and above, 91.1% were educated. The economically active population employed was 90.4%. Only 3.6% of households are into agricultural activities. About 76.3% of dwelling block walls are made of cement while 19.8% are made of wood. The flooring is mostly done with cement and tiles, while dwelling structures are mostly roofed with metal sheets (59.9%) and slates (28.2%). The major source of water for the metropolis are pipe-borne water, public standpipe and water tanker supply. The commonly used source of cooking fuel is gas (51.7%) while light is source from electricity (86.7%), kerosene lamp (5.3%) and torch (4.4%). Majority (53.1%) of the households in the metropolis use WC while 30.8% uses public toilet (W.C, KVIP, pit, pan,). Flinging waste on compounds and streets are the major media of disposing off liquid waste in Tema. Solid waste is normally collected at house levels or dumped by the individual into public refuse containers (GSS, 2014p). 71 University of Ghana http://ugspace.ug.edu.gh 3.2. Study design This is a descriptive cross-sectional and an exploratory study designed to explore the usage of geospatial and statistical techniques to address issues associated with limited use of facility- based data and to evaluate its value for small area evidence-based decision-making. Annual cumulative outpatient data or otherwise health facility visits (HFV) due to malaria and diarrhoea obtained for Greater Accra Region (GAR) in 2014, were used as case studies, but the methods and procedures may be applied to other diseases. Out of 533 health facilities in GAR offering out-patient services, data could be obtained for only 271 (51%), which submitted their data to DHIMS2 database. Data for the rest 262 (49%) health facilities had been imputed to make-up for the missing data. GDHS data, EAs base map and PHC data (to enumeration area level) were obtained from the GSS. Climatic data was obtained from Ghana Meteorological Agency. Public and private health facility locations were obtained from GHS - DHIMS2 database and USAID|Deliver Project. The base map of GAR including district boundaries was obtained from RS/GIS Lab, University of Ghana. The outpatient (HFV) data for malaria and diarrhoea together with census data on attending populations were joined at the regional, district and health facility catchment levels. The regional level join was necessary to enable comparison with the GDHS data. The resultant number of HFV were compared with the GDHS to assess consistency of patterns in HFV due to malaria and diarrhoea. Annual cumulative health facility visit rates (HFVR) were computed based on the attending populations for each district and catchment area. Catchment areas where attending population were not available were dropped. Choropleth and dot density maps were respectively used to visualise malaria and diarrhoea rates and counts at district and catchment levels to examine their plausibility. Cluster detection analysis was performed on malaria and diarrhoea rates using Spatial Autocorrelation (Global Moran's I) and Getis-Ord Gi* statistic. 72 University of Ghana http://ugspace.ug.edu.gh Ordinary Least Squares (OLS) and GWR models used to see if plausible relationships emerged with risk factors extracted from the PHC data and to determine geographic heterogeneities for these relationships. The statistical analyses were performed using Microsoft Excel version 16, Stata software package, version 14.4, and geospatial analyses were performed using GWR4 © (Newcastle University, UK) and ArcGIS (10.4.1, ESRI Inc., Redlands, USA). 3.3. Study population The study population for this study was the entire population of Greater Accra Region in 2014. 3.4. Inclusion and exclusion Criteria The study involved all private, public and quasi-government health facilities offering OPD services in GAR in 2014. However, Maternity Homes, Teaching Hospitals and Specialist Medical Centres were excluded from the study. 3.5. Operational definition of main study outcome The main study outcome, enhanced value of facility-based data, was measured using plausibility of indicators calculated from DHIMS2 facility-based data (HFV due to malaria and diarrhoea) and existence of plausible HFVR and risk factor relationships in GAR. Thus, it was defined as the use of DHIMS2 HFV (outpatient data) for: • determining completeness of facility reporting rate for Greater Accra Region into DHIMS2. • finding out the extent of agreement between DHIMS2 facility-based data and GDHS; 73 University of Ghana http://ugspace.ug.edu.gh • displaying variability in HFV due to malaria and diarrhoea among districts and showing plausible indicators by means of tables, graphs and maps; • comparing variations in HFVR for malaria and diarrhoea at the district and catchment levels; • determining plausible spatial relationships between HFVR ( for malaria and diarrhoea) and environmental / household factors extracted from census; • observing the existence of geographical heterogeneities of the relationship between HFVR for malaria and environmental / household risk factors at the community level. 3.6. Data types and sources 3.6.1. Routine health facility-based data This study utilized HFV due to malaria and diarrhoea recorded by all health facilities in GAR in 2014 for illustration purposes. It is worth noting that in Ghana all health facility visits are suspected visits but not confirmed visits and therefore are only clinically diagnosed. All such visits are defined according to Ministry of Health Standard Treatment Guidelines (MOH, 2010). HFV due to malaria (Clinical malaria) is defined as an individual with fever (axillary temperature ≥37.5 °C) or with other malaria-related symptoms such as chills, severe malaise, headache or vomiting at the time of examination or 1–2 days prior to the examination in the presence of P. falciparum parasitaemia of any density. HFV due to diarrhoea (Clinical diarrhoea) is defined as the passage of three or more loose or liquid stools per day or more frequent passage than is normal for the individual. The facility-based data used in this study consisted of observed health facility visits and imputed health visits (HFV due to malaria and diarrhoea imputed for non-reporting health facilities). 74 University of Ghana http://ugspace.ug.edu.gh The observed data were obtained from DHIMS2 database through GHS headquarters and GAR Health Directorate. The database is maintained by Information Monitoring and Evaluation Unit of the PPME Division of the Ghana Health Service (GHS). It contains information on all routine health services including over 62 diseases collected and reported through mission, private, quasi-government and government health facilities across the country. Data are captured daily into registers, forms, files and ledgers in consultation rooms and departments of the various health facilities. The data are subsequently collated and summarized onto DHIMS2 compatible reporting format at the end of the month by the research assistants at the health facilities’ biostatistics department across the nation. It is then captured electronically into DHIMS2 database within one to two weeks of the ensuing month. The characteristics of the data include: disease name, district, facility name, sex, age group, month of collection, referrals and re-attendances. Geographic coordinates of the health facilities were obtained from two sources namely GHS - DHIMS2 database and USAID|Deliver Project, Ghana. The coordinates could be obtained for 513 health facilities out of the 533 health facilities under study. The majority of the coordinates (70%) were obtained from the USAID | Deliver Project. 3.6.2. Population and housing census data The study made use of two levels of 2010 PHC data for GAR namely a micro-data sample and aggregated district level data which were obtained from the GSS. The micro-data sample consist of information at the individual respondent and household levels. The variables contained in this national census include sex, age, date of birth, nationality, ethnicity, birthplace, religion, internal migration, marital status, literacy, fertility, mortality, economic activity, occupation, industry, employment status, employment sector, disability, use of ICT, 75 University of Ghana http://ugspace.ug.edu.gh agricultural activity, dwelling type, income, education level and several dozen data items. With the micro data, it was possible to aggregate variables considered in this study to Enumeration Areas (EAs) and further aggregated to health facility catchment areas. However, records for only a random 10% of the enumerated households were released as micro-data’ therefore the count of all variables extracted had to be multiplied by 10 to arrive at the total count per unit/EA. EAs are the operational geographic units or blocks for the collection of the census data. The total number of EAs in GAR according to the 2010 PHC was 5,425. The average size of an EA based on the 2010 PHC figures in GAR was 739 persons, whilst the least EA size was 22 persons and the maximum EA size was 9,880 persons. The aggregated district level data on the other hand consisted of projected 2014 population distribution by districts, sex and 5-year age groups from the 2010 PHC data obtained from Ghana Statistical Service. The average population size for each of the 16 districts in GAR was 275,131 with the least populated district being Shai-Osudoku with a population size of 56,089 and AMA with the highest population size of 1,841,091. 3.6.3. Household Survey data The GDHS of 2014 is the sixth in a series of national-level population and health surveys conducted in Ghana as part of the global Demographic and Health Surveys (DHS) programme. The field work component of the 2014 GDHS started on 12th September 2014 and ended on 25th December 2014. The survey obtained detailed information on fertility, marriage, sexual activity, fertility preferences, awareness and use of family planning methods, breastfeeding practices, nutritional status of women and young children, childhood mortality, maternal and child health, awareness and behaviour regarding HIV/AIDS, and other sexually transmitted infections (STIs). In addition, 2014 GDHS like the 2008 collected 76 University of Ghana http://ugspace.ug.edu.gh information on domestic violence, malaria and use of mosquito nets, and carried out anaemia testing and anthropometric measurements for women and children (GDHS, 2015). The GDHS data of interest in this study were the under-5years old children: who were taken to a medical facility with fever and/or cough and had received any anti-malaria treatment – malaria cases or who were taken to a medical facility for diarrhoea and had received any diarrhoea drug – diarrhoea cases. In both the women and children questionnaires of 2014 GDHS, mothers were asked if their children under 5 years old had experienced an episode of fever and diarrhoea in the two weeks preceding the survey and if so, whether treatment or advice was sought and where the treatment was sought for, and whether they have received any anti-malarial or diarrhoea drug. For example, among children under age 5, the percentage who had a fever in the two weeks preceding the survey and among children with fever, the percentage for whom advice or treatment was sought from a health facility or provider, the percentage who took anti-malarial medicines as treatment, by background characteristics extracted from 2014 GDHS were compared with what is contained in 2014 GDHS report for confirmation. 3.6.4. Meteorological Data Meteorological data was obtained from Meteorological Agency of Ghana. The types of data collected include monthly rainfall, maximum temperatures and minimum temperatures for the year 2014. In all, the data was collected from 5 stations namely Accra Airport, Accra Mpehuasem, Ada, Tema and Pokuase. 77 University of Ghana http://ugspace.ug.edu.gh 3.6.5. Administrative boundary Data Administrative boundary data for this study was obtained from RS/GIS Laboratory of Geography Department, University of Ghana and GSS. Administrative boundary data obtained included GAR base map and EAs map of GAR. The GAR base map consist of shapefiles of GAR boundary, district boundaries, towns and river bodies. The EAs map contains the shapefile of the delineated EAs of GAR. Out of a total of 5,423 enumeration areas in GAR during the 2010 PHC, 2,916 (54%) have been delineated as individual EAs and the remainder are as groups of EAs. Thus, some districts have their individual EAs either partially delineated or not delineated at all and a typical example is Ashaiman municipality with all its 260 EAs not delineated. Only 3 districts have complete delineation of individual EAs which include Accra Metropolitan Area, La Dade-Kotopon and Ledzokuku-Krowor municipalities, out of the 16 districts in the GAR. These delineated EAs in the above three mentioned districts constituted 63% of all the individual delineated EAs in the region. 3.7. Data Processing 3.7.1. Health facility visits data The data used in this study consist of observed HFV due to malaria and diarrhoea. Data was extracted from DHIMS2 database as monthly statistics for each facility by districts and imputed data for non-reporting facilities. However, for the sake of simplicity both the observed and imputed data together would be referred to in this document interchangeably as outpatient data or DHIMS2 data or HFV data or facility-based data. Each of the diseases has been re-grouped by facilities, sex and age groups for the entire year per district. Most of the age groupings of the original data were maintained except under-1year, 0-5years and 15-19 78 University of Ghana http://ugspace.ug.edu.gh years’ age groups which were created by regrouping some age groups to enable linking the age groupings in the DHS and the census data. The under-1year age group consist of under- 28 days and under-1 year, under-5 years age group consist of under-28 days, under-1 year and 1-4 years’ age group of the original dataset and age groups 15-17years and 18-19years were merged to give 15-19 years’ age group. Data has also been grouped into months i.e. from January to December at district and regional levels by facilities. It was necessary to account for gaps in the facility-based data to enable comparison with GDHS data. There were two issues related to missing data concerning HFV due to malaria and diarrhoea obtained from the DHIMS2 database. The first has to do with facilities that did not report data for any of the 12 months in 2014 on the two diseases under consideration. In all 262 facilities were identified and their missing annual cumulative HFV were imputed, using the marginal long multiple imputation technique in Stata (see details of this technique in section 3.6.1). However, in processing the data for imputation, using malaria as an example, the observed HFV for malaria and all other covariates were arranged in a long form in excel spreadsheet. The data was later exported to Stata for imputation. Twenty possible values were imputed for each missing data and the datasets were exported back to an Excel spreadsheet for further processing. The missing values were finally determined by calculating the mean number of HFV for each health facility from the 20 different data sets imputed. The observed and imputed data were then combined to constitute the full or complete DHIMS2 dataset. The second type of missing data concerned those facilities that do report, but record zero outpatient cases for some of the months. The difficulty here was that there was no code for a missing monthly record, so it was unclear whether a zero meant no HFV or a missing record. Where a zero value was reported for a month for malaria or diarrhoea, the number of reported outpatient visits for other diseases were examined. If there were records for other diseases it 79 University of Ghana http://ugspace.ug.edu.gh was assumed that the facility has reported data for the month and the zero is considered as no malaria or diarrhoea case (s) and if otherwise it was considered as missing data. The number of months for each of the diseases where zeros were considered as missing data were less than 1% and therefore were ignored. 3.7.2. Population census data Independent variables used in this study, were extracted from the 10% micro-data sample of 2010 PHC data. A Series of steps were taken to extract the variables for the district and catchment areas. First and foremost, with Stata command “keep if’ all other regions were dropped from the national data, keeping only GAR. Following from that, variables that are not relevant to this study were also dropped from the database through the same procedure. To extract the variables by EAs, unique 10-digit codes were generated for each EA. The code was generated using GSS’s procedure i.e. by joining region code, district code, sub-district code and EA number. A typical example of such code is EA number one (0306200001) located in Ledzokuku-Krowor municipality of GAR. The Stata command “concatenate” was used to join the components of the code together. The EA codes generated were then used to extract the variables. The variables extracted from the 2010 Housing and Population Census data include population counts and all the independent variables (see section 3.6.5). These EA level variables, after the extraction from the micro-data were multiplied by 10 to represent the entire population rather than a 10% sample. The EA-level totals from the micro- data were joined to EA attributes attached to the map layer by linking on EA codes. 80 University of Ghana http://ugspace.ug.edu.gh 3.7.3. Ghana Demographic and Health Survey Data Many steps have been taken to extract the data for children under five years of age whose parents claimed they had experienced fever/malaria and diarrhoea in the last two weeks preceding the survey and had sought treatment or advice from a health facility excluding pharmacies from GDHS data. In the case of malaria, the key variable used was H32Z which was defined as “Whether the child was taken to a medical facility for treatment of the fever and/or cough”. The medical facilities referred to by this variable were all public-sector facilities and all medical private sector facilities except for pharmacy. This variable was used in the final reports to summarise prevalence and treatment of fever (GDHS, 2015). Since fever is a symptom of many diseases, children who sought treatment or advice from the medical facility were further screened by the kind of medicines received. Using the kind of medication or treatment given to patients confirmed whether the child had been clinically diagnosed with malaria or not, and only those who received any anti-malarial drugs (i.e. fansidar, chloroquine, quinine, combination with artemisinin, artemisinin or other anti- malarial drugs) were considered malaria cases. In finding the number of children under-five diarrhoea cases from the GDHS data, the key variable used was H12Z which was defined as “Whether the child was taken to a medical facility for treatment of the diarrhoea”. The medical facilities referred to by this variable were all public-sector facilities and all medical private sector facilities except for pharmacy. This variable was also used in the final reports to summarise treatment of diarrhoea (GDHS, 2015). Medication or treatment given to patients who sought medical treatment or advice from a health facility was used as an indicator to confirm whether the patients was clinically diagnosed with diarrhoea or not. Only those who received the main or basic diarrhoea treatment drugs (i.e. oral rehydration or recommended home solution) were clinically diagnosed of having diarrhoea. 81 University of Ghana http://ugspace.ug.edu.gh 3.7.4. Linking of HFV to district -level population census data. The locations of the 513 health facilities (public and private), with known geographic coordinates, have been geo-coded. The geocoding process involved transforming the health facility location as a pair of coordinates to a location on the earth's surface using ArcGIS. The health facility coordinates were first entered in an excel table and uploaded onto ArcGIS software. The resulting locations were output as geographic features with attributes and were spatially linked to the GAR district boundaries using ArcGIS spatial join functionality. At the district level, the total HFV due to malaria and diarrhoea were arrived at by summing the number of visits recorded by months together from each health facility in the district for the year. Population data for the districts was obtained directly from GSS. 3.7.5. Creating Thiessen polygon catchment areas Routine health facility-based data in Ghana contains information mainly on disease characteristics and some basic information on facility location and type. Information such as patient-use and capacity related information are absent. In such a context, Thiessen polygons (otherwise known as Voronoi polygon) are recommended as an intuitive and most appropriate approach to the delineation of facility catchment areas (Black et al., 2004; Gething et al., 2004). Literature has shown several instances where Thiessen polygons were used to define catchment areas for small area analysis (Judge et al., 2009b; Martin & Williams, 1992). The label points for the Thiessen polygons in this study were health facility locations. The polygons were defined for all the health facilities in GAR. The Thiessen polygon concept have a unique property to apportion space to each polygon such that all points in that space 82 University of Ghana http://ugspace.ug.edu.gh are closest to that reference health facility than any other facility in other polygons (ESRI, 2017a; Tatalovich, 2005; Wilkinson & Tanser, 1999). “The theoretical background for creating Thiessen polygons is as follows: • Where S is a set of points in coordinate or Euclidean space (x, y), for any point p in that space, there is one point of S closest to p, except where point p is equidistant to two or more points of S. • A single proximal polygon (Voronoi cell) is defined by all points p closest to a single point in S, that is, the total area in which all points p are closer to a given point in S than to any other point in S.”(ESRI, 2017a). “Thiessen proximal polygons are constructed as follows: • All points are triangulated into a triangulated irregular network (TIN) that meets the Delaunay criterion. • The perpendicular bisectors for each triangle edge are generated, forming the edges of the Thiessen polygons. The location at which the bisectors intersect determine the locations of the Thiessen polygon vertices.” (ESRI, 2011) This method assumes that patients will travel to the facility that is closest in Euclidean space. 3.7.6. Linking of population census data to health facility catchment areas. The EA polygons were first converted to points forming a map layer of EA centroids. This map layer of census EA centroids was overlaid on a layer of health facility catchment areas and the two layers were spatially joined to determine the EA centroids that fall within each catchment area. The population size of a catchment area was determined by aggregating all the population counts of the EA centroids that fall within each catchment polygon. The 83 University of Ghana http://ugspace.ug.edu.gh known risk factors of malaria and diarrhoea extracted from the micro-data are also joined to the EA attributes attached to the facility catchment map layer by linking them by EA codes. The limitation to the use of the centroid method to determine the size/counts of variables for a catchment must do with the fact that some of the original EA boundaries close to the edges of the catchments polygons have overlapped the catchment boundaries, hence creating mismatches between EA and catchment boundaries. Whilst this will add uncertainty to individual catchment populations, it is expected that any under or over-estimation that may occur would be minimal, due to overall cumulative effect of positive and negative deviations. 3.7.7. Data flow diagram Figure 3.2 is a flow diagram that shows the graphical representation of the flow of data through the different stages of the analysis process. Data types represented by this diagram included population size (referred to as population at risk), health facility information, HFV (malaria was used as an example) and 2014 GDHS under-5 years old children HFV due to malaria information. With the current GAR EA map layer, 48.5% of the total number of EAs (5,425) were not delineated as single EA polygons, they exist as a clump of EAs ranging from two EAs to as much as 260 EAs per clump. Thirteen (13) of these clumped EAs which consisted of 50 to 260 EAs per clump were dropped because of their share sizes which were likely to inflate the population of the catchments in which they may fall. The total number of individual EAs dropped (as part of the 13 clumps of EAs) amounted to 721 EAs which had a total population of 695,457(see Figure 3.2), representing approximately 16% of the projected 2014 population of GAR. 84 University of Ghana http://ugspace.ug.edu.gh HF: health facility, HFV: health facility visit, Pop – population size Figure 3.2: Data flow diagram showing GAR population at risk, HFV due to malaria and number and type of HF at different levels of the analysis. Health facility catchment areas with no attending population as a result of either the clumping together of EAs as explained above and or that no EA centroids fell in them were also excluded from the analysis. The number of such facility catchment areas identified were 52. Similarly, 120 facility catchment areas with no record of malaria or diarrhoea risk factors were also excluded from the regression analysis. In all analyses, facility catchment areas with coordinates but excluded from the analysis due to any of the reasons above have been represented by blank spaces. Most of these blank catchemt areas were located in AMA, Ashaiman, TMA, Ga South, and La Nkwantanang-Madina districts. 85 University of Ghana http://ugspace.ug.edu.gh 3.8. Data Analysis Analyses for this study involved both geospatial and statistical approaches. The statistical analysis included frequency distributions, ratios, proportions, scatter plots and OLS regression. One Way ANOVA test at 95% confidence level was used to test the statistical significance of sex and age group differences observed in the distribution of annual cumulative HFV/HFVR for malaria and diarrhoea. The geospatial analysis involved visualization, exploration and modelling of spatial data. Analyses were carried out at the regional, district and catchment levels to show geographical differences in the results. This study used OLS and GWR models to analyse spatial relationships between HFVR for malaria and diarrhoea annual cumulative rate (dependent variable) and independent variables (environmental / household factors extracted from 2010 PHC data). Microsoft Excel 2016, Stata version 14.0 and ArcGIS 10.4.2 were the software used for the analyses. 3.8.1. Determining the level of completeness in reporting into DHIMS2 The determination of the level of completeness in facility reporting into DHIMS2 for Greater Accra Region involves two components. These include completeness of facility reporting and estimating the number of visits for non-reporting facilities. All health facilities ( public, private, non-governmental, faith, etc) are expected to report data on key service outputs on monthly basis in to DHIMS2. If facility reporting completeness is less than 100%, this implies that there is incomplete information on health indicators. Hence, it is very crucial to know the facility reporting completeness rate to make informed interpretation of the indicators (WHO, 2014). Completeness of facility reporting (%), defined by (WHO, 2014) as “the number of reports received, according to schedule, from all health facilities nationally, divided by the total expected reports from all facilities that are supposed 86 University of Ghana http://ugspace.ug.edu.gh to report to the RHIS for a specified time period (usually one year)”. This can be expressed mathematically as: 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑓𝑎𝑐𝑖𝑙𝑖𝑡𝑖𝑒𝑠 𝑡ℎ𝑎𝑡 𝑠𝑢𝑏𝑚𝑖𝑡 𝑎 𝑟𝑒𝑝𝑜𝑟𝑡 Completeness of facility reporting = x 𝑡𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑓𝑎𝑐𝑖𝑙𝑖𝑡𝑖𝑒𝑠 𝑡ℎ𝑎𝑡 𝑎𝑟𝑒 𝑒𝑥𝑝𝑒𝑐𝑡𝑒𝑑 𝑡𝑜 𝑠𝑢𝑏𝑚𝑖𝑡 𝑎 𝑟𝑒𝑝𝑜𝑟𝑡 100 % (3a) At the district level, a facility reporting completeness rate is computed using actual and expected reports. Hence, 𝐴𝑐𝑡𝑢𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑟𝑒𝑝𝑜𝑟𝑡𝑠 𝑟𝑒𝑐𝑒𝑖𝑣𝑒𝑑 𝑖𝑛 12 𝑚𝑜𝑛𝑡ℎ𝑠 Facility reporting rate of a district = x 100 % (3b) 𝐸𝑥𝑝𝑒𝑐𝑡𝑒𝑑 𝑟𝑒𝑝𝑜𝑟𝑡𝑠 (𝑇𝑜𝑡𝑎𝑙 𝑓𝑎𝑐𝑖𝑙𝑖𝑡𝑖𝑒𝑠 𝑥 12 𝑚𝑜𝑛𝑡ℎ𝑠) In estimating the number of HFV for non-reporting facilities, the HFV were treated as missing. The non-reporting health facilities were first identified and their attributes listed. In this study, 262 non-reporting health facilities were identified and multiple imputation by chain equation (MICE) method (Buuren & Groothuis-Oudshoorn, 2011; Raghunathan et al., 2001) was used to estimate the missing number of HFV. Multiple Imputation by Chained Equation (MICE) MICE was used to impute missing values for the annual cummulative HFV due to malaria for each non-reporting health facility in Greater Accra region in 2014. Imputation of the data was necessary because a complete HFV data of Greater Accra region was required to enable comparison of routine facility-based data with the GDHS data and for tha calculation of population-based rates for small area analysis. Compared to ordinary imputation techniques such as single mean imputation and regression imputation, MICE approximate the true figure and reduces bias in the estimation. All other multiple imputation methods such as Fully Bayesian, Maximum Likely Estimation and Substantive Model Compactible Fully Conditional Specification (SMCFCS) will give similar results. Since HFV due to malaria and diarrhoea are count outcome measures, Poisson Conditional imputation model was evaluated 87 University of Ghana http://ugspace.ug.edu.gh by generating missing indicator variables where 1 represented missing value and 0 otherwise. Binary logistic regression models were fitted with the fully observed covariates. The covariates used in this statistical technique included: district type, facility ownership, facility type, National Health Insurance Scheme (NHIS) membership, population size of district, population size by sex, age group, month of the year and observed annual cumulative HFV due to malaria or diarrhoea by reporting facilities. Most of the variables were found to be related to binary indicator of missingness, an indicative of the fact that Missing At Random (MAR) is plausible. MAR assauption is applicable if the propensity for a data point to be missing is not related to the missing data, but it is related to some of the observed data (Schafer & Graham, 2002). Based on the Markov chain error estimates this study used 20 imputations. The results were combined using Rubins rule (Rubin, 1987) which involves averaging of the 20 complete-data estimates (in the case of this study) 3.8.2. Comparison of facility-based DHIMS2 and GDHS data To compare GDHS to DHIMS2 (routine facility-based) data, treatment information was collected on only children with fever and diarrhoea for whom advice or treatment was sought from a health facility (public or private) not pharmacies, drug stores and other sources. And the child was treated and received anti-malaria drug or diarrhoea drug. The 2014 GDHS data were self-reported (by mothers of children under 5-years who were part of the sample selected) malaria and diarrhoea cases of children under 5 years who had malaria or diarrhoea within the 2 weeks preceding the survey and sought advice or treatment in a health facility (both public and private). Based on this background information about the GDHS, HFV due to malaria and diarrhoea (observed and imputed) were extracted for the period September to December 2014. Under each month the data for children <5 years was extracted and 88 University of Ghana http://ugspace.ug.edu.gh categorized into age groups (< 1year and 1-4 years) and finally into sex categories (male and female). The sum of cases under each category at the final stage of grouping was divided by 8 (there were 4 months involved amounting to 16 weeks and dividing by 8 gives 2 weeks average) to arrive at the 2 weeks average which can be compared with its corresponding 2014 DHS report. To compute for the expected number of children who received anti-malaria/diarrhoea treatment in a medical facility at the population level in 2014, these procedures were employed: 1. The proportion of sample who received anti-malaria treatment in a medical facility as well as its 95% confidence interval were calculated using these formulae No. of sample of children who received anti − malaria treatment from a medical facility 𝑝 = 𝑇𝑜𝑡𝑎𝑙 𝑠𝑎𝑚𝑝𝑙𝑒 (3c) 2. 95% Confidence Limit (CL): 𝑝(1−𝑝) 𝑝 ± 𝑧𝛼/2 ( √ ) (3d) 𝑛 Where, • 𝑝 = proportion of sample who received anti-malaria treatment in a medical facility; • n = total sample size; • α =significance level = 5%; • 𝑧𝛼/2 = 𝑧0.05/2 = 𝑧 − 𝑠𝑐𝑜𝑟𝑒 𝑓𝑟𝑜𝑚 𝑡ℎ𝑒 𝑠𝑡𝑎𝑛𝑑𝑎𝑟𝑑 𝑛𝑜𝑟𝑚𝑎𝑙 𝑑𝑖𝑠𝑡𝑟𝑖𝑏𝑢𝑡𝑖𝑜𝑛 = 1.96. 3. The expected number of children who received anti-malaria treatment in a medical facility at the population level was then estimated by multiplying the proportion of 89 University of Ghana http://ugspace.ug.edu.gh sample who received anti-malaria treatment in a medical facility (𝑝) - by the 2014 population (projected). 4. The lower and upper limits of the expected numbers were also estimated by multiplying the lower or upper confidence limit by the projected 2014 population size. External consistency ratio and other descriptive statistics tools such as ratios, percentages, and proportions were used to examine the level of agreement between facility-based DHIMS2 HFV due to malaria and diarrhoea (this included both observed and imputed visits) and self-reported malaria and diarrhoea HFV obtained from GDHS dataset of 2014. The external consistency ratio was calculated as the population coverage of say <5 years old DHIMS HFV due to malaria divided by the population coverage of the self-reported <5 HFV due to malaria based on GDHS data. The ratio of say HFV due to malaria is expressed mathematically as: 𝑁𝑜. 𝑜𝑓 <5 𝑦𝑒𝑎𝑟𝑠 𝑜𝑙𝑑 𝐻𝐹𝑉 𝑑𝑢𝑒 𝑡𝑜 𝑚𝑎𝑙𝑎𝑟𝑖𝑎 𝑤𝑖𝑡ℎ𝑖𝑛 2 𝑤𝑒𝑒𝑘𝑠 𝑝𝑒𝑟𝑖𝑜𝑑 𝑓𝑟𝑜𝑚 𝐷𝐻𝐼𝑀𝑆2 Ratio= 𝑁𝑜.𝑜𝑓 𝑒𝑥𝑝𝑒𝑐𝑡𝑒𝑑<5 𝑦𝑒𝑎𝑟𝑠 𝑜𝑙𝑑 𝐻𝐹𝑉 𝑑𝑢𝑒 𝑡𝑜 𝑚𝑎𝑙𝑎𝑟𝑖𝑎 𝑤𝑖𝑡ℎ𝑖𝑛 2 𝑤𝑒𝑒𝑘𝑠 𝑝𝑟𝑒𝑐𝑒𝑒𝑑𝑖𝑛𝑔 𝑡ℎ𝑒 𝑠𝑢𝑟𝑣𝑒𝑦 𝑓𝑟𝑜𝑚 𝐺𝐷𝐻𝑆 𝐷𝐻𝐼𝑀𝑆2 <5𝑦𝑒𝑎𝑟𝑠 𝑜𝑙𝑑 𝐻𝐹𝑉 𝑑𝑢𝑒 𝑡𝑜 𝑚𝑎𝑙𝑎𝑟𝑖𝑎 = (3e) 𝐺𝐷𝐻𝑆<5𝑦𝑒𝑎𝑟 𝑜𝑙𝑑 𝐻𝐹𝑉 𝑑𝑢𝑒 𝑡𝑜 𝑚𝑎𝑙𝑎𝑟𝑖𝑎 The ratio gives an idea of how consistent the <5 years old malaria or diarrhoea estimated from facility-based reports is to the number of HFV obtained from self-reported GDHS data within an average of two weeks between September and December. The closer this ratio is to 1 (or 100%), the higher the consistency. If the ratio is 1, it means that the two coverage counts or attendance rates are the same. If the ratio is >1, it means that the facility-based coverage is higher than the GDHS coverage counts. If the ratio is <1, it means that the GDHS coverage count is higher than the facility-based coverage counts. Then, the percentage difference between the DHIMS2 and GDHS was calculated as: |𝐷𝐻𝐼𝑀𝑆2 𝐻𝐹𝑉−𝐺𝐷𝐻𝑆 𝐻𝐹𝑉| % Difference = x 100 % (3f) 𝐺𝐷𝐻𝑆 𝐻𝐹𝑉 90 University of Ghana http://ugspace.ug.edu.gh This ratio was recommended by the WHO framework for internal and external consistency checks for aggregation units such as province/state/region. The document further stated that if the ratio of facility-based coverage count/rates to survey coverage counts/rates differ by at least 33%, the ratio is deemed to be inconsistent. In other words, at least 33% difference between ratio of facility-based coverage to GDHS coverage is a maker of inconsistency between the two datasets (WHO, 2014). Graphs and tables were used to display the result of the analysis. The software used was MS Excel 2016 (Microsoft Corporation, Redmond, USA) 3.8.3. Measuring health differentials using routine health facility-based data The descriptive analysis of HFVR for malaria and diarrhoea was performed at district and catchment levels in GAR to determine the geographical variations in HFV. Annual cumulative HFVR for malaria or diarrhoea, was defined as the total number of HFV due to malaria or diarrhoea divided by the size of the population of the analysis unit (district or catchment area), expressed per 1,000 population. For example, HFVR for malaria can be expressed mathematically as: 𝑡𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 ℎ𝑒𝑎𝑙𝑡ℎ 𝑓𝑎𝑐𝑖𝑙𝑖𝑡𝑦 𝑣𝑖𝑠𝑖𝑡𝑠(𝐻𝐹𝑉) 𝑑𝑢𝑒 𝑡𝑜 𝑚𝑎𝑙𝑎𝑟𝑖𝑎 𝐻𝐹𝑉𝑅𝑚𝑎𝑙𝑎𝑟𝑖𝑎 = x 1,000 (3g) 𝑠𝑖𝑧𝑒 𝑜𝑓 𝑡ℎ𝑒 𝑝𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛 𝑜𝑓 𝑡ℎ𝑒 𝑎𝑛𝑎𝑙𝑦𝑠𝑖𝑠 𝑢𝑛𝑖𝑡(𝑖.𝑒.𝑐𝑎𝑡𝑐𝑚𝑒𝑛𝑡 𝑜𝑟 𝑑𝑖𝑑𝑡𝑟𝑖𝑐𝑡) The annual cumulative rates for the district level analysis have been age-standardized to remove the effects of the different frequency distributions among the populations using a direct standardization method. A direct standardization method (Naing, 2000) is applied when age-specific morbidity rates for two or more populations are known. Procedure for direct standardization include calculating the age-specific morbidity rates for each age group in each district. Since the HFVR of the districts are compared to the regional population, the regional population was 91 University of Ghana http://ugspace.ug.edu.gh considered as a “standard” population. The age-specific HFVR of the districts were multiplied by the number of persons in each age group of the standard population. In this way, the expected HFV for each age group of each district was determined. The number of expected HFV from all age groups were added for each district. The age-adjusted morbidity rates were calculated by dividing the total number of expected HFV of each district by the standard population. This allows for comparing the age-standardized morbidity rates of two or more districts (Naing, 2000). The observed age-specific sex differences in HFV due to malaria or diarrhoea were tested for statistical significance using Welch t-test. One-way ANOVA test was conducted to determine the statistical significance of the observed HFVR for malaria and diarrhoea among the districts in GAR in 2014. All the tests were performed at 95% Confidence Interval. Average monthly rainfall and temperature for the region were also calculated from records of five stations in the region and graphed with malaria and diarrhoea case rates to examine the plausibility of seasonal patterns in reported HFV. The results were presented using frequency distribution tables, charts and graphs. 3.8.4. Visualising spatial variations in routine facility-based data (HFV due to malaria and diarrhoea Geospatial and statistical analyses techniques were used in this section. Geo-visualization techniques were used to show spatial variations in HFV due to malaria and diarrhoea in the districts and facility catchment areas in GAR. The techniques used included graduated symbol, choropleth and Hot Spot Analysis (Getis-Ord Gi*). 92 University of Ghana http://ugspace.ug.edu.gh 3.8.4.1. Graduated symbol map Graduated symbols were used to efficiently map the different counts of HFV due to malaria and diarrhoea and superimposed on a population density map to show if there were any association between them. Graduated symbol maps can reveal the relative differences in disease counts among the various districts and one can determine a pattern across the map. Size was used as the ordering visual variable where larger symbols means “more” HFV and vice versa. The data is usually classified using a scheme that reflects the data distribution. In this study, the data was classified into 5 classes using a manual scheme that establishes class breaks at rounded values adequate to ensure there was a good degree of symbol variation across the map. The disadvantages of graduated symbol maps are that it is difficult to calculate actual values if they are not shown and the size of the symbols may obscure locations or mean less accurate positioning on the maps. 3.8.4.2. Choropleth map A choropleth map (sometimes referred to as an area-value map) was used to reveal patterns in annual cumulative HFV due to malaria and diarrhoea among districts and catchment areas of GAR. The maps displaying divided geographical areas of the region are coloured or shaded in relation to size of the rate. The choropleth maps used differences in colour value to represent differences in the annual cumulative HFVR mapped. Usually, the darker colour represents the larger rate and the lighter colour represents smaller rate. Generally, choropleth maps provide a way to visualise values over a geographical area, which can show variation or patterns across the displayed location. The key limitation to this method is the Modifiable Areal Unit Problem (MAUP) (Openshaw, 1984) due to the arbitrary nature of the boundaries 93 University of Ghana http://ugspace.ug.edu.gh of the catchments. MAUP may arise when the boundaries of the catchments are drawn differently to produce different spatial patterns (see details of MAUP in section 5.3.3). 3.8.4.3. Hot Spot Analysis (Getis-Ord Gi*) Hot spot locations are places where observed patterns represent places where there are underlying spatial processes at work (Getis & Ord, 1992). This statistic calculates for each feature in the dataset by considering each feature within the context of neighbouring features. The resulting Z - score indicates the location where features (e.g. 𝐻𝐹𝑉𝑅𝑚𝑎𝑙𝑎𝑟𝑖𝑎) with either high or low values were spatially clustered. For a feature (e.g. 𝐻𝐹𝑉𝑅𝑚𝑎𝑙𝑎𝑟𝑖𝑎) to be statistically significant hot spot, it will have a high value and be surrounded by other features with high values as well. The process involves proportional comparison of the local sum for a feature and its neighbours to the sum of all features. Where the local sum is much different than the expected local sum, and that difference is too large to be the result of random chance, a statistically significant Z-score results and the 𝐺∗𝑖 statistic is the z-score. The Getis-Ord local statistic 𝐺∗𝑖 is given as: 𝑛 ∗ ∑𝑗=1 𝑤𝑖,𝑗𝑋𝑗− ?̅? ∑ 𝑛 𝐺 𝑗=1 𝑤𝑖,𝑗 𝑖 = (3h) 𝑆 𝑛 ∑𝑛 𝑤2 −(∑𝑛 2 √ 𝑗=1 𝑖,𝑗 𝑗=1 𝑤𝑖,𝑗) 𝑛−1 Where 𝑥𝑗 is the attribute value for feature j, 𝑤𝑖,𝑗 is the spatial weight between feature i and j, n is equal to the total number of features and the mean of the attribute is given by ∑𝑛 𝑥𝑗 ?̅? 𝑗=1 = 𝑛 94 University of Ghana http://ugspace.ug.edu.gh and standard deviation ∑𝑛 2 √ 𝑗=1 𝑥𝑗 S = − (?̅?)2 𝑛 If the calculated index values are greater than a threshold associated with statistical significance (5% in this study), the location of a cluster is identified as a hot spot. Therefore, facility catchment that are nearby or that encompass such a cluster are identified as hot spots. Conceptualization of spatial relationships was based on inverse distance since several authors have shown that health facility attendance is one of impedance or distance decay - closer features are weighed more heavily than features that are further away (Kadobera et al., 2012; Müller et al., 1998; Stock, 1983). That is, patients tend to attend facilities closer to them than the one farther away. Inverse Manhattan distance was used as it works best when analysis involves locations with fixed facilities such as health facilities and when road network data is unavailable (ArcGIS_Pro, 2017). 3.8.5. Analysing spatial relationship between HFVR for malaria or diarrhoea and household factors OLS and GWR models (Lin & Wen, 2011) were used to analyse spatial relationships between dependent (annual cumulative HFVR for malaria and diarrhoea) and independent variables (environmental / household factors) and to identify the existence of any spatial heterogeneity in these relationships. Moran’s I (Index) was used to measures spatial autocorrelation. 95 University of Ghana http://ugspace.ug.edu.gh 3.8.5.1. Study variables Dependent variable The dependent variables used in both OLS and GWR models were the annual cumulative HFVR for malaria and diarrhoea (crude) calculated for each catchment area.. Independent variables Two types of independent variables were considered for each condition: those reflecting risk factors, and those reflecting data system effects arising from the calculation of the rates. The risk factors were extracted from the 2010 PHC data based on available literature. The proportions were arrived at by summing up the number of persons having the characteristic under consideration in a catchment area divided by the total population of the catchment area and multiplying by 100. The independent variables used for both malaria and diarrhoea models are listed in Table 3.1. Data system effect variable ( health facility types) was added to the models to adjust for the effect of health facility capacity or attractiveness. e.g. hospitals, because of the kinds of services they provide and availability of qualified health professionals attract more patients than other types of facilities with lower capacities. 96 University of Ghana http://ugspace.ug.edu.gh Table 3.1: Independent variables of HFVR for malaria and diarrhoea Independent variables Malaria Risk factors: Proportion of persons living in walls prone to malaria Pwalpma Proportion of persons within low educational level Plowedul Proportion of persons within the lowest wealth quintile Plowweaq Proportion of persons in dwellings prone to malaria Pdwelpma Proportion of persons in houses roofed with materials prone to malaria Proofpma Proportion of persons in houses with floor prone to malaria Pflorpma Proportion of persons engaged in agriculture activities PagricAc Data system factors Facility type - CHPS Ftyp_Chps Facility type - Clinic/Health centre Ftyp_C/HC Facility type - Polyclinic Ftyp_Poly Facility type - Hospital Ftyp_Hosp Diarrhoea Risk factors: Proportion of persons whose toilet facility sharing with others prone to diarrhoea Ptoishpdiar Proportion of persons whose main source of water for domestic use prone to diarrhoea Pdowatpdiar Proportion of persons whose method of rubbish disposal prone to diarrhoea Prudispdiar Proportion of persons whose main source of drinking water prone to diarrhoea Pdrwatpdiar Proportion of persons using toilet facility prone to diarrhoea Ptoipdiar Proportion of persons whose method of liquid waste disposal is prone to diarrhoea Plwasdispdiar Proportion of persons within the lowest wealth quintile Plowweaq Proportion of persons within low educational level Plowedu Data system factors Facility type - CHPS Ftyp_Chps Facility type - Clinic/Health centre Ftyp_C/HC Facility type - Polyclinic Ftyp_Poly Facility type - Hospital Ftyp_Hosp Except for proportion of persons engaged in agriculture activities All the other independent variables (i.e. malaria and diarrhoea predictors) were derived from a group of factors or household information about the population. Table 3.2 provided the list of the household components from which the variables were derived from 97 University of Ghana http://ugspace.ug.edu.gh Table 3.2: Malaria and diarrhoea predictors and household factors Predictors Component household factors Malaria Proportion of persons living in walls Persons living in walls made of: prone to malaria • mud brick/earth • wood • palm leaf/thatch (grass)/raffia • bamboo Proportion of persons within low Education levels: educational level • nursery • kindergarten • primary Proportion of persons in dwellings Persons living in: prone to malaria • huts/buildings (same compound) huts/buildings (different compound) • tent, • improvised home (kiosk/container etc) • uncompleted building Proportion of persons in houses roofed Persons living in: houses roofed with: with materials prone to malaria • mud/mud bricks/earth • wood • bamboo • thatch/palm leaf • raffia Proportion of persons in houses with persons living in houses with floor made of: floor prone to malaria • earth/mud • wood Diarrhoea Proportion of persons whose toilet Persons sharing toilet facility with: facility sharing with others prone to • other households, diarrhoea • other households in different house • other households and locations Proportion of persons whose main Persons whose main source of water for other domestic use source of water for domestic use prone include: to diarrhoea • bore-hole/pump/tube • well • protected well • rain water • tanker supply/vendor provided, • unprotected well • unprotected spring • river • stream • dugout/pond /lake dam/canal and other Proportion of persons whose method of Persons whose method of rubbish disposal by household include: rubbish disposal prone to diarrhoea • burned by household • public dump (container) • public dump (open space) • dump indiscriminately • buried by household and others; Proportion of persons whose main Persons whose main source of drinking water include: source of drinking water prone to • bore-hole/pump/tube well diarrhoea • protected well • rain water 98 University of Ghana http://ugspace.ug.edu.gh • tanker supply/vendor provided • unprotected well • unprotected spring • river/stream • dugout/pond/lake/dam/canal and others Proportion of persons using toilet Persons using: facility prone to diarrhoea • no facility (bush/beach/field) • pit latrine • bucket (pan) latrine • public toilet (WC, KVIP, pit latrine, pan etc.) Proportion of persons whose method of Persons whose method of liquid waste disposal by household liquid waste disposal is prone to include: diarrhoea • through drainage system into a gutter • through drainage into a pit (soak away) • thrown onto the street/outside • thrown into gutter • thrown onto compound and others Health facility type • CHIPS • Clinic/Health Centre • Polyclinics • Hospitals Principal Componenet Analysis (PCA) The proportion of persons within the lowest wealth quintile (Plowweaq) was computed using Principal Component Analysis (PCA). This was calculated from socio-economic index using variables from the PHC data. It is widely accepted that measures of household wealth can be reflected by income or expenditure information (Vyas & Kumaranayake, 2006). Thus, PCA was applied using asset data from 2010 PHC data in creating socio-economic status (SES) indices named level of wealth quintile. To compute the SES, household assets such as toilet facility, piped water, finished house roof, finished house floor, finished house walls, electricity, TV ownership, bike ownership and car ownership were used. The assets/variables used may be correlated in multiple dimensions and PCA helps to reduce those dimensions by assessing similarities between these variables. New sets of variables called principal components were created to capture variation across.the variables.. The first component is typically used because of the assumption of accounting for the largest possible variation between the assets/variables used. Subsequent components account for as much of the remaining variance as possible in a reducing manner with all components not linearly 99 University of Ghana http://ugspace.ug.edu.gh correlated to preceding variables . This can be expressed mathematically say from a s set of variables 𝑋1 to 𝑋𝑛 as: 𝑃𝐶1 = 𝑎11𝑋1 + 𝑎12𝑋2 + . . . + 𝑎1𝑛𝑋𝑛 . . . 𝑃𝐶𝑚 = 𝑎𝑚1𝑋1 + 𝑎𝑚2𝑋2 + . . . + 𝑎𝑚𝑛𝑋𝑛 (3i) where 𝑎𝑚𝑛 represents the weight for the 𝑚 𝑡ℎ principal component and the 𝑛𝑡ℎ variable. Each variable is then given a factor weight corresponding to it relative importance (positively or negatively) within the component. The resulting factor weights are then used to compute a total index score which is taken as the total wealth score. The distribution of wealth scores are then split into quintiles with the lowest representing the poorest and the highest, the richest 3.8.5.2. Checking Model Assumptions OLS model diagnoses included assessing model performance, model significance, multi- collinearity , normality of residuals, stationarity and residual spatial autocorrelation. Measures of model performance included Adjusted R-squared values. Adjusted R-squared is always between 0 and 100%. Zero percent (0%) indicates that the model explains none of the variability of the response data around its mean and 100% indicates that the model explains all the variability of the response data around its mean. The overall model statistical significance was measured using both the Joint F-Statistics and Joint Wald Statistics. Multi- collinearity check was done through the assessment of variance inflation factor (VIF) values. 100 University of Ghana http://ugspace.ug.edu.gh The general rule of thumb is that VIFs beyond 4 warrant further investigation, while VIFs exceeding 10 are signs of serious multi-collinearity which require correction (O’brien, 2007; Menard, 2002). Normality of residual was determined by the statistical significance of Jarque-Bera statistic. Non-normal residuals resulted in model misspecification. The Koenker (Bruesch-Pagan) statistics was used to determine test for stationarity and statistical significance of this statistics indicated heteroscedasticity or non-stationarity. The spatial independency of residuals was evaluated by running spatial autocorrelation (Moran’s I) on the regression residuals. Global Moran’s I index is based on cross-products to measure value association, and is calculated as: 𝛴𝑛 𝑛𝑖=1𝛴𝑗=1, 𝑤𝑖,𝑗 𝑧𝑖𝑧𝑗 I = 𝑛 2 (3j) 𝑆0 𝛴𝑖=1𝑧𝑗 Where 𝑧𝑖 is the deviation of an attribute for feature i from its mean (𝑥𝑖 − ?̅?), 𝑤𝑖,𝑗 is the spatial weight between feature i and j, n is equal to the total number of features, and 𝑆0 is the aggregate of all the spatial weights. The inverse of the distance between i and j was used to specify the relationship between them. 𝑆0 = 𝛴 𝑛 𝑖=1𝛴 𝑛 𝑗=1𝑤𝑖,𝑗 The 𝑍𝐼 - score for the statistic is computed as: 𝐼−𝐸[𝐼] 𝑍𝐼 = √𝑉[𝐼] where −1 E [I] = (𝑛−1) V [I] = E [𝐼2] – E(𝐼)2 Moran’s, I tool assesses whether the pattern expressed is clustered, dispersed or random, given a set of features and accompanying attribute. The tool computes Moran’s I Index value, a Z score and evaluates the significance of the index value. A Moran's Index value near +1.0 101 University of Ghana http://ugspace.ug.edu.gh indicates clustering while an index value near -1.0 indicates dispersion. The p-value is used to determine the statistical significance of the I value (i.e. knowing if the I value is anything more than random chance). As a Spatial Autocorrelation tool, a null hypothesis is stated as: 𝐻0: “there is no spatial clustering of the values”. When the null hypothesis is rejected because the Z score is large (or small) sufficient to such that it falls outside of the desired significance, the subsequent step is to examine the value of the Moran’s I index. If the I> 0, then the set of features shows a clustered pattern, otherwise, if I<0, then the set of features shows a dispersed pattern (Griffith, 1987; Moran, 1950). All tests were conducted at α level of 5%. A GWR model was used to analyse how the independent variables relationship changed from one catchment area to another. GWR is a localized multivariate regression model which allows the parameters of a regression estimate to vary locally. Whiles OLS produces a single regression equation to summarize global relationships among the independent and dependent variables, GWR distinguishes spatial variation of relationships in a model and produced maps for exploring and interpreting spatial non-stationarity (Brunsdon et al., 2002). This study used an ADAPTIVE kernel to provide the geographic weighting in the model. This was an appropriate choice because the observations appeared to be clustered so that the density of the observations varies around the study area. Alkaike’s Information Criterion (AICc) Bandwidth method was used to allow an automatic method for finding the bandwidth which minimises the AICc value. This is written as: 𝑛+𝑡𝑟 (𝑆) AICc = 2n log e (?̂?) + n log e (2π) + { ] (3k) 𝑛−2−𝑡𝑟(𝑆) Where ?̂? is the estimated standard deviation of the error, and tr(S) is the trace of the matrix of covariates and n is the total number of catchments. 102 University of Ghana http://ugspace.ug.edu.gh 3.8.5.3. Fitting of Models Ordinary Least Square Models Prior to fitting the GWR model, OLS - a generalized linear modelling technique was fitted to establish a relationship between annual cumulative HFVR for malaria and diarrhoea (Y) and the explanatory variables (𝑋𝑖) and help select statistically significant independent variables for GWR model. This relationship was represented using a line of best-fit, where Y is predicted by values (𝑋𝑖). It is represented mathematically using the straight-line equation: Y = α + 𝛽1𝑋1 + 𝛽2𝑋2 + 𝛽3𝑋3 … + ε (3j) with α indicating the value of Y when 𝑋𝑖 is equal to zero (referred to as the intercept) and β indicating the slope of the line (refer to as the regression coefficient). The regression coefficient β describes the change in Y that is associated with a unit change in 𝑋𝑖 whereas ε was the model random error. This study runs 6 OLS regression models which consisted of two primary or overall models and 4 secondary models. The primary or overall models were run with dependent variables (HFVR for malaria and diarrhoea) computed from observed and imputed HFV obtained from reporting and non-reporting health facilities respectively in GAR. The secondary models were run to test for the effect of imputation on HFVR for malaria and diarrhoea. The dependent variables (HFVR for malaria and diarrhoea) were run against environmental / household determinants extracted from 2010 PHC data. Type of health facility variable has been added to each model to control for the effect of facility attractiveness as variable which is a potential confounder to HFV. 103 University of Ghana http://ugspace.ug.edu.gh Geographically Weighted Regression (GWR) Model GWR regression model is like OLS and can be rewritten as: 𝑌𝑖(𝑔) = 𝛽0𝑖(𝑔) + 𝛽1𝑖𝑋1𝑖(𝑔) + 𝛽2𝑖𝑋2𝑖(𝑔) + 𝛽3𝑖𝑋3𝑖 (g) +… + 𝜀𝑖 (3l) where (g) indicate the parameters that were estimated at each catchment area in which the coordinates were given by vector g; i represent each catchment area (Brunsdon et al., 2002) The GWR diagnostics were examined to evaluate the fit of the model. The diagnostics assessed local collinearity, independency and normality of residuals of GWR model. The local collinearity was assessed by scatter plots of the local coefficient estimates for the independent variables and condition number. With a condition number greater than 30, in the presence of strong collinearity, the result may be unreliable. The Adjusted R-squares (coefficient of determination), were used for comparing OLS and GWR models. AIC generated for OLS and corrected AICc calculated for GWR were also used for model comparison. The comparison was to determine which model could interpret data better. 3.8.6. Visualising GWR output for spatial heterogeneities Geographically Weighted Regression (GWR) tool was used for exploring the spatial heterogeneity. Spatial heterogeneity exists when the structure of the variable being modelled varies across the study area. The GWR tool created a report and a DataBase File ( DBF ) table which contains the diagnostic statistics. The attribute table for the output feature class contained the coefficient estimates, their standard errors, and a range of diagnostic statistics.. The local coefficient estimates of intercept and that of the independent variables (proportion of persons living in walls prone to malaria and proportion of persons within the lowest wealth quintile) were mapped. These maps showed the variation in the coefficient estimates for the 104 University of Ghana http://ugspace.ug.edu.gh variables. The map for the local coefficients reveal whether the influence of the variables in the model vary considerably over Greater Accra region or not. The range of the local coefficients serve as the evidence which points to heterogeneity in the model structure within Greater Accra region. Other maps of the diagnostics that were also analysed included the local weighted R2 value which indicated how well the GWR model replicates the HFVR for malaria aouund the independent variables and maps of pseudo-t-values for the intercept and the independent variabes, which explain the fitting level for each specific variable under the GWR. 3.9. Ethical considerations DHIMS2 database stores no patient identifiers. Approval for the research using DHIMS2 anonymised data was by Ghana Health Service Ethics Review Committee with reference number GHS-ERC: 15/11/15. The approval was renewed yearly during the period of the study. 105 University of Ghana http://ugspace.ug.edu.gh CHAPTER FOUR RESULTS AND FINDINGS 4.1. Health facility distribution in GAR Five hundred and thirty-three health facilities (both public and private) have been identified in GAR providing outpatient services in 2014. Table 4.1 shows the distribution of these health facilities by districts, ownership, type and NHIS memberships. The table also presents the number of facilities in each district that submitted their data to DHIMS2. The table shows that majority of the health facilities (72%) were clinics and health centres (these include maternity homes offering outpatient services) while 19% were hospitals, 3% were polyclinics and 6% were CHIPS compounds. Table 4.1. Distribution of health facilities in GAR by districts, NHIS status, ownership and type, 2014 DHIMS2 NHIS Owner Health facility type District Total No Yes No Yes Go Q- Priv CHPS C/H Pol Hosp v Gov y AMA 182 141 41 108 74 25 8 149 0 133 7 42 Ada East 9 1 8 1 9 8 0 1 3 5 0 1 Ada West 6 0 6 0 6 6 0 0 3 3 0 0 Adentan 20 7 13 10 10 5 0 15 2 16 0 2 Ashaiman 24 7 17 7 17 2 0 22 0 21 1 2 Ga Central 21 3 18 8 13 2 0 19 1 17 0 3 Ga East 22 6 16 11 11 2 1 19 0 19 1 2 Ga South 40 9 31 23 17 11 0 29 3 32 0 5 Ga West 31 9 22 8 23 16 0 15 6 19 0 6 Kpone- Katamanso 20 5 15 4 16 5 1 14 1 18 0 1 La Dade- Kotopon 15 9 6 6 9 1 4 10 0 12 0 1 LANKMA 33 8 25 15 18 5 0 28 0 25 2 6 LEKMA 20 10 10 10 10 7 0 13 1 13 1 5 Ningo Prampram 14 2 12 4 10 6 0 8 6 7 1 0 Shai-Osudoku 17 3 14 4 13 12 0 5 7 7 0 3 TMA 59 42 17 35 24 7 5 47 1 38 1 19 Total 533 262 271 254 279 120 19 394 34 385 14 100 Gov: Government, Q-Gov: Quasi-Government, Priv: Private, CHPS: Community-based Health and Planning Services, C/H: Clinic/Health Centre, Poly: Polyclinic, Hosp: Hospital Source: Ghana Health Service-DHIMS 2 database, USAID|Deliver Project 106 University of Ghana http://ugspace.ug.edu.gh The private health facilities constituted 74% of the entire health facilities in GAR offering outpatient service, 23% were public health facilities and the rest 3% are jointly owned by the government and its agencies. A little more than half (52%) of the health facilities subscribed to the NHIS. Figure 4.1 shows spatial distribution of the health facilities by districts. The majority of the health facilities were clustered in AMA (34%) and TMA (11%) and most of them were secondary level health facilities (i.e. hospitals and polyclinics). Some of the districts such as Ningo-Prampram and Ada West did not have any secondary level health facilities. The implication of this lack of secondary level facility for such districts is that in case of any complication, the patient needs to be referred to another district with a secondary healthcare facility to continue treatment. Figure 4.1. Geographical distribution of health facilities in GAR in 2014 107 University of Ghana http://ugspace.ug.edu.gh 4.2. Determining the level of completeness in the coverage of DHIMS2 dataset As presented in Table 4.2, 262 (49%) of health facilities in GAR did not report to DHIMS2 resulting in their HFV missing. In terms of distribution of these non-reporting health facilities by facility type, 1 facility was a CHPS compound, 202 were clinics/health centres, 1 was a polyclinic and 58 were hospitals. Classified by ownership type, 225 were private facilities made up of 172 clinics/health centres and 53 hospitals; 21 were public health facilities which consisted of 1 CHPS, 17 clinics/health centres, 1 polyclinic and 2 hospitals and 16 were quasi-public facilities which included 13 clinics/health centres and 3 hospitals. Completeness of facility reporting have been determined for each of the districts in GAR and the overall facility reporting rate for GAR has also been determined. Units that have a completeness rate below 80% were considered to have poor reporting (WHO, 2014). Summary of results are shown in Table 4.2. and districts with facility reporting rate less than 80% are shown in pink. Table 4.2: Facility reporting rate within districts in GAR Total Expected reports Actual number Facility number of (Total facilities x of reports completeness facilities 12 months) received in 12 rate(%) DISTRICT (a) (b) = (a) x 12m months (c) c/b x 100% AMA 182 2,184 504 23.08 Ada East 9 108 96 88.89 Ada West 6 72 72 100.00 Adentan 20 240 216 90.00 Ashaiman 24 288 204 70.83 Ga Central 21 252 240 95.24 Ga East 22 264 192 72.73 Ga South 40 480 468 97.50 Ga West 31 372 264 80.00 Kpone-Katamanso 20 240 168 70.00 La Dade-Kotopon 15 180 72 40.00 LANKMA 33 396 240 60.61 LEKMA 20 240 132 55.00 Ningo-Prampram 14 168 156 92.86 Shai-Osudoku 17 204 168 82.35 TMA 59 708 204 28.81 Greater Accra Region 533 6,396 3,396 53.10 108 University of Ghana http://ugspace.ug.edu.gh At the regional level, by the end of 2014 only 3,396 reports have been received (shown in Table 4.2). GAR completeness of facility reporting rate is 53% leaving a gap of 47%. Facility reporting rates within each of the 16 districts have been examined and half of the districts have facility reporting rates between the range of 23.08% to 72.73% which were less than than 80%. The worst performing districts were TMA, AMA, La Dade-Kotopon and LEKMA which are urban districts with several quasi-government and private health facilities. The missing annual cumulative HFV data for the 262 facilities have been imputed and compared to observed visits. The total HFV due to malaria used in this study was 880,735, which included 495,197(56%) observed visits and 385,538 (44%) imputed visits. In the case of diarrhoea, the annual cumulative HFV due to diarrhoea was 233,003 which consisted of 131,568 (57%) reported visits and 101,435 (43%) imputed visit. Figure 4.2 shows the percentage distribution of observed and imputed annual cumulative HFV due to (a) malaria and (b) diarrhoea by districts. The red colour showed percentage of the HFV imputed and the blue colour showed percentage of actual reported HFV by health facilities in the districts. It could be observed that districts such as AMA, TMA and La Dade- Kotopon Municipal had more than 50% of their annual HFV due to malaria and diarrhoea imputed. On the other hand, the imputed annual cumulative HFV due to malaria and diarrhoea were very low (i.e. below 20%) in districts such as Shai-Osudoku, Ningo- Prampram, Ada East and even 0% for Ada West. 109 University of Ghana http://ugspace.ug.edu.gh (a) 120 Imputed Observed 100 0 8 14 13 23 20 18 25 25 21 31 34 80 41 56 54 72 60 100 92 86 87 40 8277 8075 75 79 69 66 59 44 4620 28 0 (b) 120 Imputed Observed 100 0 11 8 10 18 17 19 14 33 30 27 27 80 35 62 57 64 60 100 89 92 90 40 82 83 81 86 67 73 7365 70 20 38 43 36 0 Figure 4.2. Percentage comparison of observed and imputed HFV due to (a) malaria and (b) diarrhoea by districts in GAR in 2014 110 Annual cumulative HFV due to malaria (%) Annual cumulative HFV due to diarrhoea (%) University of Ghana http://ugspace.ug.edu.gh Figure 4.3 (a) showed that approximately 70% (603,383) of the annual cumulative HFV due to malaria - which consisted of 281,832 (47%) observed and 321,551 (53%) imputed visits came from private health facilities. Public health facilities on the other hand contributed 241,856 (27%) HFV due to malaria of which majority 205,095 (85%) were observed visits and the rest 36,761(15%) were imputed. (a) HFV due to malaria 350,000 321,551 Observed 300,000 281,832 Imputed 250,000 205,095 200,000 150,000 100,000 50,000 36,761 27,226 8,270 0 Private Public Quasi-Public (b). HFV due to diarrhoea 80,000 74,413 74,212 70,000 Observed Imputed 60,000 55,622 50,000 40,000 30,000 20,000 15,797 11,225 10,000 1,734 0 Private Public Quasi-Public Figure 4.3. Distribution of observed and imputed HFV due to (a) malaria and (b)diarrhoea by facility ownership, in GAR,2014 111 No. of HFV due to diarrhoea No. of HFV due to malaria University of Ghana http://ugspace.ug.edu.gh The pattern of HFV due to diarrhoea in GAR was not different from what have been recorded in the case of malaria. Figure 4.3 (b) showed that out of 233,003 annual cumulative HFV due to diarrhoea in 2014 for GAR, more than half 130,035 (56%) - consisting of 55,622 (43%) observed visits and 74, 413 (57%) imputed visits, came from private health facilities. Majority (77%) of HFV due to diarrhoea, obtained by the public health facilities were observed visits. The imputed HFV due to malaria and diarrhoea for quasi-public facilities were higher than the observed visits. Figure 4.4 shows the distribution of observed and imputed annual cumulative HFV due to malaria by health facility types. The distribution pattern appeared to be similar for HFV due to both malaria and diarrhoea. It could be observed from Figure 4.4 that Clinic/HC contributed the highest number of HFV due to malaria (555,903; 63%) and diarrhoea (129,558; 56%), followed by hospitals - malaria (244,010; 28%) and diarrhoea (69,679; 30%), then by polyclinics - malaria (53,579; 6%) and diarrhoea (30,495;13%) and CHPS - malaria (25,284; 3%) and diarrhoea (3,273; 1%). Perhaps, the above distribution pattern was influenced largely by the number of such type of health facilities in the region (see Table 4.1). Figure 4.4 also showed how much each facility type contributed to observed and imputed HFV due to malaria and diarrhoea in 2014. For malaria, the proportion of observed and imputed visits obtained by each health facility type respectively were Clinic/HC (49%; 51%), hospitals (60%; 40%), polyclinics (95%; 5%) and CHPS (93%; 7%). In the case of diarrhoea, the number of observed and imputed HFV obtained respectively were clinic/HC (44%; 56%), hospitals (60%; 40%), polyclinics (96%;4%) and CHPS (100%; 0%) 112 University of Ghana http://ugspace.ug.edu.gh (a) HFV due to malaria 300,000 273,561 282,252 Observed 250,000 Imputed 200,000 145,333 150,000 98,677 100,000 50,929 50,000 25,284 2,650 1,959 0 Clinic/HC Hospital Polyclinic CHPS (b) HFV due to diarrhoea 80,000 72,535 70,000 Observed Imputed 60,000 57,023 50,000 41,967 40,000 29,305 30,000 27,712 20,000 10,000 3,273 1,188 0 0 Clinic/HC Hospital Polyclinic CHPS Figure 4.4. Distribution of observed and imputed HFV due to (a) malaria and (b) diarrhoea by facility type in GAR, 2014 4.3. Comparing DHIMS2 data to GDHS data 4.3.1. Expected number of children <5 years who were treated for malaria and diarrhoea in medical facilities (public and private) during the two weeks preceding 2014 GDHS The 2014 GDHS report has shown that 858 (<5year old) children were sampled from GAR during the 2014 GDHS and their mothers were interviewed under section 5 (i.e. child immunization, health and nutrition) of the women questionnaire and children questionnaire 113 No. of HFV due to diarrhoea No. of HFV due to malaria University of Ghana http://ugspace.ug.edu.gh (GDHS, 2015). Table 4.2 to 4.5 show the results of the various stages of analysis to determine the expected number of under-5 years old children by age groups and sex, who had fever or diarrhoea in the two weeks preceding the survey, sought treatment in a health facility (public and private) and were treated with anti-malaria or diarrhoea drugs. Table 4.3 also shows that 91 (10.6%) of the children were claimed to have episode of fever during the two weeks preceding the survey, and out of this number, 57(63%) were claimed to have sought treatment or advice at a health facility (public and private). Among the children who sought treatment at a medical facility, 17(30%) received anti-malaria treatment. Table 4.3. Sample distribution and treatment status of children <5year old with fever during 2014 GDHS in GAR Age Total Total sample Number of Number of Number of sample of by sex children who had children taken children who group chilldren fever in the 2 to medical received anti- <5years Male Female weeks preceding facility for malaria treatment the survey treatment <1year 156 (18%) 80(51%) 76 (49%) 14 14 3 (M=2, F=1) 1-4years 702 (82%) 373(53%) 329 (47%) 77 43 14 (M=9, F= 5) Total 858 453 (53%) 405 (47%) 91 57 17(M=11, F= 6) M = male and F = female Table 4.4 provides details of how the expected number of children <5 years old in the population of GAR who were treated for malaria after having sought treatment or advice at medical facilities (public and private) during the two weeks preceding the survey was arrived at. These include sample proportions of the various age and sex groups and their corresponding ranges for the true population proportions. The expected numbers for the various categories were also shown with the corresponding lower and upper confidence limits. In total, 10,539 children <5 years old were expected to have experienced the episode of fever, sought treatment in a medical facility, clinically diagnosed with malaria and treated with any anti-malaria drug. Out of the 10,539 expected number of children <5years old who 114 University of Ghana http://ugspace.ug.edu.gh were treated for malaria during the two weeks preceding the 2014 GDHS, 6,630 (63%) of these children were males and 3,909 (37%) were females, and also 2,468 (23%) were < 1 year and 8,071(77%) were between 1-4years. Table 4.4. Determination of expected number of children <5year old from the population who received anti-malaria treatment at a medical facility(public and private) for the two weeks preceding the GDHS Age 2014 Total Number of Proportion of sample Expected Number Group Population sample sample of who received anti- of children who of Greater children who malaria treatment in a received anti- Accra received anti- medical facility [95% malaria treatment in region malaria CI] a medical facility at treatment in a the population level medical facility Male <1year 64,044 (a) 80(b) 2(c) 0.025(c/b) 1,601 ((c/b) *a) [0.0030, 0.0874] [192 , 5,597] 1-4years 206,508(d) 373(e) 9(f) 0.0241 (f/e) 4,983 ((f/e) *d) [0.0111, 0.0453] [2,292 , 9355] Sub-total 270,552(g) 453(h) 11(i) 0.0243 (i/h) 6,630((i/h)*g) [0.0122, 0.0430] [3,301 , 11,634] Female <1year 64,038 (j) 76(k) 1(l) 0.0132(l/k) 843 ((l/k) *j) [0.0003, 0.0711] [19 , 4,553] 1-4years 197,319 (m) 329(n) 5(o) 0.0152 (g/f) 2,999 ((o/n)*m) [0.0050, 0.0351] [987, 6,926] Sub-total 261,357(p) 405(q) 6(r) 0.0148(r/q) 3,909((r/q)*p) [0.0055, 0.0320] [1,437, 8,363] Both Sexes <1year 128,082(s) 156(t) 3(u) 0.0192(u/t) 2,468 ((u/t)*s) [0.0040, 0.0552] [512, 7070] 1-4year 403,827(v) 702(w) 14(x) 0.0199(x/w) 8,071 ((x/w)*v) [0.0109, 0.0332] [4,402, 13,407] <5year 531,909(y) 858(z) 17(α) 0.0198(α/z) 10,539 ((α/z)*y) [0.0116, 0.0315] [6,170 , 16,755] a is <1year 2014 male population (projected) for GAR; b is <1 year male sample size used in GDHS in 2014; c is number of male children aged <1year who received anti-malaria treatment at a medical facility; d is 1-4 years 2014 male population (projected) for GAR; e is 1-4 years male sample size used in GDHS in 2014; f is number of male children aged 1-4years who received anti-malaria treatment at a medical facility; g is 2014 male population (projected) for GAR; h is the male sample size used in GDHS in 2014; i is number of male children who received anti-malaria treatment at a medical facility; j is <1year 2014 female population (projected) for GAR; k is <1year female sample size used in GDHS in 2014; l is number of female children aged <1year who received anti-malaria treatment at a medical facility; m is 1-4 years 2014 female population (projected) for GAR; n is 1-4 years female sample size used in GDHS in 2014; o is number of female children aged 1-4years who received anti-malaria treatment at a medical facility; p is 2014 female population (projected) for GAR; q is the female sample size used in GDHS in 2014; r is number of female children who received anti- malaria treatment at a medical facility; s is 2014 <1year population (projected) for GAR; t is the <1year sample size used in GDHS in 2014; u is number of <1year children who received anti-malaria treatment at a medical facility; v is 2014 1-4year population (projected) for GAR; w is children 1-4year sample size used in GDHS in 2014; x is number of 1-4year children who received anti-malaria treatment at a medical facility; y is 2014 <5year population (projected) for GAR; z is the <5year sample size used in GDHS in 2014; α is number of <5year children who received anti-malaria treatment at a medical facility 115 University of Ghana http://ugspace.ug.edu.gh Table 4.5 shows the sample distribution of the number of children <5 years old whom their had diarrhoea, sought treatment or advice at a medical facility (private and public) and were given diarrhoea drugs. It could be noted that 63 (7%) of these children had the episode of diarrhoea during the two weeks preceding the survey, and out of this number, 37 ( 59%) sought treatment or advice at medical facilities (public and private). Out of the 37 children who sought treatment or advice at medical facilities 19 (51%) received diarrhoea drugs. It could be observed that more male children (74%) visited the medical facilities (public and private) for diarrhoea as compared to their counterpart female children (26%) during the two weeks preceding the survey. Table 4.5. Sample distribution and treatment status of children <5year old who visited medical facility due to diarrhoea during two weeks preceding 2014 GDHS in GAR Age Total Total sample by sex Number of Number of Number of group sample of children who children taken children who chilldren had diarrhoea to medical received under 5 Male Female in the two facility for diarrhoea years weeks treatment treatment preceding the survey <1year 156 (18%) 80(51%) 76 (49%) 9 5 4 (M=2, F=2) 1-4years 702 (82%) 373(53%) 329 (47%) 54 32 15 (M=12, F= 3) Total 858 453 (53%) 405 (47%) 63 (7.3%) 37 19 (M=14, F=5) M = male and F = female Table 4.6 shows the expected number of children (<5 years old) from the population of the GAR who had sought treatment or advice at medical facilities (private and public) and were treated for diarrhoea during the two weeks preceding the 2014 GDHS. In all, 11,778 children <5 years old were expected to have experienced the episode of diarrhoea, sought treatment in a medical facility, clinically diagnosed with diarrhoea and received diarrhoea treatment. Out of the 11,778 expected number of children <5year old who were treated for malaria during 116 University of Ghana http://ugspace.ug.edu.gh the two weeks preceding the 2014 GDHS, 8,498 (72%) of these children were males and 3,280 (28%) were females. Table 4.6. Determination of expected number of children from the population who received diarrhoea treatment at a medical facility(public or private) for the two weeks preceding the GDHS Age group 2014 Total Number of Proportion of sample Expected Number Population sample sample of who received diarrhoea of children who (projected) children who treatment in a medical received diarrhoea received facility treatment in a diarrhoea [95% CI] medical facility at treatment in a the population level medical facility Male <1year 64,044 (a) 80(b) 2(c) 0.0250(c/b) 1,601 ((c/b) *a) [0.0030 - 0.0874] [192 – 5,597] 1-4years 206,508(d) 373(e) 12(f) 0.0322 (f/e) 6,644 ((f/e) *d) [0.0167 – 0.0555] [3,449 – 11,461] Sub-total 270552(g) 453(h) 14(i) 0.0309(i/h) 8,498((i/h)*g) [0.0170 – 0.0513] [4,599 – 13,879] Female <1year 64,038 (j) 76(k) 2(l) 0.0263(l/k) 1,685 ((l/k)*j) [0.0032 – 0.0918] [205 – 5,879] 1-4years 197,319 (m) 329(n) 3(o) 0.0091 (n/m) 1,799 ((o/n)*m) [0.0019 – 0.0264] [375 – 5,209] Sub-total 261,357(p) 405(q) 5(r) 0.0123(r/q) 3,280((r/q)*p) [0.0040 – 0.0286] [1,045 – 7475] Both Sexes <1year 128,082(s) 156(t) 4(u) 0.0256(u/t) 3,248((u/t)*s) [0.0070 – 0.0643] [897 – 8,236] 1-4year 403,827(v) 702(w) 15(x) 0.0214(x/w) 8,530((x/w)*v) [0.0120 – 0.0350] [4,846 – 14,134] <5year 531,909(y) 858(z) 19(α) 0.0220(α/t) 11,778((α/t)*s) [0.0134 – 0.0344] [7,128 – 18,298] a is <1year 2014 male population (projected) for GAR; b is <1 year male sample size used in GDHS in 2014; c is number of male children aged <1year who received diarrhoea treatment at a medical facility; d is 1-4 years 2014 male population (projected) for GAR; e is 1-4 years male sample size used in GDHS in 2014; f is number of male children aged 1-4years who received diarrhoea treatment at a medical facility; g is 2014 male population (projected) for GAR; h is the male sample size used in GDHS in 2014; i is number of male children who received diarrhoea treatment at a medical facility; j is <1year 2014 female population (projected) for GAR; k is <1year female sample size used in GDHS in 2014; l is number of female children aged <1year who received diarrhoea treatment at a medical facility; m is 1-4 years 2014 female population (projected) for GAR; n is 1-4 years female sample size used in GDHS in 2014; o is number of female children aged 1-4years who received diarrhoea treatment at a medical facility; p is 2014 female population (projected) for GAR; q is the female sample size used in GDHS in 2014; r is number of female children who received diarrhoea treatment at a medical facility.; s is 2014 <1year population (projected) for GAR; t is the <1year sample size used in GDHS in 2014; u is number of <1year children who received diarrhoea treatment at a medical facility; v is 2014 1-4year population (projected) for GAR; w is children 1-4year sample size used in GDHS in 2014; x is number of 1- 4year children who received diarrhoea treatment at a medical facility; y is 2014 <5year population (projected) for GAR; z is the <5year sample size used in GDHS in 2014; α is number of <5year children who received diarrhoea treatment at a medical facility. 117 University of Ghana http://ugspace.ug.edu.gh Table 4.6 also shows 95% confidence limits that would contain the true expected number of children from the population by sex and age groups who visited a medical facility (private and public) and had received diarrhoea treatment during the two weeks preceding the GDHS survey. 4.3.2. Comparing DHIMS2 and GDHS data A comparison of HFV data with population based data obtained from household surveys was an important way to assess the quality of estimates generated by health facility data (WHO, 2011a). Malaria Table 4.7 shows the results of comparison of children <5 years old HFV due to malaria extracted from DHIMS2 database and GDHS self-reported HFV due to malaria during the two weeks preceding the 2014 GDHS. Table 4.7 shows the overall consistency ratios, percentage difference between the two datasets and confidence interval that would contain the true expected numbers of the GDHS by age and sex groups. The overall ratio of DHIMS2 HFV due to malaria to GDHS self-reported HFV due to malaria was 0.56. This ratio translates to 44% difference between DHIMS2 HFV due to malaria and GDHS self-reported HFV due to malaria for <5 years old children. It could be observed that the difference between DHIMS2 HFV due to malaria and GDHS self-reported HFV due to malaria for male and female children less than 5years old were 49% and 37% respectively. The least difference between DHIMS2 HFV due to malaria and GDHS self-reported HFV due to malaria was 35% and was recorded by 1-4 years old female children. The highest difference was 56% recorded by <1year male children. It could also be observed that the differences between DHIMS2 and GDHS estimates for all sex by age groups (Table 4.7) were above 33% (WHO accepted standard) which suggest an unacceptable difference between the two 118 University of Ghana http://ugspace.ug.edu.gh datasets. However, taking into account that the above differences were computed from a point estimate the 95% confidence limit of the expected numbers were considered. Table 4.7 column three shows that total number of under-5 years DHIMS2 HFV due to malaria fell outside the 95% confidence limits on the expected numbers of GDHS self-reported HFV due to malaria, suggesting a difference between the two datasets. Table 4.7: Ratio of DHIMS2 to GDHS HFV due to malaria of <5 years old in GAR for a two-week average between September to December in 2014 DHIMS2 GDHS [CL] Ratio % Difference = [DHIMS2/GDHS] [1 – (DHIMS2/GDHS)]*100 Age group <1year old 1168 2444 0.48 52 [512, 7070] 1-4year old 4626 7982 0.58 42 [4402, 13407] <5year old 5794 10539 0.56 44 [6170, 16755] Sex group Male 3384 6630 0.51 49 [3301, 11634] Female 2410 3909 0.63 37 [1437, 8363] Age by sex group <1year male children 699 1601 0.44 56 [192, 5597] <1year female children 469 843 0.56 44 [19, 4553] 1-4year male children 2685 4983 0.54 46 [2292, 9355] 1-4yearfemale children 1941 2999 0.65 35 [987, 6926] CL is the upper and lower confident limit on the expected numbers However, all other age by sex groups show DHIMS2 estimates overlapping the confidence limits of the expected GDHS estimates, suggesting consistency of the two datasets. Figure 4.5 shows that DHIMS2 HFV due to malaria and GDHS self-reported HFV due to malaria were consistent in terms of patterns shown by age groups. 119 University of Ghana http://ugspace.ug.edu.gh 6000 5000 4000 3000 Male Female 2000 1000 0 DHIMS2 GDHS DHIMS2 GDHS <1year 1-4 years Figure 4.5: Column chart showing age by sex comparison of DHIMS2 and GDHS data of <5 year old HFV due to malaria in GAR for two-week average between September to December in 2014 It is observed from both datasets that more male children in the < 1year and 1-4 years’ age groups visited medical facilities for malaria treatment compared to their counterpart females. Diarrhoea Table 4.8 shows the overall consistency ratios, percentage differences between DHIMS2 HFV due to diarrhoea to GDHS self-reported HFV due to diarrhoea and confidence interval that would contain the true expected numbers of the GDHS by age and sex groups for the two weeks preceding the 2014 GDHS in GAR. The Table shows that the overall consistency ratio of DHIMS2 HFV due to GDHS self-reported HFV due to diarrhoea for all children <5 year old was 0.20. This ratio translates to 80% difference between DHIMS2 HFV due to diarrhoea and GDHS self-reported HFV due to diarrhoea for children <5year old. 120 No. of <5 years old HFV due to malaria University of Ghana http://ugspace.ug.edu.gh Table 4.8: Ratio of DHIMS2 to GDHS HFV due to diarrhoea of <5 year old reported for GAR for an average of two weeks between September to December in 2014 DHIMS2 GDHS [CL] Ratio % Difference [DHIMS2/GDHS] [1 – (DHIMS2/GDHS)]*100 Age group <1year old 878 3286 0.27 73 [897, 8236] 1-4year old 1521 8443 0.18 82 [4846, 14134] <5year old 2399 11588 0.20 80 [7128, 18298] Sex group Male 1242 8498 0.15 85 [4599, 13879] Female 1157 3280 0.33 67 [1045, 7475] Age by sex group <1year male children 335 1601 0.21 79 [192, 5597] <1year female children 543 1685 0.32 68 [205, 5879] 1-4year male children 907 6644 0.14 86 [3449, 11461] 1-4yearfemale children 614 1799 0.34 66 [375, 5209] CL is the upper and lower confident limit on the expected numbers It is observed that the difference between DHIMS2 HFV due to diarrhoea and GDHS self- reported HFV due to diarrhoea for male and female children <5year old were 85% and 67% respectively. The female children 1-4year recorded the least difference between DHIMS2 HFV due to diarrhoea and GDHS self-reported HFV due to diarrhoea the highest difference of 85% was recorded by <1year male children. From Table 4.8 the differences between DHIMS2 and GDHS estimates for all sex by age groups were above 30% (accepted WHO standard), which suggested that the differences between the two datasets were unacceptably high. However, taking into account that the above differences were computed from a point estimate, 95% confidence limit of the expected numbers were considered. Table 4.8 column three showed that the total <5 years old DHIMS2 HFVdue to diarrhoea lie outside the 95% confident limits of the expected numbers calculated for GDHS self-reported HFV due to diarrhoea, suggesting existence of a difference between the two datasets. Similar observations were made for <1year old, 1-4year old, <5year male and 1-4year male groups. Age and sex 121 University of Ghana http://ugspace.ug.edu.gh groups whose CL results showed no difference between DHIMS2 and GDHS include <5 years female, <1year males, <1year female and 1-4 years female. Figure 4.6 shows that the two datasets were consistent with patterns shown by age groups. It is observed from the graph that for both DHIMS2 HFV due to diarrhoea and GDHS self- reported HFV due to diarrhoea, more female children < 1 year were treated for diarrhoea more than their counterpart the males, and similarly in the age group 1-4 years, more male children were treated for diarrhoea more than their female counterparts. 7000 6000 5000 4000 Male 3000 Female 2000 1000 0 DHIMS2 GDHS DHIMS2 GDHS <1year 1-4 years Figure 4.6: Column chart showing age by sex comparison between DHIMS2 and GDHS HFV due to diarrhoea for <5 years old reported for GAR for an average of two weeks between September to December in 2014 122 No. of <5 HFV due to diarrhoea University of Ghana http://ugspace.ug.edu.gh 4.4. Measuring variations in HFV due to malaria and diarrhoea among population of GAR 4.4.1. Overview of HFVR due to malaria and diarrhoea in GAR in 2014 The annual cumulative HFV due to malaria recorded from all medical facilities in GAR in 2014 amounted to 880,735. The majority of the visits (70%) were reported by private health facilities, while government and quasi-government facilities reported 27% and 3% respectively. Clinic/Health Centres constitute the highest (63%) reporters of the annual cumulative HFV for 2014. This was followed by hospitals (27%), polyclinics (6%) and CHPS compounds (3%). Data analysis has revealed that overall, fewer males (46%) visited health facilities in the region with malaria episodes than their female counterparts (54%). On the other hand, annual cumulative HFV due to diarrhoea in 2014 was 233,003 which were reported for GAR by both public and private medical facilities. Approximately 56% of the visits due to diarrhoea were reported by private health facilities, while government and quasi- government facilities reported 37% and 8% respectively. In terms of facility type distribution, CHPS compounds reported 1% of the total HFV due to diarrhoea morbidity, Clinic/Health Centres 56%, Polyclinics 13% and Hospitals 30%. 56% of the total number of HFV due to diarrhoea reported in 2014 were females. 4.4.2. Age by sex distribution of HFV due to malaria and diarrhoea for GAR in 2014 Figure 4.7 shows the distribution of annual cumulative HFV due to malaria and diarrhoea reported for GAR in 2014, expressed as a proportion of 2014 population projected from 2010 census for the various age groups. The pattern of age distribution of the proportion of the visits for the age groups were generally similar (U-shaped) for the two diseases under consideration. For both malaria and diarrhoea, there were high HFV due to the diseases among children <5 years old and the proportion of the visits decrease for the consequent age groups till age group 35-49 years when it starts to rise again. It is observed that the 123 University of Ghana http://ugspace.ug.edu.gh percentage of male children under-5 years old who visited health facilities due to malaria were more than their female counterparts. Similar pattern was observed for diarrhoea as well. (a) HFV due to malaria 45.0 41.8 40.0 35.7 35.4 35.0 32.4 32.4 30.0 28.928.4 26.2 26.5 24.925.2 24.7 24.7 25.0 Female 20.0 18.1 14.8 Male 15.0 12.5 10.3 10.0 8.2 5.0 0.0 (b) HFV due to diarrhoea 14.0 12.4 12.0 10.4 10.0 8.6 7.8 8.0 7.3 7.4 6.8 7.06.56.5 6.4 Female 6.0 5.1 5.4 Male 3.9 4.0 3.1 3.2 2.2 2.4 2.0 0.0 Figure 4.7: Bar chart showing distribution of HFV due to (a) malaria) and (b) diarrhoea reported for GAR in 2014 expressed as a proportion of 2014 population projected from 2010 census 124 HFV due to diarrhoea (%) HFV due to malaria (%) University of Ghana http://ugspace.ug.edu.gh For all age groups above 14years, more females visited the medical facilities (public and private) with malaria episodes than their male counterparts. A similar trend was observed for HFV due to diarrhoea. Using the Welch t-test, significant difference was observed between percentage of males and females HFV due to malaria for the age groups 50 – 59 years and 70+ years (p<0.05). Whilst in the case of diarrhoea, a borderline significant difference was observed in the percentage HFV between males and females for ages 70 years and above (p<0.06). There was no sufficient evidence to show that statistically significant difference existed between percentage HFV of males and females in the other age groups. Appendix 1 shows details of the Welch t-test. 4.4.3. Percentage distribution of HFV due to malaria and diarrhoea at the district levels in GAR, 2014 Figure 4.8 showed the distribution of annual cumulative HFV due to malaria and diarrhoea by districts and sex, expressed as a percentage of 2014 population projected from 2010 census. The study observed that apart from AMA, higher percentage of females visited medical facilities for malaria treatment than the males in all the districts. The overall female’s HFV ranges from 63% to 42% and for males 45% to 33%. The female-bias HFV due to malaria were not different from that of diarrhoea, as majority of the districts exhibited same characteristics (see Figure 4.8 (b)). Percentage visits to medical facilities by inhabitants of the districts for diarrhoea treatment ranges from 13.2% to 1.6% among females and 10.4% to 1.4% among males. This observed female-bias in the percentage of HFV due to diarrhoea found in many districts were absent in some urban districts such as AMA, TMA and La Dade-Kotopon Municipal. From the Welch t-test, there is no sufficient evidence to show that there is significant difference in the average percentage HFV due to malaria or diarrhoea 125 University of Ghana http://ugspace.ug.edu.gh among males and females per each district (p>0.05). Appendix 2 shows details on the Welch t-test. (a) HFV due to malaria 70.0 60.0 50.0 40.0 30.0 Female Male 20.0 10.0 0.0 (b) HFV due to diarrhoea 14.0 12.0 10.0 8.0 6.0 Female 4.0 Male 2.0 0.0 Figure 4.8: Bar chart showing percentage distribution of annual cumulative HFV due to (a) malaria and (b) diarrhoea by sex and districts in GAR in 2014. 126 HFV due to diarrhoea (%) HFV due to malaria (%) University of Ghana http://ugspace.ug.edu.gh Figure 4.9 shows the annual cumulative HFVR for malaria and diarrhoea by districts in GAR in 2014. The rates have been age-standardized to remove the effects of differences between age groups to enable comparison among the districts of the region. It is observed in Figure 4.9 (a) that Ashaiman Municipal district topped all other districts with HFV rate 577 visits per 1,000 population, followed by Adentan municipality (515 visits per 1,000 population) and Kpone-Katamanso district (412 visits per 1,000 population). The districts whose inhabitants experienced less health HFVR for malaria in 2014 included Ga East, Ada East and Ga South which recorded 90, 109 and 118 per 1,000 population respectively. Figure 4.9 (b) shows that Shai-Osudoku district was in the lead this time for diarrhoea-related visits to medical facilities by the inhabitants. The district recorded 118 HFV due to diarrhoea per 1,000 population, followed by Ashaiman (114 visits per 1,000 population) and La-Nkwantana- Madina (113 visits per 1,000 population). Ga West (16 visits per 1,000 population), Ningo- Prampram (22 visits per 1,000 population) and Ga South (25 visits per 1,000 population) were the districts with less HFVR for diarrhoea in 2014. One-way ANOVA test was conducted to determine the statistical significance in the observed differences of HFVR for malaria and diarrhoea among GAR districts in 2014. The results showed that the variations in the rates among the districts were statistically significant for both malaria and diarrhoea related HFV (p<0.0001). Details of the oneway-ANOVA results can be found in appendix 3 127 University of Ghana http://ugspace.ug.edu.gh (a) HFVR for malaria 700 600 500 400 300 200 100 0 (b) HFVR for diarrhoea 140 120 100 80 60 40 20 0 Figure 4.9: Bar chart showing distribution of Age-adjusted annual cumulative HFVR for (a) malaria and (b) diarrhoea by districts in GAR in 2014. 128 HFVR for diarhoea (per 1,000 population) HFVR for malaria (per 1000 population) University of Ghana http://ugspace.ug.edu.gh 4.5. A-spatial analysis of HFV and HFVR for malaria and diarrhoea and environmental factors 4.5.1. Relationships between climatic conditions and HFV due to malaria and diarrhoea Figure 4.10 shows the association between annual cumulative HFV due to malaria and diarrhoea and mean monthly rainfall of GAR in 2014. (a) HFV due to malaria 50000 250 45000 40000 200 35000 30000 150 25000 20000 100 15000 10000 50 5000 0 0 Mean Monthly Rainfall OPD Malaria Cases (b) HFV due to diarrhoea 20000 250 18000 16000 200 14000 12000 150 10000 8000 100 6000 4000 50 2000 0 0 Mean Monthly Rainfall OPD Diarrhoea Cases Figure 4.10: Mean monthly rainfall and HFV due to (a) malaria and (b) diarrhoea reported for GAR in 2014 Source: Ghana Meteorological Agency(GMA) and DHIMS2 of GHS. 129 No. of HFV due to diarrhoea No. of HFV due to malaria Mean Monthly Rainfall ( mm) Mean Monthly Rainfall ( mm) University of Ghana http://ugspace.ug.edu.gh Figure 4.10 (a) shows that HFV due to malaria started rising from January, got to its peak in April, stabilised relatively throughout the main raining season (i.e. May to August) and declined simultaneously with the rainfall from August to December. HFV due to diarrhoea rather had an interesting association with rainfall distribution in 2014. The HFV recorded were relatively high in January and when rainfall was at its lowest. It declined since until May when rainfall was also on the increase by month except in April. The HFV increased in June and peaked in August as rainfall value was on the decline. It is worth noting that the peak of HFV for diarrhoea occurred in August after the major rainfall in June. Figure 4.11 shows the relationship between HFV due to malaria and diarrhoea and temperature for the period for GAR. For malaria (Figure 4.11(a)), the general observation was that HFV was high when the minimum and maximum temperatures were high and declined with a fall in minimum and maximum temperatures from January to August. The visits fluctuated for the rest of the year while minimum and maximum temperatures rose from September to December. Figure 4.11 (b) shows that HFV due to diarrhoea declined when mean maximum and minimum temperatures were high between January and May. The HFV increased from June and peaked in August and thereafter gradually declined to its lowest in December. Both mean maximum and minimum temperatures were low in August and rose to their highest in December. 130 University of Ghana http://ugspace.ug.edu.gh (a) HFV due to malaria 50000 35 45000 30 40000 35000 25 30000 20 25000 15 20000 15000 10 10000 5 5000 0 0 Mean Max. Temp Mean Min. Temp Malaria cases (b) HFV due to diarrhoea 20000 50 18000 45 16000 40 14000 35 12000 10000 30 8000 25 6000 20 4000 15 2000 0 10 Mean Max. Temp Mean Min. Temp OPD Diarrhoea Cases Figure 4.11: Monthly distribution of maximum and minimum temperatures and HFV due to (a) malaria and (b) diarrhoea reported for GAR in 2014 Source: Ghana Meteorological Agency(GMA) and DHIMS2 of GHS 131 No. of HFV due to diarrhoea No. of HFV due to malaria Mean Monthly Temperatures (oC) Mean Monthly Temperature in ( oC) University of Ghana http://ugspace.ug.edu.gh 4.5.2. Relationship between HFV due to malaria and diarrhoea and population density by districts This analysis showed the relationship between population density in the 16 districts of GAR and the HFV due to malaria and diarrhoea in 2014. Figure 4.12 showed the population density map of GAR districts overlaid by HFV for malaria and diarrhoea. It was observed that while AMA and Ashaiman Municipal recorded the highest population density shown by the deep brown colour, both districts also recorded the highest HFVR for malaria shown by the size of circles used. On the other hand, districts such as Ningo Prampram, Ada East and Ada West which recorded the least population densities as shown on the map by light yellow colouration also recorded the lowest number of HFVR for malaria in 2014. The pattern suggests that population density is associated with HFVR for malaria. The distribution pattern of HFVR for diarrhoea and population density by districts as shown in Figure 4.12 indicate a close relationship between the two variables. That is districts with high population densities recorded comparatively very high HFVR for diarrhoea. However, there were few semi-urban districts such as Adenta, Ga East and Ga West with high population densities but very low HFV due to malaria in 2014. Similarly, Shai-Osudoku. (a rural district) has a very low population density but very high HFV due to malaria. Similar observations were made for HFV due to diarrhoea for Kpone- Katamanso (a rural district with relatively low population density) which recorded much higher HFV than districts such as Ga East, Ga West, Ga Central, Adentan, LEKMA and LANKMA which were semi-urban districts and as such have much higher population densities. 132 University of Ghana http://ugspace.ug.edu.gh (a) HFV due to malaria (b) HFVdue to diarrhoea Figure 4.12: Graduated symbol maps of HFVR of malaria and diarrhoea superimposed on population density maps by districts in GAR in 2014. 133 University of Ghana http://ugspace.ug.edu.gh 4.6. Spatial distribution of HFVR for malaria and diarrhoea by districts The choropleth maps in Figure 4.13 shows spatial variation in annual cumulative HFVR for malaria and diarrhoea. (a) HFVR for malaria (b) HFVR for diarrhoea Figure 4.13: Choropleth map showing distribution of age-adjusted annual cumulative HFVR for (a) malaria and diarrhoea per 1,000 population by GAR districts in 2014 134 University of Ghana http://ugspace.ug.edu.gh It is observed by the deep brown colour (see Figure 4.13 (a,)), that Ashaiman and Adenta Municipal recorded more than 400 HFV per 1,000 population for malaria. These were followed by Shai-Osudoku districts with less brown colouring indicating between 300 to 400 HFV per 1,000 population for malaria. The district with the least HFVR for clinical malaria in 2014 was Ga East Municipal with 90 visits per 1,000 population represented by yellowish colouring. The distribution of HFVR for diarrhoea, is shown by Figure 4.13 (b). Shai-Osudoku district, Ashaiman Municipal district and Adenta Municipal, recorded the highest HFV due to diarrhoea in 2014, shown by the deep brown colouring with HFVR between 71 to 117 visits per 1,000 population. Ningo-Prampram, Ga East and Ga South municipalities indicated by light yellowish colouring on the map were among districts with the least HFVR for diarrhoea in GAR in 2014 with less than 25 visits per 1,000 population 4.7. Spatial distribution of HFVR for malaria and diarrhoea by catchments The HFV data was disaggregated to give insight into the distribution of the annual cumulative HFVR for malaria and diarrhoea at the local level. Figure 4.16 shows the distribution of annual cumulative HFVR for malaria and diarrhoea by catchments in the GAR using choropleth map. With the choropleth map, the annual HFVR for malaria and diarrhoea were represented on the map by grades of yellowish colour. The catchments with the highest visit rates were shaded in the darkest tone; those with the lowest visit rates were shaded in the lightest tone. For malaria, the darkest tone catchments by districts include: AMA - Asamoah clinic, Nyaho Medical Centre, Achimota hospital, Opoku Ware clinic, Dr Eustance Akwei Memorial clinic; Ada East district - Dordoekope Health Centre, Shai-Osudoku district - Shai- Osudoku Hospital, Kordiabe St Andrews Catholic Clinic and Valley View University; La- 135 University of Ghana http://ugspace.ug.edu.gh Nkwantanang-Madina Municipal - Madina Polyclinic (Rawlings Circle); Ledzokuku-Krowor Municipal Family health hospital and LEKMA hospital. (a) HFVR for malaria (b) HFVR for diarrhoea Figure 4.14: Choropleth map showing distribution of HFVR for (a) malaria and (b) diarrhoea by catchment areas in GAR, 2014. 136 University of Ghana http://ugspace.ug.edu.gh With respect to diarrhoea, catchment areas with the highest HFVR in 2014 indicated by the darkest tone colour in Figure 4.14 (b) includes AMA Civil Service - Clinic, Bob Freeman Clinic, Princess Marie Louise hospital; Ledzokuku-Krowor Municipal - LEKMA Hospital; La Dade-Kotopon Municipal - Police Hospital; Shai-Osudoku district - Osudoku Health Centre, Kordiabe St Andrews Catholic Clinic. 4.8. Cluster detection using spatial autocorrelation (Global Moran’s I) tool The Global Moran's I was used to evaluate whether the patterns expressed in Figure 4.14 were clustered, dispersed, or random. The Moran’s I Index value, a Z score and evaluates the significance of the index value. There is spatial clustering of the values associated with the geographic features in the study area when p-value is small and the absolute value of the Z score is large (or small) enough that it falls outside of the desired confidence level. If I > 0, then the set of features shows a clustered pattern, otherwise, if I<0, then the set of features shows a dispersed pattern. The result of spatial autocorrelation test was presented in Figure 4.15 below. (a) HFVR for malaria (b) HFVR for diarrhoea Moran's Index: 0.078197 Moran's Index: 0.050024 Expected Index: -0.002174 Expected Index: -0.002174 Variance: 0.000048 Variance: 0.000045 z-score: 11.612179 z-score: 7.754789 p-value: 0.000001 p-value: 0.000001 Figure 4.15: Moran’s, I spatial autocorrelation test results for the distribution of HFVR for (a) malaria and (b) diarrhoea 137 University of Ghana http://ugspace.ug.edu.gh The results, HFVR for malaria (z = 11.61, p-value < 0.001) and HFVR for diarrhoea (z =7.75, p-value < 0.001). In this case, all Moran's Indices indicate that there were spatial clustering with significant p value. These indicate that the spatial distribution of high values and/or low values in both datasets were more spatially clustered than would be expected if underlying spatial processes were random. 4.9. Getis-Ord Gi* (hotspot) analysis of HFVR for malaria and diarrhoea The Getis-Ord Gi* analysis was conducted on the distribution of HFVR for malaria and diarrhoea to determine statistically significant hotspots or locations, where lots of HFVR due to malaria and diarrhoea have clustered spatially or to determine where unexpectedly high HFVR for malaria and diarrhoea were located across the study area. Figure 4.16 shows the resulting map outputs. A high z-score and a low p-value for a feature indicates a significant hotspot. The higher (or lower) the z-score, the stronger the clustering. The red catchment areas represented the hotspots. A low negative z-score and a small p-value indicates a significant cold spot. Figure 4.16 shows that there were no cold spots. However, there were several catchment areas found to be statistically significant hotspots at 99%, 95% and 90% confidence level for both malaria and diarrhoea HFVR (see appendix 4). Appendix 4(a) provide the list of hotspots of HFVR for malaria by district and towns of location and their respective confidence levels as hotspots. Table 4.8 provide a summary of the contents of appendix 4. In all, 33 of these hotspots were identified across GAR. Majority of these hotspots were in AMA (42%), followed by La-Nkwantanang-Madina Municipal (15%), Kpone-Katamanso District (12%). Figure 4.16 (b) shows the spatial distribution of HFVR hotspots for diarrhoea. In all, 22 of these hotspots were identified. The districts with the most 138 University of Ghana http://ugspace.ug.edu.gh hotspots included AMA (41%), Kpone-Katamanso district (9%), La-Nkwantanang-Madina Municipal (9%) and Shai-Osudoku district (9%). (a) HFVR for malaria (b) HFVR for diarrhoea Figure 4.16: Getis-Ord Gi*’s maps showing hotspots of HFVR for(a) malaria and (b) diarrhoea in GAR, 2014 139 University of Ghana http://ugspace.ug.edu.gh Table 4.9: Summary of HFVR hotspots for malaria and diarrhoea in GAR, 2014 Facility type No. of catchments No. of HFVR for No. of HFVR for (catchment) analysed malaria hotspots diarrhoea hotspots CHPS 32 1 (3%) 0 Clinic/Health centre 326 19 (6%) 12(4%) Polyclinic 12 2 (17%) 2(17%) Hospital 91 11(12%) 8 (9%) Total 461 33(7%) 22(5%) Table 4.9 provide a summary of HFVR hotspots for malaria and diarrhoea in GAR, the details were captured by appendix 4. For both HFVR for malaria and diarrhoea, there were more hotspots centred on polyclinics and hospitals. This result was expected because a hospital will likely pull patients from further away, increasing the numerator and therefore raising the rate. 4.10. Distribution of HFVR for malaria and diarrhoea by health facility types From Figure 4.17 (a) among the four facility types the highest HFVR for malaria observed was about 600 per 1,000 population by hospitals and clinics/health centers while zero rate was observed as the minimum by both clinics/Health Centers and CHPS. The box plot for polyclinic is relatively the shortest among all the plots, indicating that there was less dispersion among HFVR for malaria reported by the various polyclinics. Hospitals and clinics/Health Centers boxplots are almost of the same size and comparatively the tallest. However, the box plot for hospitals show approximately a normal distribution while that of CHPS and clinics/Health Centers are positively skewed and negatively skewed for polyclinics. Outliers were observed in the HFVR for malaria reported by CHPS and polyclinics. Hospitals had the highest median HFVR for malaria while CHPS had the least. Kruskal-Wallis equality-of-populations rank test (Chisquare 𝜒2 test) confirmed a significant 140 University of Ghana http://ugspace.ug.edu.gh difference between the median HFVR due to malaria of the various health facility types (𝜒2 = 18.334, 𝑝 = 0.0004). Figure 4.19 (b) below shows that, among the facilities types, hospitals recorded the maximum HFVR for diarrhoea of little less than 200 visits per 1,000 population while both CHPS and Clinic/HC recorded a minimum rate of zero excluding outliers. (a) HFVR for malaria (b) HFVR for diarrhoea Figure 4.17: Side-by-side boxplot comparing the distribution of HFVR for (a) malaria and (b) diarrhoea and facility types in GAR in 2014 141 University of Ghana http://ugspace.ug.edu.gh The box plot for CHPS is relatively the shortest among all the plots, indicating that most of the HFVR for diarrhoea reported by the various CHPS were relatively very similar while polyclinic has the tallest box plot comparatively, showing relatively more dispersed HFVR for diarrhoea reported by the various polyclinics. The length of the 4 quartiles in the box plot for polyclinic are relatively of the same size suggesting that rates reported were evenly distributed across all the quartiles except for the one outlier. However, the length of the 4 quartiles in the box plot for remaining 3 facility types were of an uneven size. All the facility types had their observations positively skewed. With the exception of the CHPs facility type, the remaining three facility types had at least one outlier. The box plots for polyclinic tends to have the highest median HFVR for diarrhoea as Clinic/HC had the lowest median HFVR for diarrhoea. Kruskal-Wallis equality-of-populations rank test shows a significant difference between the median HFVR for diarrhoea of the various health facility types (χ2 =19.269 ,p=0.0002). 4.11. Model variables and relationships 4.11.1. Summary Statistics of malaria and diarrhoea risk factors used The independent variables were household risk factors of malaria and diarrhoea extracted from the 2010 PHC data. They represent the proportion of the population in Greater Accra region who had the highest probability of being infested by malaria or diarrhoea per their ranking of the possession the variable. Statistical analysis was conducted to determine the characteristics of each of the variables used for the models. The summary statistics for the variables were presented in the Table 4.10 142 University of Ghana http://ugspace.ug.edu.gh Table 4.10: Summary statistics of study variables Variable Mean Std. Dev. Minimum Maximum Malaria HFVR for malaria 208.01 168.81 1.19 563.09 Risk factors Proportion of persons living in walls prone to 21.97 21.05 0.00 100.00 malaria Proportion of persons within low educational 43.69 14.13 5.63 92.13 level Proportion of persons within the lowest wealth 36.35 17.01 0.00 100.00 quintile Proportion of persons in dwellings prone to 13.16 12.63 0.00 94.56 malaria Proportion of persons in houses roofed with 9.68 13.96 0.35 100.00 materials prone to malaria Proportion of persons in houses with floor 13.62 13.32 0.00 97.66 prone to malaria Proportion of persons engaged in agriculture 13.64 22.66 0.00 100.00 activities Diarrhoea HFVR for diarrhoea 53.99 59.60 0.00 518.25 Risk factors Proportion of persons using toilet facility prone to diarrhoea 50.87 26.30 0.00 100.00 Proportion of persons whose toilet facility sharing with others prone to diarrhoea 29.70 16.65 0.00 100.00 Proportion of persons whose main source of drinking water prone to diarrhoea 86.51 18.94 7.32 100.00 Proportion of persons whose main source of water for domestic use prone to diarrhoea 15.87 24.72 0.00 100.00 Proportion of persons whose method of rubbish disposal prone to diarrhoea 51.15 30.26 0.00 100.00 Proportion of persons whose method of liquid waste disposal is prone to diarrhoea 85.73 17.99 0.00 100.00 Proportion of persons within low educational level 42.53 15.79 0.01 96.60 Proportion of persons within the lowest wealth quintile 36.12 17.95 0.00 97.92 Data system effects Facility type - CHPS - - - - Facility type – Clinic/ Health Centre - - - - Facility type - Polyclinic - - - - Facility type - Hospital - - - - 4.11.2. Bivariate analysis of the dependent variables and the corresponding risk factors Some associations have been revealed by the correlation analysis shown in Figures 4.18 and Figure 4.19. Each scatter plot show the relationship between HFVR for malaria and diarrhoea (dependent variable) and environmental/household risk factors (independent variables). 143 University of Ghana http://ugspace.ug.edu.gh Proportion of persons in walls prone to malaria Proportion of persons in dwellings prone to malaria Proportion of persons in houses roofed with materials Proportion of persons in houses with floor prone to prone to malaria malaria Proportion of persons within low educational level Proportion of persons within the lowest wealth quintile Proportion of persons engaged in agriculture activities Figure 4.18: Scatter plot vs fitted graph showing relationship between annual cumulative HFVR for malaria and environment /household factors in GAR, 2014. 144 University of Ghana http://ugspace.ug.edu.gh Proportion of persons whose main source of drinking Proportion of persons whose main source of water water prone to diarrhoea for domestic use prone to diarrhoea Proportion of persons whose method of rubbish Proportion of persons whose method of liquid waste disposal prone to diarrhoea disposal is prone to diarrhoea Proportion of persons within low educational level Proportion of persons within the lowest wealth quintile Proportion of persons whose toilet facility sharing Proportion of persons using toilet facility prone to with others prone to diarrhoea diarrhoea Figure 4.19: Scatter plot vs. fitted graph showing relationship between HFVR for diarrhoea and environmental/ household risk factors in GAR, 2014. 145 University of Ghana http://ugspace.ug.edu.gh Pairwise correlation coefficients between HFVR for malaria and diarrhoea (dependent variables) and their environmental/household determinants (independent variables) were also presented in Table 4.11 and 4.12. Positive relationships were expected between the dependent variables and the independent variables for both HFVR for malaria and diarrhoea With HFVR for malaria (see Table 4.11), five out of the seven environmental / household factors (excluding facility types) were correlated with HFVR for malaria. The majority of the independent variables were significantly correlated linearly with the HFVR for malaria. However, in general the correlations were weak. In the case of diarrhoea (see Table 4.12), three of the independent variables were correlated. None of the independent variables (excluding facility types) had statistically significant HFVR for diarrhoea (excluding facility types). Like malaria, a very weak correlation could be observed between the independent variables and HFVR for diarrhoea 146 University of Ghana http://ugspace.ug.edu.gh Table 4.11: Correlation Coefficients among study variables of HFVR for malaria. HFVR for malaria Pwalpma Pdwelpma Proofpma Pflorpma PeduL Plowweaq PagricAc Ftyp_Chps Ftyp_Chps Ftyp_Chps Ftyp_Chps HFVR for malaria 1.00 risk factors Pwalpma 0.171* 1.00 (0.002) Pdwelpma 0.138* 0.495* 1.00 (0.011) (<0.001) Proofpma 0.120* 0.728* 0.508* 1.00 (0.027) (<0.001) (<0.001) Pflorpma 0.137* 0.634* 0.483* 0.628* 1.00 (0.012) (<0.001) (<0.001) (<0.001) PeduL -0.093 0.567* -0.027 0.390* 0.219* 1.00 (0.088) (<0.001) (0.614) (<0.001) (<0.001) Plowweaq -0.176* 0.067 -0.023 -0.174* -0.227* 0.317* 1.00 (0.001) (0.215) (0.672) (0.001) (<0.001) (<0.001) PagricAc -0.006 0.599* 0.052 0.455* 0.369* 0.739* -0.082 1.00 (0.912) (<0.001) (0.338) (<0.001) (<0.001) (<0.001) (0.129) Data system factors Ftyp_Chps -0.121* 0.345* 0.068 0.290* 0.207* 0.485* -0.025 0.550* 1.00 (0.026) (<0.001) (0.210) (<0.001) (<0.001) (<0.001) (0.651) (<0.001) Ftyp_C/H -0.112* -0.144* -0.047 -0.162* -0.221* -0.082 0.158* -0.234* -0.462* 1.00 (0.039) (0.008) (0.384) (0.003) (<0.001) (0.132) (0.003) (<0.001) (<0.001) Ftyp_Poly 0.016 -0.066 -0.058 -0.050 -0.024 -0.058 0.025 -0.065 -0.051 -0.245* 1.00 (0.763) (0.226) (0.290) (0.354) (0.665) (0.285) (0.652) (0.229) (0.347) (<0.001) Ftyp_Hosp 0.210* -0.052 0.030 0.002 0.120* -0.227* -0.176* -0.094 -0.154* -0.736* -0.081 1.00 (<0.001) (0.340) (0.583) (0.972) (0.027) (<0.001) (0.001) (0.084) (0.005) (<0.001) (0.134) *p<0.05, Proportion of persons living in walls prone to malaria, Pdwelpma: Proportion of persons in dwellings prone to malaria, Proofpma: Proportion of persons in houses roofed with materials prone to malaria, Pflorpma: Proportion of persons in houses with floor prone to malaria, Plowedu: Proportion of persons within low educational level, Plowweaq: Proportion of persons within the lowest wealth quintile, PagricAc: Proportion of persons engaged in agriculture activities; Health facility types - Ftyp_Chps: CHPS; Ftyp_C/H: Clinic/Health Centre; Ftyp_Poly: Polyclinic; Ftyp_Hosp: Hospitals 147 University of Ghana http://ugspace.ug.edu.gh Table 4.12: Correlation Coefficients among study variables for HFVR for diarrhoea. HFVR for Ptoipdiar Ptoishpdia Pdrwatpdia Pdowatpdi Prudispd Plwasdis Plowedu Plowwe Ftyp_Chps Ftyp_C/ Ftyp_Pol Ftyp_Hos diarrhoea a ia pd aq H y p HFVR for 1.00 diarrhoea risk factors Ptoipdiar -0.001 1.00 (0.991) Ptoishpdia -0.013 -0.273* 1.00 (0.811) (<0.001) Pdrwatpdia -0.027 0.049 0.248* 1.00 (0.615) (0.372) (<0.001) Pdowatpdia 0.035 0.333* 0.202* 0.002 1.00 (0.515) (<0.001) (<0.001) (0.970) Prudispdia 0.002 0.757* -0.211* -0.041 0.402* 1.00 (0.967) (0.522) (<0.001) (0.447) (<0.001) Plwasdispd -0.019 0.522* 0.289* 0.618* 0.525* 0.433* 1.00 (0.730) (<0.001) (<0.001) (<0.001) (<0.001) (<0.001) Plowedu 0.003 0.272* -0.126* 0.005 0.088 0.239* 0.082 1.00 (0.962) (<0.001) (0.020) (0.921) (0.104) (<0.001) (0.130) Plowweaq -0.011 0.256* -0.044 0.042 0.003 0.183* 0.074 0.906* 1.00 (0.847) (<0.001) (0.421) (0.435) (0.954) (<0.001) (0.174) (<0.001) Data system factors Ftyp_Chps -0.080 0.320* -0.066 -0.203* 0.262* 0.331* 0.113* 0.143* 0.049 1.00 (0.139) (<0.001) (0.224) (<0.001) (<0.001) (<0.001) (0.037) (0.008) (0.366) Ftyp_C/H -0.170* -0.003 0.005 0.089 -0.124* -0.026 0.009 -0.052 0.004 -0.462* 1.00 (0.002) (0.963) (0.931) (0.102) (0.023) (0.639) (0.867) (0.336) (0.945) (<0.001) Ftyp_Poly 0.183* -0.030 0.035 0.037 -0.041 -0.062 -0.017 -0.044 -0.028 -0.051 -0.245* 1.00 (<0.001) (0.579) (0.517) (0.493) (0.452) (0.256) (0.749) (0.414) (0.609) (0.347) (<0.001) Ftyp_Hosp 0.182* -0.213* 0.027 0.026 -0.0270 -0.181* -0.084 -0.023 -0.028 -0.154* -0.736* -0.081 1.00 (<0.001) (<0.001) (0.614) (0.630) (0.624) (<0.001) (0.120) (0.670) (0.604) (0.005) (<0.001) (0.134) *p<0.05, Ptoipdiar: Proportion of persons using toilet facility prone to diarrhoea, Ptoishpdiar: Proportion of persons whose toilet facility sharing with others prone to diarrhoea, Pdrwatpdiar: Proportion of persons whose main source of drinking water is prone to diarrhoea, Pdowatpdiar: Proportion of persons whose main source of water for domestic use is prone to diarrhoea, Prudispdiar: Proportion of persons whose method of rubbish disposal is prone to diarrhoea, Plwasdispdiar: Proportion of persons whose method of liquid waste disposal is prone to diarrhoea, Plowedu: Proportion of persons within low educational level, Plowweaq: Proportion of persons within the lowest wealth quintile; Health facility types - Ftyp_Chps: CHPS; Ftyp_C/H: Clinic/Health Centre; Ftyp_Poly: Polyclinic; Ftyp_Hosp: Hospit 148 University of Ghana http://ugspace.ug.edu.gh 4.11.3. Checking assumptions underlying OLS model for HFVR for malaria and diarrhoea To determine the best environmental/household predictors of HFVR for malaria and diarrhoea by OLS regression model approach, assumptions underlying the use of the model were evaluated. Normality of the residuals was assessed using histogram composed with HFVR for malaria and diarrhoea residuals superimposed with normal distribution curve as illustrated in Figure 4.20. The test for residual normality was conducted on square root transformed HFVR for malaria and diarrhoea (outcome variables) as the original HFVR were not normally distributed. For HFVR for malaria, Figure 4.20 shows an approximately normal distribution of the residuals produced by the model. For HFVR for diarrhoea, the graph shows a skewness at the upper tail, a minor and trivial deviation from normality. Therefore it was concluded that the residuals were close to a normal distribution and was accepted. (a) HFVR for malaria (a) HFVR for diarrhoea Figure 4.20: Histograms showing square root transformed HFVR for (a) malaria and (b) diarrhoea residuals superimposed with normal density functions Breusch-Pagan test was used to test the null hypothesis 𝐻0: that the variance of the residuals was homogenous. The estimated Chi-square test statistic and the corresponding p-value 149 University of Ghana http://ugspace.ug.edu.gh (HFVR for malaria: Chi2 =1.52, p= 0.2173 and HFVR for diarrhoea: Chi2 = 1.06, p= 0.3029) showed that the homoscedasticity assumption was not violated (p>0.05). Variance inflation factor (VIF) was used to determine whether the estimated coefficients were inflated leading to the existence of multi-collinearity . The general rule of thumb is that VIFs beyond 4 warrant further investigation, while VIFs exceeding 10 are signs of serious multi-collinearity which require correction (O’brien, 2007). Table 4.13 showed the reported variance inflation factors among HFVR for malaria and diarrhoea predictor variables. Table 4.13: Variance inflation factor (VIF) values for malaria and diarrhoea predictors. Predictors VIF HFVR for malaria risk factors Proportion of persons living in walls prone to malaria 3.85 Proportion of persons within low educational level 3.98 Proportion of persons within the lowest wealth quintile 1.78 Proportion of persons in dwellings prone to malaria 1.89 Proportion of persons in houses roofed with materials prone to malaria 2.76 Proportion of persons in houses with floor prone to malaria 2.15 Proportion of persons engaged in agriculture activities 3.49 Data System factors Facility type - CHPS Facility type – Clinic/ Health Centre 3.99 Facility type - Polyclinic 1.47 Facility type - Hospital 3.71 HFVR for diarrhoea risk factors Proportion of persons whose toilet facility sharing with others prone to diarrhoea 1.67 Proportion of persons whose main source of water for domestic use prone to diarrhoea 1.97 Proportion of persons whose method of rubbish disposal prone to diarrhoea 2.61 Proportion of persons whose main source of drinking water prone to diarrhoea 2.75 Proportion of persons using toilet facility prone to diarrhoea 3.63 Proportion of persons whose method of liquid waste disposal is prone to diarrhoea 5.20 Proportion of persons within the lowest wealth quintile 6.55 Proportion of persons within low educational level 6.56 Data system factors Facility type - CHPS Facility type - Clinic/ Health Centre 3.42 Facility type - Polyclinic 1.39 Facility type - Hospital 3.28 150 University of Ghana http://ugspace.ug.edu.gh For HFVR for malaria, the VIF has a mean value of 2.91, a minimum value of 1.47 and a maximum value of 3.99. In the case of HFVR for diarrhoea the mean VIF value was 3.55, with a minimum value of 1.39 and a maximum value of 6.63. Three of the variables for diarrhoea had their VIF above 4 but were below 10 which was acceptable and therefore required no further correction. These outcomes suggest that there was less collinearity among the predictor variables of HFVR for malaria and diarrhoea and therefore the variables used presented no limitation for this study. Therefore the null hypothesis was accepted. 4.12. Ordinary Least Square (OLS) Regression Analysis Six OLS regression models were run consisting of two primary/overall models and 4 secondary models. Three of the OLS models, one primary/overall (consisted of observed + imputed data) and two secondary models and one each for observed and imputed data, were run for each of the diseases under consideration (i.e. HFVR for malaria and diarrhoea). The OLS tests results are presented in the following sub-sections: 4.12.1. HFVR for malaria OLS regression models The relationship between HFVR for malaria on one hand and environmental/household risk factors and data system effects on the other hand for the primary/overall model (Table 4.14) was tested. The t-statistics test the hypothesis that the value of an individual coefficient estimate is not significantly different from zero. The OLS test result shows that the proportion of persons living in walls prone to malaria and proportion of persons within the lowest wealth quintile were the only two household risk factors that have a relationship with HFVR for malaria. However, all the data system effect variables (health facility types) were significantly associated with HFVR for malaria. The summary of OLS results show proportion of persons living in walls prone to malaria, proportion of persons within the 151 University of Ghana http://ugspace.ug.edu.gh lowest wealth quintile and health facility types have p-values that were lower than the 0.05 level at which they were tested suggesting that those variables were statistically significant. Except for proportion of persons living in walls prone to malaria and facility types, the coefficient estimates of other covariates were all not in the expected direction. The VIFs were all reasonably small indicating that there was no strong evidence of variable redundancy. Model 1 – Overall HFVR for malaria model (observed + imputed visits rates) Table 4.14: OLS result for the overall HFVR for malaria ( observed + imputed rates) model Summary of OLS Results Variable Coeft StdErr t- Prob [b] Robust_SE Robust_t Robust_Pr VIF [c] [a] or Statistic [b] Intercept 11.3690 2.2107 5.1426 0.000* 2.3637 4.8098 0.0000* ----- risk factors Pwalpma 0.1241 0.0297 4.1840 0.000* 0.0237 5.2291 0.0000* 3.8530 Pdwelpma -0.0047 0.0347 -0.1354 0.8924 0.0354 -0.1326 0.8946 1.8907 Proofpma -0.0296 0.0378 -0.7817 0.4349 0.0343 -0.8611 0.3898 2.7582 Pflorpma -0.0310 0.0349 -0.8856 0.3765 0.0328 -0.9438 0.3460 2.1456 Plowedul -0.0295 0.0449 -0.6561 0.5122 0.0496 -0.5944 0.5526 3.9771 Plowweaq -0.0679 0.0250 -2.7218 0.006* 0.0238 -2.8441 0.0047* 1.7819 PagricAc -0.0149 0.0263 -0.5680 0.5704 0.0269 -0.5534 0.5804 3.4939 Data system factors Ftyp_Chps Ref Ftyp_C/H 3.3672 1.3728 2.4529 0.015* 1.3550 2.4850 0.0134* 3.9917 Ftyp_Poly 5.5065 2.4075 2.2872 0.023* 2.2024 2.5001 0.0129* 1.4727 Ftyp_Hosp 6.0176 1.5416 3.9034 0.000* 1.5172 3.9663 0.00009* 3.7102 OLS Diagnostics Input Features: Primary HFVR for malaria Dependent Variable: Square root of HFVR for malaria Number of Observations: 341 Akaike's Information Criterion (AICc) [d]: 2188.8130 Multiple R-Squared [d]: 0.147426 Adjusted R-Squared [d]: 0.121591 Joint F-Statistic [e]: 5.706337 Prob(>F), (7,333) degrees of freedom: 0.00000* Joint Wald Statistic [e]: 89.529757 Prob(>chi-squared), (7) degrees of freedom: 0.00000* Koenker (BP) Statistic [f]: 9.529358 Prob(>chi-squared), (7) degrees of freedom: 0.0482* Jarque-Bera Statistic [g]: 11.302808 Prob(>chi-squared), (2) degrees of freedom: 0.0035* * An asterisk next to a number indicates a statistically significant p-value (p < 0.01). [a] Coefficient: Represents the strength and type of relationship between each explanatory variable and the dependent variable. [b] Probability and Robust Probability (Robust_Pr): Asterisk (*) indicates a coefficient is statistically significant (p < 0.01); if the Koenker (BP) Statistic [f] is statistically significant, use the Robust Probability column (Robust_Pr) to determine coefficient significance [c] Variance Inflation Factor (VIF): Large Variance Inflation Factor (VIF) values (> 7.5) indicate redundancy among explanatory variables. [d] R-Squared and Akaike's Information Criterion (AICc): Measures of model fit/performance. [e] Joint F and Wald Statistics: Asterisk (*) indicates overall model significance (p < 0.01); if the Koenker (BP) Statistic [f] is statistically significant, use the Wald Statistic to determine overall model significance. [f] Koenker (BP) Statistic: When this test is statistically significant (p < 0.01), the relationships modeled are not consistent (either due to non-stationarity or heteroskedasticity). You should rely on the Robust Probabilities (Robust_Pr) to determine coefficient significance and on the Wald Statistic to determine overall model significance. [g] Jarque-Bera Statistic: When this test is statistically significant (p < 0.01) model predictions are biased (the residuals are not normally distributed) [h] Ref: Reference category for health facility type 152 University of Ghana http://ugspace.ug.edu.gh The OLS diagnostics result shows Adjusted R-Squared value of 0.12 which indicated that the model (independent variables modelled using linear regression) explained only 12% of the variation in HFVR for malaria with the AICc value of 2188.81. The joint F-statistic and Wald statistic were significant indicating that the overall model and independent variables in the model were effective (CI: 95%, p-value <0.05). The test for heteroskedasticity using the Koenker (BP) statistics showed that the dependent variable was non-stationary (Chi2=9.5294, p-value <0.05). This Koenker test of non-stationarity shows local relationships for which GWR analysis could be considered. The Jarque-Bera statistics (Chi2=11.3028, p<0.05) shows that model residuals were not normally distributed, indicating model misspecification (which meant a key variable was missing from the model). Hence the OLS result cannot be trusted. The OLS results of the secondary models were presented in Table 4.15 and 4.16. Table 4.16 shows that four household risk factors and all the data system factors of the observed HFVR for malaria model were in the hypothesised direction. It was observed that only one household risk factor.i.e. the proportion of persons living in walls prone to malaria was statistically significant (CI: 95%, p-value <0.05). For data system factors, polyclinics and hospitals were significantly related to the observed HFV due to malaria. The Adjusted R- Squared value for this model was 0.18, which means that 18% of the total variation in the observed HFVR for malaria could be explained by the linear relationship between observed HFVR for malaria one hand and household risk factors factors and data system effects on the other on the other hand, with AICc value of 1061.06. The joint F-statistic and Wald statistic of the observed HFVR for malaria were significant suggesting that the independent variables in the models were effective (CI: 95%, p-value <0.05). In Table 4.16 it has been observed that three of the household risk factors of the imputed HFVR for malaria were in the hypothesised direction. The proportion of persons in the low 153 University of Ghana http://ugspace.ug.edu.gh wealth quintile was the only risk factor that is statistically significant in the model together with one data system factor - hospital. The Adjusted R-Squared value was 0.21, which suggest that 21% of the total variation in the imputed HFVR for malaria model could be explained by the household factors and data system factors. The joint F-statistic was significant but the Wald statistic could not be explained. Model 2 – Observed HFVR for malaria Table 4.15: OLS model result for observed HFVR for malaria Summary of OLS Results Variable Coeft StdErr t- Prob [b] Robust_SE Robust_t Robust_Pr VIF [c] [a] or Statistic [b] Intercept 1.9573 3.2908 0.5948 0.5529 3.2456 0.6030 0.5474 ------- risk factors Pwalpma 0.1565 0.0466 3.3579 0.0001* 0.0390 4.0116 0.0001* 5.5913 Pdwelpma 0.0087 0.0464 0.1885 0.8507 0.0422 0.2071 0.8362 1.7351 Proofpma -0.0997 0.0507 -1.9680 0.0509 0.0370 -2.6932 0.0079* 3.1689 Pflorpma -0.0245 0.0473 -0.5185 0.6047 0.0438 -0.5604 0.5761 2.1666 Plowedul 0.0757 0.0699 1.0830 0.2805 0.0630 1.2016 0.2314 5.1364 Plowweaq 0.0097 0.0435 0.2404 0.8103 0.0364 0.2668 0.7900 1.9473 PagricAc -0.0282 0.0377 -0.7496 0.4546 0.0317 -0.8902 0.3748 4.7909 Data system effects Ftyp_Chps Ref Ftyp_C/H 2.8841 1.5679 1.8394 0.0678 1.4406 2.0020 0.0471 2.5321 Ftyp_Poly 7.835 2.7129 2.8879 0.0045* 2.5419 3.0823 0.0024 1.4974 Ftyp_Hosp 8.1558 1.9394 4.2054 0.0000* 1.8301 4.4565 0.0000* 2.1963 OLS Diagnostics Input Features: Observed HFVR Dependent Variable: Square root of observed HFVR for malaria for malaria Number of Observations: 162 Akaike's Information Criterion (AICc) [d]: 1061.0559 Multiple R-Squared [d]: 0.229320 Adjusted R-Squared [d]: 0.178282 Joint F-Statistic [e]: 4.493098 Prob(>F), (7,333) degrees of freedom: 0.000015* Joint Wald Statistic [e]: 81.241181 Prob(>chi-squared), (7) degrees of freedom: 0.00000* Koenker (BP) Statistic [f]: 9.738396 Prob(>chi-squared), (7) degrees of freedom: 0.463737 Jarque-Bera Statistic [g]: 9.209455 Prob(>chi-squared), (2) degrees of freedom: 0.010004* 154 University of Ghana http://ugspace.ug.edu.gh Model 3 – Imputed HFVR for malaria Table 4.16: OLS model result for imputed HFVR for malaria. Summary of OLS Results Variable Coeft StdErr t- Prob [b] Robust_SE Robust_t Robust_Pr VIF [c] [a] or Statistic [b] Intercept 18.4278 5.1096 3.6065 0.0004* 1.2802 14.3942 0.0000* ----- risk factors Pwalpma 0.0472 0.0356 1.3233 0.1875 0.0246 1.9188 0.0567 2.6542 Pdwelpma -0.0272 0.0553 -0.4919 0.6234 0.0605 -0.4496 0.6536 3.0288 Proofpma 0.1035 0.0609 1.7000 0.0910 0.0687 1.5070 0.1337 2.9478 Pflorpma -0.0224 0.0551 -0.4065 0.6849 0.0589 -0.3805 0.7041 3.0178 Plowedul 0.0037 0.0550 0.0674 0.9564 0.0585 0.0634 0.9496 2.4863 Plowweaq -0.1353 0.0287 -4.7118 0.0000* 0.0301 -4.4910 0.0000* 1.8140 PagricAc -0.0289 0.0369 -0.7846 0.4338 0.0406 -0.7126 0.4771 1.7725 Data system effects Ftyp_Chps Ref Ftyp_C/H -0.7318 4.9310 -0.1484 0.8822 0.6236 -1.1735 0.2423 33.2121 Ftyp_Poly -1.7311 6.9448 -0.2491 0.8036 0.7854 -2.2041 0.0289 2.0073 Ftyp_Hosp 0.5576 4.9748 0.1121 0.9109 0.9623 0.5794 0.5631 32.7057 OLS Diagnostics Input Features: Imputed HFVR Dependent Variable: Square root of imputed HFVR for for malaria malaria Number of Observations: 179 Akaike's Information Criterion (AICc) [d]: 1090.7692 Multiple R-Squared [d]: 0.251771 Adjusted R-Squared [d]: 0.207233 Joint F-Statistic [e]: 5.653015 Prob(>F), (7,333) degrees of freedom: 0.00000* Joint Wald Statistic [e]: NaN Prob(>chi-squared), (7) degrees of freedom: NaN Koenker (BP) Statistic [f]: 11.508130 Prob(>chi-squared), (7) degrees of freedom: 0.319323 Jarque-Bera Statistic [g]: 2.322328 Prob(>chi-squared), (2) degrees of freedom: 0.313121 4.12.2. Comparing HFVR for malaria primary/overall and secondary models Table 4.17 and Figure 4.22 show how the observed and imputed HFVR for malaria OLS results were compared with results of the primary/overall model (observed + imputed HFVR). It will be seen from Table 4.16 that the two statistically significant independent variables - proportion of persons living in walls prone to malaria and proportion of persons within the lowest wealth quintile, in the primary model - were contributed by the observed and imputed HFVR for malaria models respectively. While 12% of the total variation in HFVR for malaria could be explained by the linear relationship between independent and 155 University of Ghana http://ugspace.ug.edu.gh dependent variables in the primary/overall model, this value have increased marginally in the observed (18%) and imputed (22%) HFVR for malaria models AICc value for the primary model was approximately twice the values obtained from the observed and imputed HFVR for malaria models. Table 4.17: Comparison OLS diagnostics of HFVR for malaria primary and secondary models Observed model Imputed model Overall (observed + imputed) model Predictors Coefficients Intercept 1.9573 18.4278* 11.3690* risk factors Pwalpma 0.1565* 0.0472 0.1241* Pdwelpma 0.0087 -0.0272 -0.0047 Proofpma -0.0997 0.1035 -0.0296 Pflorpma -0.0245 -0.0224 -0.0310 Plowedul 0.0757 0.0037 -0.0295 Plowweaq 0.0097 -0.1353* -0.0679* PagricAc -0.0282 -0.0289 -0.0149 Data system factors Ftyp_C/H 2.8841 -0.7318 3.3672* Ftyp_Poly 7.8347* -1.7311 5.5065* Ftyp_Hosp 8.1558* 0.5576 6.0176* Other diagnostics Adjusted-R2 0.18 (18%) 0.21 (21%) 0.12 (12%) AICc 1061.06 1090.77 2188.81 *p<0.05, Proportion of persons living in walls prone to malaria, Pdwelpma: Proportion of persons in dwellings prone to malaria, Proofpma: Proportion of persons in houses roofed with materials prone to malaria, Pflorpma: Proportion of persons in houses with floor prone to malaria, Plowedu: Proportion of persons within low educational level, Plowweaq: Proportion of persons within the lowest wealth quintile, PagricAc: Proportion of persons engaged in agriculture activities; Health facility types - Ftyp_Chps: CHPS; Ftyp_C/H: Clinic/Health Centre; Ftyp_Poly: Polyclinic; Ftyp_Hosp: Hospitals 156 University of Ghana http://ugspace.ug.edu.gh (a) Overall ( observed + imputed) HFVR for malaria (b) Observed HFVR for malaria (c) Imputed HFVR for malaria Figure 4.21: Comparison of OLS StdResid maps of HFVR for malaria:(a) overall (b) observed and (c) imputed 157 University of Ghana http://ugspace.ug.edu.gh Figure 4.21 (a), (b) and (c) above show the maps of standardised residuals of the OLS models for HFVR for malaria primary/overall, observed and imputed models respectively. The map shows that the imputed data coverage for catchment areas of AMA, TMA, LEKMA and La Dade-Kotopon Municipality (see Figure 4.21 (c)) which are largely urban districts. The observed data coverage was largely concentrated in districts other than the 4 districts mentioned above (See Figure 4.21 (b)). This information is critical in explaining the contributions of the observed and imputed HFVR for malaria models. Figure 4.21 also identify the catchment areas in GAR which have positive standardised residuals greater than 2 (model underprediction) or negative standard standardised residuals less than -2 (model over prediction) in all the three maps. Red clusters correspond to catchments with high HFVR for malaria, where the observed values were under predicted, and blue clusters were catchments with low HFVR for malaria where the observed values were over predicted Some of the clustering patterns identified in the observed and imputed HFVR for malaria model standardised residual maps were also seen in the primary HFVR for malaria model standardised residual map. 4.12.3. HFVR for diarrhoea OLS regression models Table 4.18 showed the relationship between HFVR for diarrhoea on one hand and environmental/household risk factors and data system factors on the other hand for the primary/overall model. 158 University of Ghana http://ugspace.ug.edu.gh Model 4 – Primary HFVR for diarrhoea model Table 4.18: OLS model result for primary HFVR for diarrhoea ( observed + imputed visit rates) Summary of OLS Results Variable Coeft StdError t- Prob [b] Robust_SE Robust_t Robust_Pr VIF [c] [a] Statistic [b] Intercept 5.4558 1.0612 5.1375 0.0000* 0.8539 6.3890 0.0000* ---- risk factors Ptoipdiar 0.0054 0.0124 0.4365 0.6628 0.0139 0.3884 0.6980 3.6312 Ptoishpdia -0.0001 0.0132 -0.0731 0.9418 0.0119 -0.0811 0.9354 1.6717 Pdrwatpdia -0.0088 0.0129 -0.6823 0.4955 0.0115 -0.7663 0.4440 2.7501 Pdowatpdi 0.0090 0.0082 1.0959 0.2739 0.0071 1.2718 0.2044 1.9694 Prudispdia 0.0035 0.0091 0.3850 0.7005 0.0106 0.3303 0.7414 2.6063 Plwasdispd -0.0032 0.0173 -0.1842 0.8540 0.0160 -0.1988 0.8425 5.2017 Plowedul 0.0043 0.0044 0.9704 0.3325 0.0041 1.0541 0.2926 6.5649 Plowweaq - 0 .0049 0.0047 -1.0377 0.3002 0.0035 -1.3965 0.1635 6.5476 Data system factors Ftyp_Chps Ftyp_C/H 1.2379 0.6957 1.7795 0.0761 0.6085 2.0345 0.0427* 3.4181 Ftyp_Poly 4.8747 1.2779 3.8114 0.0002* 1.8430 2.6450 0.0086* 1.3858 Ftyp_Hosp 3.0715 0.7938 3.8692 0.0001* 0.7362 4.1721 0.0000* 3.2800 OLS Diagnostics Input Features: Primary HFVR for diarrhoea Dependent Variable: Square root of HFVR for diarrhoea Number of Observations: 341 Akaike's Information Criterion (AICc) [d]: 1779.301545 Multiple R-Squared [d]: 0.092644 Adjusted R-Squared [d]: 0.062307 Joint F-Statistic [e]: 3.053820 Prob(>F), (8,332) degrees of freedom: 0.000653* Joint Wald Statistic [e]: 34.142252 Prob(>chi-squared), (8) degrees of freedom: 0.000343* Koenker (BP) Statistic [f]: 14.04873 Prob(>chi-squared), (8) degrees of freedom: 0.230309 Jarque-Bera Statistic [g]: 111.845166 P rob(>chi-squared), (2) degrees of freedom: 0.00000* Notes on Interpretation * An asterisk next to a number indicates a statistically significant p-value (p < 0.01). [a] Coefficient: Represents the strength and type of relationship between each explanatory variable and the dependent variable. [b] Probability and Robust Probability (Robust_Pr): Asterisk (*) indicates a coefficient is statistically significant (p < 0.01); if the Koenker (BP) Statistic [f] is statistically significant, use the Robust Probability column (Robust_Pr) to determine coefficient significance [c] Variance Inflation Factor (VIF): Large Variance Inflation Factor (VIF) values (> 7.5) indicate redundancy among explanatory variables. [d] R-Squared and Akaike's Information Criterion (AICc): Measures of model fit/performance. [e] Joint F and Wald Statistics: Asterisk (*) indicates overall model significance (p < 0.01); if the Koenker (BP) Statistic [f] is statistically significant, use the Wald Statistic to determine overall model significance. [f] Koenker (BP) Statistic: When this test is statistically significant (p < 0.01), the relationships modeled are not consistent (either due to non-stationarity or heteroskedasticity). You should rely on the Robust Probabilities (Robust_Pr) to determine coefficient significance and on the Wald Statistic to determine overall model significance. [g] Jarque-Bera Statistic: When this test is statistically significant (p < 0.01) model predictions are biased (the residuals are not normally distributed). 159 University of Ghana http://ugspace.ug.edu.gh The summary OLS result (see Table 4.18), reveal that majority of the environmental/ household risk factors as well as all the data system factors were in the hypothesised direction. However, none of the environmental/household predictors had a statistically meaningful relationship with HFVR for diarrhoea . In the case of the data system factors, polyclinic and hospitals have statistically significant relationship with the dependent variable. The OLS diagnostics results also show that the independent variables had a combined R- squared of 0.06. This means that 6% of the total variation in HFVR for diarrhoea could be explained by the linear relationship between HFVR for diarrhoea on one hand and environmental/ household factors and data system factors on the other hand. The other 94% of the total variation in HFVR for diarrhoea remains unexplained. However, joint F-statistic and Wald statistic were all significant which implies that the overall model was statistically significant. The test for heteroscedasity using the Koenker (BP) statistics showed that the dependent variable was stationary (Chi2=6.7526, p>0.05). The Jarque-Bera statistics (Chi2=111.845, p<0.05) indicated that model predictions were biased (i.e. residuals were not normally distributed). Since the tests show that the dependent variable was stationary, none of the environmental/household variables were statistically significant and the residuals were not normally distributed (conditions necessary for implementing GWR), and therefore GWR was not considered for further analysis of HFVR for diarrhoea. The OLS results for the secondary models are presented in Table 4.18 and 4.19. It was observed in Table 4.19 that with the HFVR for diarrhoea model, majority of the environmental /household risk factors and all the data system factors were in the hypothesised direction. The proportion of persons using toilet facility prone to diarrhoea was the only environmental/ household factor having statistically significant relationship with the observed HFV for diarrhoea. However, all the data system variables were statistically 160 University of Ghana http://ugspace.ug.edu.gh significant. The Adjusted R-Squared value for this model was 0.09 with AICc value of 899.72. Model 5 –Observed HFVR for diarrhoea Table 4.19: OLS model result for observed HFVR for diarrhoea. Summary of OLS Results Variable Coeft StdError t- Prob [b] Robust_SE Robust_t Robust_Pr VIF [c] [a] Statistic [b] Intercept 4.5956 1.5447 2.9750 0.0034 1.1577 3.9697 0.0001* ----- risk factors Ptoipdiar 0.0546 0.0249 2.1959 0.0296* 0.0270 2.0207 0.0451* 5.1553 Ptoishpdia 0.0283 0.0257 1.1012 0.2726 0.0246 1.1471 0.2532 2.5479 Pdrwatpdia 0.0096 0.0237 0.4048 0.6862 0.0243 0.3952 0.6932 5.7793 Pdowatpdia 0.0182 0.0136 1.3390 0.1826 0.0121 1.5029 0.1350 3.0192 Prudispdia -0.0052 0.0161 -0.3206 0.7490 0.0184 -0.2810 0.7791 2.9645 Plwasdispd -0.0526 0.0372 -1.4140 0.1594 0.0389 -1.3506 0.1789 12.318 Plowedul 0.0036 0.0060 0.6048 0.5462 0.0043 0.8510 0.3961 4.7856 Plowweaq -0.0026 0.0077 -0.3390 0.7351 0.0068 -0.3838 0.7017 4.1450 Data system factors Ftyp_Chps Ftyp_C/H 1.4383 0.8879 1.6199 0.1074 0.7079 2.0318 0.0439* 2.2159 Ftyp_Poly 5.5654 1.6143 3.4475 0.0007 1.9607 2.8384 0.0052* 1.4468 Ftyp_Hosp 4.2953 1.1352 3.7838 0.0002 1.0110 4.2486 0.0000* 2.0534 OLS Diagnostics Input Features: Observed HFVR for diarrhoea Dependent Variable: Sqrt of observed HFVR for diarrhoea Number of Observations: 162 Akaike's Information Criterion (AICc) [d]: 899.717180 Multiple R-Squared [d]: 0.156146 Adjusted R-Squared [d]: 0.094263 Joint F-Statistic [e]: 2.523256 Prob(>F), (7,333) degrees of freedom: 0.006030* Joint Wald Statistic [e]: 30.602652 Prob(>chi-squared), (7) degrees of freedom: 0.001273* Koenker (BP) Statistic [f]: 12.528409 Prob(>chi-squared), (7) degrees of freedom: 0.325261 Jarque-Bera Statistic [g]: 33.883146 Prob(>chi-squared), (2) degrees of freedom: 0.00000* With the imputed HFVR for diarrhoea model, it is observed from Table 4.20 that four of the environmental / household risk factors were in the hipothesized direction but none of them were statistically significant. Similarly, only one of the data system factors was statistically significant but they were in the hypothesized direction. The Adjusted R-Squared value was 0.07 with AICc value of 867.79 161 University of Ghana http://ugspace.ug.edu.gh Module 6 – Imputed HFVR for diarrhoea Table 4.20: OLS model result for imputed HFVR for diarrhoea. Summary of OLS Results Variable Coeft StdError t- Prob [b] Robust_SE Robust_t Robust_Pr VIF [c] [a] Statistic [b] Intercept 1.2053 3.2560 0.3702 0.7117* 2.2898 0.5264 0.5993* risk factors Ptoipdiar -0.0247 0.0128 -1.9394 0.0541 0.0137 -1.8032 0.0732 2.6024 Ptoishpdia -0.0175 0.0145 -1.2129 0.2269 0.0140 -1.2489 0.2134 1.3765 Pdrwatpdia -0.0224 0.0229 -0.9796 0.3287 0.0251 -0.8936 0.3728 1.9095 Pdowatpdia -0.0106 0.0156 -0.6779 0.4988 0.0122 -0.8616 0.3902 1.7219 Prudispdia 0.0121 0.0105 1.1590 0.2481 0.0120 1.0115 0.3132 1.9644 Plwasdispd 0.0247 0.0171 1.4455 0.1502 0.0150 1.6416 0.1026 2.3272 Plowedul 0.0049 0.0090 0.5455 0.5861 0.0081 0.6073 0.5445 19.5866 Plowweaq -0.0080 0.0086 -0.8876 0.3760 0.0071 -1.0709 0.2857 20.1072 Data system effects Ftyp_Chps Ftyp_C/H 6.2148 2.6339 2.3595 0.0194 0.3261 19.0531 0.0000* 33.1631 Ftyp_Poly 6.3242 3.7179 1.7010 0.0908 0.4031 15.6900 0.0000* 2.0112 Ftyp_Hosp 7.4937 0.2393 2.869 0.0054* 0.6597 11.3596 0.0000* 32.7307 OLS Diagnostics Input Features: Imputed HFVR Dependent Variable: Square root of imputed HFVR for diarrhoea for diarrhoea Number of Observations: 179 Akaike's Information Criterion (AICc) [d]: 867.795542 Multiple R-Squared [d]: 0.124033 Adjusted R-Squared [d]: 0.066334 Joint F-Statistic [e]: 2.149675 Prob(>F), (7,333) degrees of freedom: 0.019472* Joint Wald Statistic [e]: NaN Prob(>chi-squared), (7) degrees of freedom: 0.000001* Koenker (BP) Statistic [f]: 17.700630 Prob(>chi-squared), (7) degrees of freedom: 0.088792 Jarque-Bera Statistic [g]: 38.556986 Prob(>chi-squared), (2) degrees of freedom: 0.000001* 4.12.4. Comparing HFVR for diarrhoea primary and secondary models Table 4.21 and Figure 4.22 compared the observed and imputed HFVR for diarrhoea OLS results were compared with results of the primary model (combined observed and imputed HFVR). From Table 4.21, it was observed that none of the environmental/household variable was statistically significant in imputed HFVR for diarrhoea model. In the observed HFVR for diarrhoea model, proportion of persons using toilet facility prone to diarrhoea statistically significant but failed to be significant in the primary HFVR for diarrhoea model. None of the data system factors were significant for the observed HFVR model while two of them were significant in both imputed and primary modes. The AICc value for the primary HFVR for diarrhoea model was twice the observed and imputed HFVR for diarrhoea models. The 162 University of Ghana http://ugspace.ug.edu.gh Adjusted R-square values for the primary, observed and imputed models were 0.09, 0.07 and 0.06 respectively, meaning none of the three models could explain at least 10% of the total variation in HFVR for diarrhoea in each model. Table 4.21: Comparison of OLS diagnostics of HFVR for diarrhoea primary and secondary models. Predictors Coefficients Observed model Imputed mode Primary model (observed + imputed) Intercept 4.5956 1.2053 5.4558* risk factors Ptoipdiar 0.0546 -0.0247 0.0054 Ptoishpdia 0.0283 -0.0175 -0.0001 Pdrwatpdia 0.0096 -0.0224 -0.0088 Pdowatpdia 0.0182 -0.0106 0.0090 Prudispdia -0.0052 0.0121 0.0035 Plwasdispd -0.0526 0.0247 -0.0032 Plowedul 0.0036 0.0049 0.0043 Plowweaq -0.0026 -0.0080 -0.0049 Data system factors Ftyp_Chps Ftyp_C/H 1.4383 6.2148* 1.2379 Ftyp_Poly 5.5654 6.3242 4.8747* Ftyp_Hosp 4.2953 7.4937* 3.0715* Other diagnostics Adjusted-R2 0.09(9%) 0.07(7%) 0.06 (6%) AICc 899.717 867.796 1779.30 *p<0.05, Ptoipdiar: Proportion of persons using toilet facility prone to diarrhoea, Ptoishpdiar: Proportion of persons whose toilet facility sharing with others prone to diarrhoea, Pdrwatpdiar: Proportion of persons whose main source of drinking water is prone to diarrhoea, Pdowatpdiar: Proportion of persons whose main source of water for domestic use is prone to diarrhoea, Prudispdiar: Proportion of persons whose method of rubbish disposal is prone to diarrhoea, Plwasdispdiar: Proportion of persons whose method of liquid waste disposal is prone to diarrhoea, Plowedu: Proportion of persons within low educational level, Plowweaq: Proportion of persons within the lowest wealth quintile, Health facility types - Ftyp_Chps: CHPS; Ftyp_C/H: Clinic/Health Centre; Ftyp_Poly: Polyclinic; Ftyp_Hosp: Hospitals 163 University of Ghana http://ugspace.ug.edu.gh (a) Overall ( observed + imputed) HFVR for diarrhoea (b) Observed HFVR for diarrhoea (c) Imputed HFVR for diarrhoea Figure 4.22: Comparison of OLS Residual maps for HFVR for diarrhoea primary and secondary models 164 University of Ghana http://ugspace.ug.edu.gh Figure 4.22 presents maps of standardised residuals of the HFVR for diarrhoea OLS models. Like the distribution in HFVR for malaria models, majority of the observed data were collected from catchment areas outside Accra Metropolis, Tema Metropolis, Ledzokuku- Krowor Municipality and La Dade-Kotopon Municipality (see Figure 4.22 (b)). On the other hand, the imputed data came mainly from 4 districts mentioned above (see Figure 4.22 (c)). The figure identified the locations which have positive standardised residuals greater than 2 (model underprediction) or negative standard standardised residuals less than -2 (model overprediction), suggesting evidence of clustering of over-and /or underpredictions in some areas. Red clusters represent catchments with high HFVR for diarrhoea, where the observed values were under predicted, and blue clusters were catchments with low HFVR for diarrhoea, where the observed values were over predicted. There were similarities in clustering patterns identified in the observed and imputed HFVR for diarrhoea model’s standardised residual maps if compared to the primary HFVR for diarrhoea model standardised residual map. 4.13. Testing for clustering of residuals using spatial autocorrelation (Global Moran’ I) OLS test recommends that spatial autocorrelation (Global Moran’ I) be run on regression residuals to assess residual spatial autocorrelation or ensure they are spatially random. Thus, spatial autocorrelation (Global Moran's I) tool was used to test the spatial independence of the regression residuals. The tool computes Moran’s I Index value, a Z score and evaluates the significance of the index value. There is spatial clustering of the values associated with the geographic features in the study area when the p-value is small and the absolute value of the Z score is large (or small) enough that it falls outside of the desired confidence level. If I> 0, then the set of features show a clustered pattern, otherwise, if I<0, then the set of features 165 University of Ghana http://ugspace.ug.edu.gh shows a dispersed pattern. The result of this test presented in Figure 4.23 shows that both tests were statistically significant with HFVR for malaria (z = 5.0957, p-value < 0.001) and HFVR for diarrhoea (z =4.6811, p-value < 0.001). These results indicate that the spatial distribution of high values and/or low values in both datasets were more spatially clustered than would be expected if underlying spatial processes were random. The observed Moran’s value of 0.03 suggested very little spatial structure in both HFVR for malaria and diarrhoea regression residual. However, this result suggests that some important independent variables (determinants) are missing from the models. Malaria Diarrhoea Moran's Index: 0.034971 Moran's Index: 0.031773 Expected Index: -0.002941 Expected Index: -0.002941 Variance: 0.000055 Variance: 0.000055 z-score: 5.095730 z-score: 4.681106 p-value: 0.000000 p-value: 0.000000 Figure 4.23: Spatial autocorrelation on OLS regression residuals of HFVR for malaria and diarrhoea. 166 University of Ghana http://ugspace.ug.edu.gh 4.14. Geographically Weighted Regression (GWR) Model 4.14.1. Spatial distribution of dependent (HFVR for malaria) and independent variables (Pwalpma and Plowweaq) Prior to the GWR model, choropleth maps was used to show the spatial distribution of the two statistically significant independent variables (from the OLS test) across the study area. Figure 4.26 shows the map of the spatial distribution of (a) HFVR for malaria (b) proportion of persons living in walls prone to malaria (Pwalpma); and (c) proportion of population within low wealth quintile level (Plowweaq). The blank spaces represents either catchment areas with very large group of EAs that were not delineated and therefore were excluded from the analysis or catchment areas with no independent variables. Figure 4.24 (a) shows that HFVR for malaria were high among catchment areas located in the central part of Shai-Osudoku district, eastern part of Ada West district, western part of Ningo-Prampram district and across Kpone-Katamanso district. Catchment areas around north-eastern parts of AMA and Ga East have also shown high HFVR for malaria. As shown in Figure 4.24 (b), majority of the catchments with more than 50% of the population living in walls prone to malaria were common in districts such as Shai-Osudoku, Ningo-Prampram, Ada West, Ada East and Ga South. Pockets of such catchments areas could also be found in Accra and Tema Metropolitan Assemblies. It will appear that the distribution of proportion of populations within the low wealth quintile was fairly distributed by catchment across all the districts in GAR (see Figure 4.24 (c)). In AMA, catchments with very high proportion of persons within the low wealth quintile level were found along the coast in the towns such as Glefe, Chorkor, James Town and Korle- Gorno and along river Ordorna in the towns such as Agblogboshie and Ashiedu-kete. South Adenta and adjoining northern parts of Ledzokuku- Krowor district have also been identified to have high proportion of persons living within low wealth quintile level in GAR. 167 University of Ghana http://ugspace.ug.edu.gh (a) HFVR for malaria (b) Proportion of persons living in walls prone to malaria (c) Proportion of persons within the lowest wealth quintile Figure 4.24: Spatial distribution of (a) HFVR for malaria; (b) Proportion of persons living in walls prone to malaria (Pwalpma); and (c) Proportion of population within low wealth quintile level (Plowweaq). 168 University of Ghana http://ugspace.ug.edu.gh 4.14.2. GWR model for HFVR for malaria It was evident from the results of the OLS model that HFVR for malaria regression fitting did not perform very well. The statistically significant Jarque-Bera statistics showed that the model residuals were not normally distributed, indicating that the model was biased. There was also spatial clustering of residuals (i.e. autocorrelation in the residuals). All the above- mentioned characteristics of the HFVR for malaria OLS regression model were at variance with the assumption of OLS model. Hence, the HFVR for malaria OLS model results were not trustworthy. As a result, the study would be justified in using GWR to attempt to improve on the reliability of the predictions, especially when the Koenker (BP) statistic indicated that the model was non-stationary. Recall that the study was interested in mapping the values of the catchment specific coefficient estimates to explore whether the process be spatially heterogeneous in small areas across the study area or not. The GWR local model was run to analyse how the annual cumulative HFVR for malaria vis-à-vis the proportion of persons living in walls prone to malaria (Pwalpma) and proportion of population within low wealth quintile level (Plowweaq) relationship changed from one catchment to another. Table 4.22 compared the summary information about model variables and parameters for OLS and GWR models. Each local regression equation was calibrated using 104 neighbours. By comparing the fit of the OLS and GWR models, the Adjusted R-square for OLS was approximately 12% and the GWR Adjusted R-squared was 22% which suggested that there has been significant improvement in model performance when GWR was used. AICc was chosen to measure the model fit, the OLS model’s AICc value was 2186.6358, and the GWR model’s AICc value was 2166.3718, a reduction of 20.2640 which was a strong evidence of an improvement in the fit of the GWR model to the data. Other indicators that have shown 169 University of Ghana http://ugspace.ug.edu.gh that the GWR has given a better fitting model than the OLS were the residual sum of squares and the -2 Log Likelihood; both statistics have reduced in GWR model as compared to OLS. Table 4.22: Comparison between OLS and GWR model for HFVR for malaria. Parameter OLS GWR Residual sum of squares 11450.8690 8941.3572 Sigma 5.8729 5.5093 AICc 2186.6358 2166.3718 R square 0.14 0.33 Adjusted R square 0.12 (12%) 0.22(22%) -2log-likelihood 2165.9691 2081.6131 4.14.3. Visualising the GWR output The values of the standardised residuals (StdResid) were mapped – these are shown in Figure 4.25. This was done to show whether the residuals were unusually high or low and to determine whether the residuals were spatially autocorrelated. Locations where the models were over predicting and under predicting are shown by difference in colour coding. Figure 4.25: GWR standard residual map for HFVR for malaria 170 University of Ghana http://ugspace.ug.edu.gh Red clusters correspond to areas of high HFVR for malaria where the observed values were under predicted, and blue clusters were areas of low HFVR for malaria where the observed values were over predicted. There were very few large positive residual (StdResid > 2.5) catchments which include New Crystal clinic in Kakasunaka at Kpone-Katamanso district, Royal MMR clinic in Dome at Ga East District and Bortianor Health Center in Bortianor at Ga South district. There was no catchment with large negative values (StdResid < - 2.5). This result suggests that the GWR model has improved over the OLS model. The report from the spatial autocorrelation tool applied on the GWR residuals was shown in Figure 4.26. At 95% confidence level Moran’s I for the residual was 0.0031(p >0.05) suggesting that there was no evidence of any autocorrelation. Moran’s Index 0.0031 Expected Index: -0.0029 Variance: 0.00005 z-score: 0.8176 p-value: 0.4135 Figure 4.26: Spatial autocorrelation result for HFVR for malaria GWR model residual As a result, any spatial correlation which might have existed in the residuals of the OLS model have been taken care of with the geographical weighting in the GWR model. However, given that only 22% of the total variation in the dependent variable could be explained by this model, additional variable(s) must be added to the model in the future. Since there was no 171 University of Ghana http://ugspace.ug.edu.gh clustering as suggested by the spatial autocorrelation (Moran’s I) result, the GWR model has predicted well for the data used. 4.15. GWR model result and spatial variation in HFVR for malaria The local coefficients of HFVR for malaria were mapped. Figure 4.27 shows the maps of (a) locally weighed R2 between the observed and fitted values and (b) condition number (a) Local R2 (b) Condition Number Figure 4.27: Spatial mapping of (a) locally weighted coefficient of determination (R2) and (b) condition number by GWR modelling. 172 University of Ghana http://ugspace.ug.edu.gh Locally weighted R2 value (goodness of fit statistics) indicate how well the GWR model replicates the HFVR for malaria around proportion of persons living in walls prone to malaria and proportion of persons within the lowest wealth quintile. R2 value by catchment area vary between the ranges of 10% to 46% with a mean value of 30%. Though R2 was generally low across the catchment areas in the study area (Figure 4.27 (a)), it was evident that the value of R2 was not homogeneously distributed in all the catchment areas – evidence which point to heterogeneity in the model structure within GAR. It has been observed from the map that the overall GWR regression fitted well in most districts. However, the model did not fit very well in districts such as Ada West, Ada East and Ga South, and this could imply additional independent variables are needed to explain the HFVR for malaria in those districts. Figure 29 (b) shows multi-collinearity by catchment area, expressed by the condition number. Condition number measures the level of dependencies in the data. Values around 5 to 10 suggest weak dependencies in the data and a larger value greater than 30 indicates an increased degree of multi-collinearity . The condition number in this model produced values of between the ranges of 8.6 to 13.2 with a mean of 10.0, suggesting that the model is less affected by multi-collinearity . Figure 4.28 (a) – (c) shows the mapped pseudo t-values for intercept and each independent variable to represent the fitting level for each specific variable under GWR. The blue and red areas representing the significant t-values, showing that the parameter estimations in these areas were reliable. 173 University of Ghana http://ugspace.ug.edu.gh (a) Pseudo t-values of regression fitting for Intercept (b) Pseudo t-values of regression fitting for Proportion of persons living in walls prone to malaria (c) Pseudo t-values of regression fitting for Proportion of persons within the lowest wealth quintile Figure 4.28: Spatial mapping of pseudo-t values of (a) intercept (b) wall prone to malaria (Pwalpma) and (c) Low wealth quintile (Plowweaq) for each catchment by GWR modelling 174 University of Ghana http://ugspace.ug.edu.gh The spatial variations in coefficient estimations for intercept, proportion of persons living in walls prone to malaria and proportion of persons within the lowest wealth quintile are shown in Figure 4.29 (a)-(c). The map of intercept term represent the distributions of HFVR for malaria when Pwalpma and Plowweaq were zero. Figure 4.29 (a), shows that higher intercept values were found in La Dade-Kotopon, LEKMA, AMA and part of Shai-Osudoku. On the other hand very low intercept values were found in Ada East, Ada West, Ga West and parts of Ga South and Adenta. The observed spatial heterogeneity in the intercept term implied that in addition to proportion of persons living in walls prone to malaria and proportion of persons within the lowest wealth quintile, there were some other variables that would influence HFVR for malaria pattern in GAR. The relationship between HFVR for malaria and Pwalpma shown in Figure 4.29 (b) suggests that, in all the districts in GAR, an increased HFVR for malaria would relate to a corresponding increase in the proportion of persons living in walls prone to malaria but with varying rates across the region. For example, the rate of increase in HFVR for malaria in AMA, La Dade-Kotopon and LEKMA were less as compared to Ningo Prampram, Ada East and Ada West. The distribution of proportion of persons within low quintile level parameter show a more interesting pattern (Figure 4.29 (c)). The influence of proportion of persons within low quintile level parameter on HFVR for malaria was negative in the urban districts such as AMA, TMA, La Dade-Kotopon and LEKMA representing higher HFVR for malaria associated with lower proportion of persons within low quintile level. However, in the remaining districts, positive relationships existed between HFVR for malaria and Plowweaq, indicating higher proportion of persons within low quintile level tended to associate with higher HFVR for malaria and vice versa. 175 University of Ghana http://ugspace.ug.edu.gh (a) Estimated coefficient of intercept (b) Estimated coefficient of Pwalpma (c) Estimated coefficient of Plowweaq Figure 4.29: Spatial mapping of (a)estimated coefficients of intercept, (b) wall prone to malaria (Pwalpma) and (c) low wealth quintile (Plowweaq) for each catchment by GWR modelling. 176 University of Ghana http://ugspace.ug.edu.gh 4.16. Key findings a) A total of 533 health facilities (both public and private) have been identified in GAR in 2014, offering outpatient services and out of which 271(51%) reported their data into DHIMS2 and the remaining 262 (49%) did not. b) GAR recorded completeness of facility reporting rate of 53%, that is 47% incompleteness. Half of the districts had completeness of facility reporting rate between the ranges 23.08% to 72.73% which translate to between 77% to 23% incompleteness of facility reporting. c) The total number of HFV due to malaria and diarrhoea used in this study were 880, 735 and 233,003 respectively which included imputed HFV for non-reporting health facilities. Imputed HFV due to malaria was 385,538 (44%) and that of diarrhoea was 101,435(43%). d) More than 50% of non-reporting health facilities were private and located in the three urban districts of GAR which include AMA, TMA and La Dade-Kotopon municipality. e) GDHS data has revealed that the number of 5-year-old and below male children reported more HFV due to malaria (63%) than their female counterpart (37%). Like malaria, HFV due to diarrhoea has similar pattern with 74% male children visiting the medical facilities compared to their counterpart female (26%) for the same age under discussion. f) Data consistency check has revealed an overall ratio of DHIMS2 to GDHS of 0.56. This ratio translates to 44% difference between DHIMS2 HFV due to malaria and GDHS self-reported HFV due to malaria for children <5 years old HFV due to malaria on the average of two weeks between September and December in 2014. 177 University of Ghana http://ugspace.ug.edu.gh g) There is a large discrepancy between HFV in GAR estimated through the GDHS versus estimates through DHIMS2. For diarrhoea visits, consistency check has revealed an overall ratio of DHIMS2 to GDHS of 0.20. This ratio translated to 80% difference between DHIMS2 HFV due to diarrhoea and GDHS self-reported HFV due to diarrhoea for children <5 years HFV due to diarrhoea in an average of two weeks between September and December in 2014. h) On the contrary, the 95% confidence level of the upper and lower confidence range that would contain the true expected population numbers for all sex and age categories for both HFV due malaia and diarrhoea overlaps the corresponding estimates from the DHIMS2 data, suggesting no difference between the two datasets. i) Both the DHIMS2 and GDHS HFV data were consistent with patterns shown by children less than 1 year and for 1-4 years age groups HFV for malaria and diarrhoea for the two weeks period between September and December 2014. j) Analysis of HFV revealed that type of health facility influences HFV by patients in GAR. Hospitals and polyclinics were more attractive to both malaria and diarrhoea patients as compared to clinics and CHPS. k) Spatial distribution of HFVR for malaria showed Ashaiman municipal district with the highest rate (577 HFV per 1000 population), Adentan municipality (515 HFV per 1,000 population) and Kpone-Katamanso (412 HFV per 1,000 population). Among the districts which experienced less HFVR for malaria in 2014 were Ga East, Ada East and Ga South which recorded values of 90, 109 and 118 HFV per 1,000 populations respectively. l) With regards to HFVR for diarrhoea, Shai-Osudoku districts recorded the highest rate of 118 visits per 1,000 population, followed by Ashaiman (114 visits per 1,000 population) and La-Nkwantana-Madina (113 visits per 1,000 population). Ga West 178 University of Ghana http://ugspace.ug.edu.gh (16 visits per 1,000 population), Ningo-Prampram (22 visits per 1,000 population) and Ga South (25 visits per 1,000 population) were the districts with less severe HFVR for diarrhoea in 2014. m) Interpretation of results has suggested that there were male-biased malaria and diarrhoea-related medical facility visits among children <5 years old and female- biased malaria and diarrhoea-related medical facility visits among adults aged 15 years and above. n) Analysis of data reveal several small areas (catchments) with statistically significant hotspots of HFVR for malaria and diarrhoea across GAR o) The observed relationships between HFVR for malaria and diarrhoea and climatic factors (rainfall, temperature and population density) were consistent with earlier literature. p) The OLS test result showed that proportion of persons living in walls prone to malaria and proportion of persons within the lowest wealth quintile were the only two variables that have a statistically significant relationship with HFVR for malaria. q) In summary OLS result for diarrhoea show that none of the environmental/ household determinants have statistically significant relationship with HFVR for diarrhoea. r) The local GWR model for HFVR for malaria improved the model fit considerably and HFVR for malaria relationship in different areas of GAR with proportion of persons living in walls prone to malaria and proportion of persons within the lowest wealth quintile were found to be positively or negatively associated. s) Both OLS and GWR tests have suggested that some important independent variables (determinants) were missing from the model t) HFVR for malaria was found to vary spatially over GAR. 179 University of Ghana http://ugspace.ug.edu.gh u) GWR results indicate that proportion of persons living in walls prone to malaria and proportion of persons within the lowest wealth quintile were significantly spatially non-stationary – evidence that spatial heterogeneity exists in these relationships. v) Geospatial tools such as choropleth map, dot density maps, hotspot analysis, spatial autocorrelation (Moran’s I) and GWR were found to be very effective for GIS-based small area analysis. 180 University of Ghana http://ugspace.ug.edu.gh CHAPTER FIVE DISCUSSION The main objective of this study was to explore the usage of geospatial and statistical techniques to address issues associated with limited use of facility-based data and to evaluate its value for small area evidence-based decision-making. The main outcome of this study is the enhancement of the value of facility-based data, which is measured by the plausible indicators estimated from the data using geospatial techniques. Studies have suggested that there is consistent evidence of spatial clustering of some diseases (Joseph et al., 2011; Sabel et al., 2003; Glaser, 1990) thus suitable data are required to give precise information on population at risk and exposures. There has also been a general awareness that survey data are less reliable for estimating indicators for small area public health decision making due to small sample sizes (Nutley & Reynolds, 2013; Buescher, 1997). Alegana et al. (2017) noted that as malaria prevalence declines, it becomes increasingly difficult to rely on community-based surveys like the MICS to measure it – the sample sizes get bigger and bigger and the costs go up. These is sufficient justification for looking at outpatient data as an alternative to currently used data sources for small area evidence-based decision making. In Ghana, considerable investments have gone into development and maintenance of the DHIMS2 database and therefore there is the need to add value to these investments for optimum healthcare delivery. 5.1. What can DHIMS2 tell us about outpatient attendance? The quality of health data in decision-making cannot be over-emphasised, as the consequences of a single wrong decision from bad data can be very devastating and cause 181 University of Ghana http://ugspace.ug.edu.gh irreparable damage to individuals and the society at large. However, one cannot agree anymore with WHO (2014) report that “No health data from any source can be considered perfect: all data are subject to a number of limitations related to quality, such as low coverage, missing values, bias, measurement error, and human errors in data entry and computation. Data quality assessment is needed to understand how much confidence can be put in the health data presented”. The report recommended that there is the need to know the reliability of national coverage estimates derived from RHISs (WHO, 2014, 2008a). In line with the above recommendation and as a quality assessment measure, this study attempted to determine the level of completeness of facility reporting of outpatient data into DHIMS2 database (WHO, 2014). External verification of consistency was also conducted to determine the comparability between DHIMS2 outpatient children under-5 year old data and that of GDHS data (WHO, 2014). The following sections discuss the resultant findings: 5.1.1. Examining completeness of DHIMS2 data This study investigated the level of completeness in the coverage of DHIMS2 data set so that the data could be manipulated to make it comparable to other data sources to enhance its usability for public health decision-making. As at December 2014, the number of health facilities in GAR providing outpatient services amounted to 533. The fact that approximately 49% of these health facilities in the region did not report their data into DHIMS2 database (see Table 4.1) was by itself a worrying situation, when this database serves as the national RHIS. WHO (2014) noted that to make an informed interpretation on key indicators, it is critical to know completeness of facility reporting rate. The study shows that facility completeness rate of GAR is 53%, a gap of 47% incompleteness (see Table 4.2) which is way below the 182 University of Ghana http://ugspace.ug.edu.gh WHO’s limit of 80%. The study also shows that completeness of health facility reporting is below the 80% for half of the districts. These results suggest that GAR has poor reporting, which is as a result of very poor facility reporting by half of the districts which are mainly urban ( AMA, TMA, La Dade-Kotopon and LEKMA) and peri-urban (Ashaiman, Ga East, Kpone-Katamanso and LANKMA), and therefore have several non-reporting private health facilities. Many developing countries face a similar challenge of poor reporting by districts. For example, a similar study conducted by Muhindo et al. (2016) in Uganda shows that completeness of district reporting is poor in 9% of districts and one-third of the districts scored below the 80% limit for completeness of health facility reporting. Private health facility are mainly blamed for such a performance. Aqil et al. (2009), noted that RHI has much been disregarded by private health facilities in decision making; as such they have less motivation for its collection and reporting. According to a report that examines the legal framework relevant to the creation and use of health information in Ghana, private health facilities were not obliged to submit any report to any health authority (MOH, 2012). According to the Ghana urban malaria report (JSI, 2013), for incompleteness of such magnitude in routine health facility-based records, manipulation of the data (e.g. imputation of missing values) is unavoidable to give a meaningful interpretation and conclusion drawn from statistics derived from the dataset. Unfortunately, most health sectors which report on disease outcomes do not manipulate their result to take care of such data challenges. This observation is line with the earlier findings by WHO (2011b) that in low income and middle- income countries tools and methods to correct bias and missing values to generate estimates and forecast for planning purposes are very key, but due to poor quality of routine facility- based data but are rarely used. The annual estimates of disease outcomes reported by key agencies in Ghana are usually based on out-patient visits as captured by the DHIMS2 database, when as a matter of fact, data coverage of health facilities (most especially the 183 University of Ghana http://ugspace.ug.edu.gh private facilities) is not documented and updated regularly by DHIMS2 database. For example, a report such as “The reported presumptive malaria visits in 2010 forms about 38% of the total outpatient visits. On regional basis, Greater Accra recorded the least proportion (27% of all outpatient visits) with the highest being Upper East region with 53%” (NMCP, 2010), can be misleading since HFV data in DHIMS2 does not cover all health facilities in GAR and for that matter the entire country. Such reports do not give the true picture of malaria morbidity in GAR and likewise any other region. According to WHO (2008c) health facilities data can provide instant and continuing information relevant to public health decision-making but only if data relate to all facilities (public and private), is of high quality, and is representative of the services available to the population. The result of imputation showed that the overall annual cumulative HFV due to clinical malaria and diarrhoea imputed for GAR in 2014 constituted approximately 44% and 43% respectively, of the total HFV used in this study. District level distribution of observed and imputed HFV due to malaria and diarrhoea showed the top three districts as AMA, TMA and La Dade-Kotopon Municipal, which are among the most urbanised districts in GAR. For these districts, the imputed annual cumulative HFV due to both malaria and diarrhoea were more than double the annual cumulative, unadjusted raw HFV totals in 2014. The reverse situation was that the lowest number of imputed HFV due to malaria and diarrhoea were recorded by Shai-Osudoku, Ningo-Prampram, Ada East and Ada West which are rural districts in GAR. The above results suggests that the problem of non-reporting leading to data incompleteness is very severe in urban districts compared to rural districts. Most imputed HFV for malaria and diarrhoea (70%) for GAR in 2014 were attributed to non- reporting by private health facilities and the majority of these private health facilities were in the urban districts. This result also suggests that public health facilities are more compliant in reporting their data to the national RHIS (DHIMS2) than the private health facilities. These 184 University of Ghana http://ugspace.ug.edu.gh findings were in agreement with what has been reported in the ‘Report of the Ghana Urban Malaria Study’ in 2012 (JSI, 2013). The study admits that data routinely reported by health facilities in Accra (Accra Metropolitan District plus all of three other neighbouring districts: Ledzokuku/ Krowor, Tema and Ashaiman) were often highly incomplete and that the completeness of these data varied significantly between urban and rural facilities. The report also stated that non-reporting health facilities were private and these have not yet been persuaded to submit their monthly data to DHIMS2. Analysis of the contribution of types of health facilities to the imputed HFV has revealed clinics/health centres as the biggest contributors. This is not surprising considering the number of clinics and health centres owned by private health providers in the urban districts who usually will not report their data to DHIMS2. This study has found that some specialised regional hospitals, quasi-government health institutions and affiliates of some tertiary health institutions do not report to the national database (DHIMS2) because they are autonomous health institutions. GHS which manages DHIMS2 does not include Teaching hospitals, University hospitals, Pivate and Mission hospitals, Armed forces hospitals and some quasi- government institutions. As a result these institutions are not obliged to report data into DHIMS2, which adversely affect completeness of reporting. For example in GAR health facilities such as Korle-Bu Teaching Hospital (KATH) and its affiliate institutions, 37 Military Hospital, Ghana Police Service hospital and University of Ghana hospital do not report their data to DHIMS2. The above-mentioned health institutions are high capacity facilities that provide diverse health services to many patients in the region. Therefore, continuous exclusion of their data from the national database would have a significant effect and impact on the quality and usability of DHIMS2 data for public health decision making especially for GAR. Furthermore, continuous non-reporting of data by such health facilities across the country into the national database would compromise the quality of DHIMS2 data 185 University of Ghana http://ugspace.ug.edu.gh which has far reaching implications for evidence-based small area public health policy decision-making for Ghana. It should also be noted that the above statistics may not reflect the extent of data incompleteness for the entire country. The situation for GAR may reflect what pertains in more urbanised regions like Ashanti and Northern region where a lot of private health facilities and tertiary health institutions exist which are the main causes of incomplete coverage (JSI, 2013). With over 90% completeness of indicator data (Amoakoh-Coleman et al., 2015), data from DHIMS2 database would attain a very high quality if steps are taken by the Ministry of Health to re-structure DHIMS2 to be directly under MOH to enable the tertiary health facilities and other health institutions owned by autonomous health institutions to submit their health services data to the national database. To take full advantage of Information Communication Technology to improve health data collection, storage and retrieval, MOH should constitute an Electronic Health Data Board. This board would be responsible for electronic data policies, health data standards and interoperability of electronic health systems. These will go a long way to reduce, if not eliminate, most of the current data quality issues affecting the national database (DHIMS2). It should be noted however, that manipulations (i.e. imputation of missing data) is necessary in the presence of significant data incompleteness for generating a meaningful health indicator but can significantly undermine the reliability of the resulting statistics. 5.1.2. Determining the external consistency of facility-based (DHIMS2) dataset The check for consistency calculated the ratio of the reported values of a two-week average of HFV due to malaria or diarrhoea from DHIMS2 and a two-week average of self-reported HFV due to malaria or diarrhoea from 2014 GDHS. This indicator looked at the percentage 186 University of Ghana http://ugspace.ug.edu.gh of sub-groups with at least 33% difference between DHIMS2 ratio and the GDHS ratio, an indication of inconsistency (WHO, 2014). It was generally observed that for all the subgroups of both HFV due to malaria and diarrhoea, GDHS estimates were much higher than DHIMS2 estimates resulting in DHIMS2 to GDHS differences of more than 33% (see Table 4.7 and 4.8). For example, the overall ratio of DHIMS2 to GDHS HFV totals for those under 5 years for malaria was 0.56. For diarrhoea, cross-checking estimates from the two data sources showed a ratio of DHIMS2 to GDHS of 0.20 for under 5 year old who visited a medical facility for diarrhoea treatment. These ratios translates to 44% and 80% differences between DHIMS2 and GDHS malaria and diarrhoea HFV totals for under 5 years respectively. The differences seen in these comparisons are at variance with WHO guidelines that limits this difference to less than 33% (WHO, 2014). These results suggest a significant underreporting of health-facility-based data (DHIMS2), the more so when the two datasets were measuring the same indicator; the two-week average (from September to December 2014) of HFV by children <5 years for malaria and diarrhoea, should be roughly equal to the number of GDHS reported HFV by children <5 years for malaria and diarrhoea during the two weeks preceding the survey (within September to December 2014). A similar study to assess health facility data quality conducted in Uganda (WHO, 2011a) by comparing Health Management Information System (HIMS) and Demographic & Health Survey (DHS) coverage rates for Diphtheria-tetanus-pertussis (DTP3) immunization, institutional deliveries and four or more antenatal care (ANC4+) visits revealed varying results. The DTP3 coverage from the HMIS reported data for children <1 year and results from the Uganda DHS in 2006-07 showed a gap of 25% between the DHS results (based on card and recall) and the HMIS in 2005 which suggests reporting bias when based on card or over-reporting in the HMIS. In addition, the comparison of annual estimates generated by the HMIS for the indicator on ANC4+ and DHS results for three years preceding the 2005-6 showed good consistency. However, the 187 University of Ghana http://ugspace.ug.edu.gh institutional delivery rates in the DHS report were higher than those reported by the HMIS (WHO, 2011a). Any such discrepancies between two datasets as observed between DHIMS2 and GDHS data for GAR or HIMS vrs DHS data in Uganda could be attributed to poor data quality of either dataset or both (WHO, 2014). First and foremost, household surveys have their own data quality challenges and if there are systematic problems, then the household survey-based coverage estimate can also fail to represent the true population value (WHO, 2014). Secondly, household surveys are usually based on a sample but not the entire population and therefore have confidence intervals for their estimates which must be considered in any comparison. Thirdly, confidence intervals for the proportions used in estimating the expected number of <5 years children who were treated at health facilities for malaria and diarrhoea in the population must be considered in the interpretation of the result. On the other hand, it is also very crucial to take note of the data reporting issues and the entire system that generated the annual cumulative HFVdata, which could be potential sources of underestimation of the facility-based data. This may have occurred because many HFV were inaccurately tallied, not all records were included in the process, or recorded data were summated before the end of the month and hence not all data included. In accounting for non-reporting, the annual cumulative HFV count for facilities that never reported data to DHIMS2 were imputed. This manipulation of data (i.e. imputation of missing data) could also be a probable source of underestimation of the facility-based data. Also, there may be instances in which missing data were not clearly differentiated from true zero values in facility reports. Missing entries may be assigned a value of 0, making it impossible to distinguish between a true zero value (no events occurred) from a missing value (events occurred but were not reported); this phenomenon can also bias the estimate leading to underestimation of the facility-based data. 188 University of Ghana http://ugspace.ug.edu.gh Contrary to what has been found by the check for consistency determined by percentage difference between DHIMS2 ratio and the GDHS ratio, a confidence Interval approach has shown that all the DHIMS2 estimates calculated for the sub-groups of under 5 HFV due to malaria or diarrhoea overlapped the expected confidence limits (see Table 4.6 and 4.7 column 3). This finding suggest that since there were large overlaps between DHIMS2 and GDHS estimates, then the observed differences are not significant (at the p <0.05 level). Hence, DHIMS2 estimates are consistent with GDHS estimates. However, critical analyses of the CIs have shown that the standard errors of the sample estimates were very high resulting in unreliable CIs i.e the precisions (CI) for the point estimates were unreasonably wide. Wide confidence intervals accentuates the unreliability of conclusions based on small samples or small numbers of events. This finding could be explained judging from the three things that impact the width of a confidence interval which include sample size, variability in the data and confidence interval (Sauro, 2012). Smaller sample sizes generate wider intervals and this happens because decreasing the sample size increases the width of confidence intervals, which increases the standard error. This is evident from Table 4.4 and 4.6 where majority of the sample sizes for the sub-groups were below 10 for children under 5 years who visited a medical facility for treatment of malaria or diarrhoea. Though this study was unable to test the distribution of the sample averages to determine their variability, there was the possibility that the distribution of the sample averages did not follow a normal curve as there was potential for over-reporting of the self-reported HFV by children’s carriers in DHS. As the confidence level decreases, the width of the interval decreases. With this data, even at 90% confidence level the interval is still wide. In practice, no researcher will want to set the confidence level below 90%, therefore this finding is highly uncertain. Investigation into all the issues raised above were beyond the scope of this work, hence further study is needed to unravel the causes of the discrepancies observed. However, due to 189 University of Ghana http://ugspace.ug.edu.gh the high uncertainty of the results produced by the confidence interval approach, this study adopted the results of the consistency checks based on difference between the ratio of the DHMS2 estimates to GDHS estimates. Considering age and sex-related HFV, the resultant pattern shows a reasonable concordance between DHIMS2 and GDHS estimates for the subgroups. Similar patterns were obtained for <1 year and 1-4 year age groups for malaria and diarrhoea HFV (Figure 4.5 and 4.6). These observed patterns shown by both GDHS data and facility-based data for malaria and diarrhoea HFV agree with the development principle of the immune response through neonatal, infant, adult and old age life (Simon et al., 2015). It has been proven in the literature that early defence against several infectious diseases earlier experienced by the mother is given by the passive IgG antibody transferred from the mother through the placenta and in milk. Once that diminishes, young children become particularly susceptible to infections (Metanat, 2015; Simon et al., 2015; Gilles, 1966) resulting in low infection rate for <1year children and higher infection rate for 1-4 years children. If the children survive this critical phase they acquire a relative immunity to the infection. Similarly, both GDHS and facility-based data were consistent about under-5 male children bias in the two-week average HFV due to malaria and diarrhoea during the period (September to December 2014). This result is also consistent with earlier studies whose conclusions seemed to be based on the assumption that there is a generalized tendency to give preferential treatment to boys over girls when seeking treatment at health facilities (Pandey et al., 2002; Kurz & Johnson-Welch, 1997; Ganatra & Hirve, 1994; Gupta, 1987). Aside the HFV differences, the two datasets share similar pattern for all other characteristic which is positive and made it necessary for further investigation to understand why one source was more robust than the than the other. 190 University of Ghana http://ugspace.ug.edu.gh Most of the analyses conducted involved health facility types as covariate. This study has shown that HFV due to both malaria and diarrhoea were largely influenced by health facility type. It was shown that hospitals were more attractive to malaria patients, followed by polyclinics, clinics and CHPS. Similar observations were made in the case of HFV due to diarrhoea except that polyclinics were more attractive than all other types of health facilities. These findings were not unexpected as per the structure of health services delivery in Ghana, regional and district hospitals provide the highest form of outpatient services and are more resourced in terms of health professionals – doctors, nurses, specialist and paramedics- infrastructure and commodities. At the sub-district level, both preventive and curative services are provided by the clinics and health centres. Often clinics / health centres are manned by one doctor or sometimes senior nurses or lab technicians and a few nurses (GHS, 2017). Basic preventive and curative services for minor ailments are being addressed at the community and household level with the introduction of the CHPS. In all the OLS models run, health facility types were significantly associated with both HFVR for malaria and diarrhoea. This finding will help health managers in allocation of resources especially human and space resources in order to control cost. It will also help in making suitable policies for increasing the profitability, allocation of financial resources and improvement of quality. 5.2. Do DHIMS2-derived pseudo-incidence rates show plausible patterns? 5.2.1. Climate variables and heath facility visits due to malaria and diarrhoea The influence of climatic conditions on infectious diseases such as malaria and diarrhoea have long been established by (Patz et al., 2003) and Sharma et al.,(2004). Both malaria and diarrhoea have usually been transmitted during seasons of high rainfall and for malaria especially during elevated temperature. Certain climatic environments are required, for both 191 University of Ghana http://ugspace.ug.edu.gh malaria and diarrhoea’s growth as well as for the survival of their vector mosquitoes and virus/bacteria respectively. Hence, the HFV data was analysed to determine how HFV due to malaria and diarrhoea vary with rainfall and temperature conditions of GAR in 2014. Ghana has a tropical climate with temperature varying with season and elevation. The 2014 climate data of GAR shows two rainy seasons which occur from April to July and from September to November. The mean maximum and minimum temperature are at their highest levels from January to May and from September to December respectively. The results have shown that HFV due to malaria was high during high rainfall and elevated temperature (Figures 4.11 (a) and 4.12(a). HFV due to malaria peaked at April to June and September to November, suggesting a direct relationship between HFV due to malaria and the climatic variables (rainfall and temperature). Similar to the findings of this study, several studies have found rainfall and temperature as predictors of malaria transmission in Ghana (Asare & Amekudzi, 2017; Darkoh et al., 2017; Arab et al., 2014; Dery et al., 2010). Though numerous factors influence malaria disease, climatic variables are seen to have more direct effects on the disease through their effect on vector development rates, transmission dynamics, human activity and behaviour. After cross analysis of climate data with HFV due to diarrhoea (Figure 4.10 (b) and 4.11 (b)), temperature seemed to have an indirect influence on HFV due to diarrhoea in GAR, while rainfall seemed to positively influence HFV due to diarrhoea in the region. As observed from Figure 4.10 (b), HFV due to diarrhoea was relatively high in January when rainfall was at its lowest but fall thereafter and this was quite abnormal. However, in May HFV due to diarrhoea rose rapidly and consistently peaking in August and steadily declined thereafter to December. July and August were the months immediately after the heavy rains in June (Figure 4.10 (b)) in Accra, when low areas are normally flooded leading to contamination of water sources and causing immerse sanitation problems. Water and sanitation play critical 192 University of Ghana http://ugspace.ug.edu.gh roles in the transmission of diarrhoeal disease, and several researchers have established a link between heavy rainfall and flooding, and subsequent outbreaks of diarrhoeal diseases (WHO, 2015a; Bhavnani et al., 2014; Carlton et al., 2013). This study observed an inverse relationship between HFV due to diarrhoea and maximum/minimum temperature. This finding is contrary to some earlier studies that established association between an increase in minimum and maximum temperature, and the rate at which diarrhoea affected people (Musengimana et al., 2016; Xu et al., 2014; Zhou et al., 2013) 5.2.2. Using routine facility-based data to measure health differentials This study investigated the use of health facility based data to measure variations in HFV due to malaria and diarrhoea by age, sex and other environmental factors to below district levels. Knowing and understanding of this variation is imperative, because it allows the discovery of elevated risk groups and careful targeting of intervention measures for optimal public health decisions (Carter et al., 2000). The outcome is intended to enhance the usability of facility- based health information. The results showed significant variations by age, sex, climatic factors and geography in the distribution of annual cumulative HFV due to malaria and diarrhoea in GAR in 2014. This study observed that HFV due to malaria and diarrhoea reported for GAR in 2014 were highest by private health facilities and lowest for quasi-public health facilities. This finding was consistent with earlier reports found in literature (JSI, 2013; Basu et al., 2012). The pattern of age-specific malaria morbidity has been well established by literature (Sharma et al., 2004; Snow & Marsh, 1998). Available literature shows that in high endemic regions, malaria morbidity peaks between the ages of 1–4 years and then declines rapidly (Pathak et al., 2012; Sharma et al., 2004; Snow & Marsh, 1998; Snow et al., 1997). Figure 4.7 in this study showed a u-shaped pattern age-specific HFV due to malaria and diarrhoea. This 193 University of Ghana http://ugspace.ug.edu.gh observation is consistent with a study conducted in 543 villages in central Ethiopia over a 4- year period to examine malaria distribution, a decrease in malaria incidence which fit a quadratic model was observed with increasing age (Yeshiwondim et al., 2009). With this study, there were high HFV due to malaria among children <5 years old, which began to decline after age four, indicating perhaps immunity Plasmodium spp. with increasing age. This finding agrees with the development principle of immune response through neonatal, infant, adult and old age life, where immunity against infectious diseases like malaria begins to develop after age 4 years (Metanat, 2015; Simon et al., 2015; Gilles, 1966). The study also observed that HFV due to malaria continue to decline from age 4 sharply until age 35-34 years (Figure 4.7). Thereafter there seemed to be age-dependent malaria morbidity stabilization between ages 20-49 when age-dependent morbidity levels began to increase possibly due to old-age. These observations have been generally consistent with findings from previous studies elsewhere most especially in developing countries (Pathak et al., 2012; Ndugwa et al., 2008; Abdullah et al., 2007). The pattern of age-specific HFVdue to diarrhoea morbidity was similar to that described for malaria. However, unlike malaria, there were studies that failed to establish any maternal passive immunity against infant diarrhoea (Bourne et al., 2007; Crump et al., 2005; Citterio et al., 2003). Underlying differences between population sub-groups in susceptibility to diseases and different treatment-seeking behaviour patterns in these groups may be responsible for the observed sex difference of HFV due to malaria and diarrhoea among districts and age groups. Age groups 50-59 years and 70+ years show statistically significant age dependent sex difference in HFV due to malaria (see Appendix 1). In both age groups, more females visited health facilities for malaria treatment than males (see Figure 4.7(a)). The literature has been widely divided on the age depended sex differences of malaria transmissions. While some findings have found evidence in data to support the claim of age-sex difference in malaria 194 University of Ghana http://ugspace.ug.edu.gh morbidity, others do not. A study conducted by Pathak et al. (2012) to investigate whether clinical malaria in hypoendemic regions exhibits sex difference and whether this difference was age-dependent, concluded that clinical malaria exhibited adult male difference which is contrary to the findings of this study. Differences in malaria cause specific morbidity has been well established in the literature. For example, the rate of malaria infection is known to be higher in pregnant women because of their decreased immunity, especially in their first and second parity (Duffy & Fried, 2005; Steketee et al., 2001). There is the perception that women are more likely to visit health facilities with respect to reproductive issues but will be clinically diagnosed with malaria. An earlier study by Müller et al. (1998) found that adolescent (10-19 years) and adult (20-40 years) women were more likely than similarly aged men to walk long distances to obtain malaria treatment at a clinic. Type of occupation could also play a significant role in sex-bias in malaria infection. For example, men have a greater occupational risk of contracting malaria than women if they work in places such as fields, forest, mines etc. (Reuben, 1993). Likewise men who get up before dawn to perform household chores may also be exposed to mosquitoes and consequently to malaria infection (Vlassoff & Manderson, 1998). In 2013, HFVR for malaria per 1,000 population was 417 for Ghana with the highest rate recorded in Upper East region (700 per 1,000 populations) and the lowest in GAR being 174 visits per 1,000 population (GNA, 2015b). The rate of 220 visit per 1,000 population found by this study for GAR is comparable to 2013 rate. The above results indicate that there has been an increase of 46 per 1,000 population in HFVR for malaria from 2013 to 2014 in GAR. When the data was disaggregated to the district level (see Figures 4.9 and 4.10), statistically significant differences were observed among the districts for both HFVR for malaria and diarrhoea (see appendix 2). Ashaiman municipality topped the rest of the districts in GAR with 555 HFV due to malaria per 1,000 population which was more than twice the regional 195 University of Ghana http://ugspace.ug.edu.gh HFVR for malaria (i.e. 220 visits per 1,000 population). This was followed by Adentan and Kpone-Katamanso with rates of 515 and 412 visits per 1,000 population respectively which were also about twice the regional rate. What was noteworthy about the above-mentioned districts was that they were among the recently created districts in GAR and therefore lack adequate health infrastructure (see section 3.2 for profiles and Figure 4.14 for health facility distribution) and other amenities. The trend in diarrhoea visits from the regional level to catchment level was similar to that of malaria. At the regional level 53 persons out of every 1,000 persons visited a health facility due to diarrhoea in 2014. At the district level, Shai-Osudoku (the lead district in the region) recorded 118 HFV per 1,000 population which was more than twice the regional visit rate. This was followed by Ashaiman municipality (114 HFV per 1000 population) and La- Nkwantanang-Madina (113 HFV per 1000 population) which were all twice as much as has been recorded by the region. Further disaggregation of the annual cumulative HFVR for malaria to catchment level has increased the rates for some of the catchment areas above what has been recorded for the districts. Figure 4.14 showed some catchments with rates as high as between 600 to 900 malaria visits per 1,000 population. HFVR for diarrhoea has increased up to 630 visits per 1,000 population for some catchments areas because of the spatial disaggregation of the data. The catchment analysis has revealed several catchment areas (small areas) with statistically significant hotspots of HFVR for malaria and diarrhoea across GAR (see Figure 4.16 and Appendix 3). Some catchment areas recorded higher HFVR for malaria and diarrhoea than their districts which previously recorded low HFV rates. These catchment areas with very high HFVR for malaria and diarrhoea, were found in AMA, Ada East and LANKMA. This finding agrees with earlier researches which have shown that disaggregation improves the quality of disease mapping and highlights areas where attention is much needed (Shi et al., 196 University of Ghana http://ugspace.ug.edu.gh 2013; Vanwambeke et al., 2011; Luo et al., 2010). However, targeted follow-up studies are needed to better understand outpatient data in those facilities with high rates. In summary, further disaggregation of facility-based data revealed details of small areas with higher HFV due to malaria and diarrhoea. Districts with high HFVR for malaria and diarrhoea have been obscured by regional level HFVR, but revealed by disaggregating the data to district level. As well, catchments with high HFVR were masked by district rates and could only show-up when the data has been disaggregated. Hence, the distribution of the annual cumulative HFVR for malaria and diarrhoea tend to be highly varying according to the depth of spatial disaggregation of the data. This result for example, calls for focused follow-up activity to understand the outpatient data in greater detail data in districts such as Ashaiman, Adenta and Kpone-Katamanso and catchment areas such as Asamoah Clinic, Akai house and Kaneshie polyclinic. The result also provides motivation and evidence for further examination of morbidity occurrences at community level and offer an opportunity to seek explanation for the observed variations. It is important to state here that the high HFVR values in this study may not necessarily be interpreted as direct measure of actual risk of malaria and diarrhoea for the geographic areas of study. This is because the research study assumes that patients access health facilities closest to them, but in reality, patients may also access health facilities that are not closest to them due to many reasons which were not the focus of this study. 197 University of Ghana http://ugspace.ug.edu.gh 5.2.3. Plausibility of association between HFVR for malaria and diarrhoea and environmental /household risk factors Environmental/household variables from the census data were used to develop models to predict HFVR for malaria and diarrhoea reported visits for catchment areas in GAR. The predictors were first determined from literature and were extracted from the 2010 PHC data. The OLS result of the overall model for HFVR for diarrhoea (see Table 4.17) showed that none of the environmental/ household variables had a statistically significant relationship with the HFVR for diarrhoea. However, when the data was divided into imputed and observed, the OLS model run on observed data had a proportion of persons using toilet (sanitary) facility prone to diarrhoea variable to be statistically significant, suggesting that people in this group have a higher risk of diarrhoea infection. There is an overwhelming evidence in literature that established that poor toilet (sanitary) facility is a major risk factor of diarrhoea infection (Unicef, 2014; Roma & Pugh, 2012; Siziya et al., 2009; Huttly, 1990). Lack of proper sanitary facilities aids the transfer of bacteria and parasites found in human excreta which otherwise pollute water resources, soil and food which is a major cause of diarrhoea. The Adjusted R2 for the observed model has also improved (9%) over that of the primary model (6%), whilst the result from the imputed data made no difference to the primary model. There is no evidence to suggest that any of the variables extracted from the census data could be used to predict HFVR for diarrhoea at the catchment levels based on the primary model results; hence no further analysis was considered. In the case of malaria, the OLS result of the overall model for HFVR for malaria (see Table 4.14) showed that two of the environmental/ household variables have statistically significant relationships with the HFVR for malaria. These variables included proportion of persons living in walls prone to malaria and proportion of persons within the lowest wealth quintile (but in a negative direction). Interestingly, the two statistically significant variables in the 198 University of Ghana http://ugspace.ug.edu.gh overall model were contributed by both the observed and imputed OLS models (see Tables 4.15 and 4.16). The above findings suggest that there are plausible associations between HFVR (from DHIMS2) and census small area statistics that are consistent with published evidence. The study observed that the overall model could explain approximately 12% of variations in the HFVR for malaria. This means that the remaining 88% of the variations in the HFVR for malaria are explained by other factors that have not been captured by this model resulting in model misspecification. This revelation came as no surprise because the two main anti-vector approaches used in Ghana are Insecticide Treated Nets (ITNs) and Indoor Residual Spraying (IRS) which most often targeted poor communities with high mosquito exposures. Mosquito net use, which has not been captured by the census data, has been identified as one of the key predictors of malaria in Ghana (Nyarko & Cobblah, 2014; NMCP, 2013; Baume & Franca-Koh, 2011; Osei-Kwakye et al., 2010). The proportion of persons living in walls prone to malaria was one of the two household factors that has a statistically significant relationship with HFVR for malaria in this study. This result corroborates findings from other studies that show associations between clinical malaria incidence and housing quality. There is an overwhelming evidence in the literature that established that housing characteristics such as composition of house wall (i.e. bricks, wood, leaves and others) is a major risk factor of malaria infection (Sharma et al., 2015; Liu et al., 2014; Guthmann et al., 2001a; Ghebreyesus et al., 2000). Though many of the household risk factors of clinical malaria have been linked to low socio-economic status, studies have shown that they have independent effects on malaria infection. For example, cracked or deteriorating mud wall (Coleman et al., 2010), traditional huts (Hiscox et al., 2013; Wolff et al., 2001) and open eaves (Lwetoijera et al., 2013; Kirby et al., 2008) as composition of housing walls that have independent effect on malaria infection. There is 199 University of Ghana http://ugspace.ug.edu.gh evidence for improving housing to reduce malaria (Tusting et al., 2015; Wanzirah et al., 2015; Hackett & Missiroli, 1932). The results of this study showed an inverse relationship between the proportion of persons within the lowest wealth quintile (i.e. socio-economic status) and HFVR for malaria, indicating that the rich/elite in society were more prone to malaria compared to individuals living below the poverty line. There has been mixed evidence linking malaria incidence and socio-economic status in literature (Worrall et al., 2005). An earlier study conducted in Southeast Nigeria found that malaria was more common in richer households (Onwujekwe et al., 2008). Several reasons may have accounted for this observation. Firstly, malaria has been described as a disease of poverty because it is less prevalent among the rich than the poor who lack education, live in rural areas and lack proper sanitation. In Ghana, the two main malaria interventions such as distribution of Insecticide Treated Nets(ITN) or Long-Lasting Insecticide Treated Nets (LLITN) and Indoor Residual Spraying (IRS) target poor people and deprived communities (Nyarko & Cobblah, 2014). As a result of the success of these malaria control interventions, the burden of the diseases has reduced in those communities over the years (WHO, 2007). Secondly, by the mere fact that the poor admit that it will be extremely difficult to pay for full malaria treatment, the inherent zeal to reduce to the barest minimum the cost of treatment may propel them to fully enrol in any available malaria prevention programmes resulting in fewer number of poor people having malaria. Thirdly, malaria immunity is dependent on the period of exposure to the malaria parasite. Most poor people may have been subjected to series of mosquito bites over a period of time because most of them live in malaria prone environments (Teklehaimanot & Paola Mejia, 2008; Organization & ESPD, 2005) and may be immune over time, compared to well to-do individuals in society who may have few mosquito bites over time and are therefore susceptible to malaria. Fourthly, differentials in HFV due to malaria reporting between the rich and the poor is likely 200 University of Ghana http://ugspace.ug.edu.gh to skew this finding. The rich are more likely to report at the hospital when they have malaria than the poor because they could afford to pay their bills (Onwujekwe et al., 2008) . The OLS result is able to establish an inverse relationship between malaria and socio-economic status in GAR, but fail to determine the actual communities where these associations existed. Hence the need for location specific analysis to establish this. 5.2.4. Spatial heterogeneities in the relationship between HFVR for malaria and household risk factors The study explored the spatial heterogeneity in relationships between HFVR for malaria and household risk factors using GWR. This analysis provides a further and better understanding of the relationship between HFVR for malaria and household risk factors and can be very useful to public health managers for targeting interventions to specific geographical locations (Mwesigwa et al., 2017; Homan et al., 2016; Lin & Wen, 2011; Bousema et al., 2010) . Ghana has chalked a lot of successes in reducing malaria through strategies such as the use of LLINs and IRS. It is believed that the application of the above interventions will be more effective if complemented by health facility data and novel tools such as geospatial techniques which provides opportunities to targeting interventions at smaller and well defined populations (Chaccour et al., 2013; Cringoli et al., 2013; Pullan et al., 2012; Alonso et al., 2011). Determination of risk factors for malaria is not new. While doing so, traditional epidemiological methods more often ignore the spatial heterogeneity of the underlying risk factors of the diseases. This study explored this dimension with the available data and tools. In exploring the spatial heterogeneity of the relationship between HFVR for malaria and household factors, the factors that were directly associated with malaria risk and statistically significant by the prior OLS model were used. These factors included proportion of persons living in walls prone to malaria and proportion of population within low wealth quintile level. 201 University of Ghana http://ugspace.ug.edu.gh The analyses have shown that the concentration of the influence of the two independent variables on HFVR for malaria were different among catchment areas and across GAR with some unexpected results. Hence, suggesting an existence of geographical heterogeneity in the relationships between HFVR for malaria and the risk factors (proportion of persons living in walls prone to malaria and proportion of persons within the lowest wealth quintile). These have been demonstrated by the various analyses shown by Figures 4.25 to 4.29. Spatial heterogeneity of HFVR for malaria was observed to be within short distances between catchment areas, perhaps because some of the facilities were closer to each other or such facilities were influenced by some other risk factors. Nevertheless, data system factors such as treatment-seeking behaviour and healthcare attendance patterns which have not been considered in this study can equally be responsible for the varying spatial pattern observed. The results were consistent with earlier studies that have shown that malaria transmission is generally heterogeneous across endemic regions with clustering of infection in small areas (Mwesigwa et al., 2017; Bousema et al., 2010; Kreuels et al., 2008). This result gives more insight to the policy maker on how to strategize to prevent malaria in specific geographical locations. Spatial distribution of proportion of persons living in walls prone to malaria have shown catchment areas in districts such as Ada West, Ada East, Ningo Prampram, Ga South (rural and semi-urban districts) and parts of AMA with very high proportions of persons living in walls prone to malaria and low proportions in other districts such as Ga West, Ga East, LEKMA, Adenta, La Dade-Kotopon and some parts of AMA. These observations reflected the housing composition usually associated with rural and urban settlements in Ghana (GSS, 2012) Rural districts are generally associated with poor housing conditions and vice-versa. AMA though an urban district, has parts inhabited by slum settlements even within planned settlements. A higher proportion of persons living in walls prone malaria were associated 202 University of Ghana http://ugspace.ug.edu.gh with a higher HFVR for malaria in the OLS model, which agrees with previous findings of many studies in both urban and rural areas (Guthmann et al., 2001b; Ghebreyesus et al., 2000). However, with GWR model, in some areas a higher percentage of persons living in walls prone to malaria were not necessarily accompanied by increase in HFVR for malaria (see Figure 4.25 to 4.29). Even though people may be living in walls prone to malaria, they might deliberately be targeted by the Long Lasting Insecticide Net (LLIN) and Indoor Residual Spray (IRS) interventions undertaken by NMCP and MOH for which reason they have less visit to health facilities due to malaria. Otherwise, they might not have the funds to pay their medicals bills and may resort to alternate means of treatment rather than visiting health facilities due to malaria. Earlier studies have shown that malaria infection heterogeneities were usually influenced by several risk factors including socio-economic, demographic, household/environmental factors (Adiamah et al., 1993; Schofield & White, 1984).which must all be considered when planning for an intervention. This is where data and geospatial methods used to explore geographically varying factors are very critical. Nevertheless, improving the housing situation in both urban and rural districts of GAR can be one of the many ways to reduce high malaria- related visits to heath facilities in GAR. The “one jacket fit all” inverse relationship shown by OLS model between proportion of persons within the lowest wealth quintile and HFVR for malaria across GAR has been improved by the GWR model. The GWR model gave a more detailed pattern of specific catchment areas where the proportion of person within low wealth quintile level showed a negative association with HFVR for malaria. In other locations a positive association was observed between the two variables( see Figure 4.29 (c)). A negative association between these two variables were observed in catchment areas in AMA, La Dade-Kotopon, LEKMA and parts of Ga South, Ga East, Adenta, LANKMA and TMA (these are mainly urban catchments shown by blue to yellow colours in Figure 4.29 (c) ), with an increased HFVR for 203 University of Ghana http://ugspace.ug.edu.gh malaria. The catchment areas in the districts such as Ada East, Ada, West, Shai-Osudoku, Ningo-Prampram, Ga South, Ga West and northern parts of Adentan and LANKMA (mainly rural catchment areas shown by red and brown colours) on the other hand showed a positive association between HFVR for malaria and the proportion of persons within low quintile level (see Figure 4.29 (c)). For the catchment areas in those urban districts, this result suggests that since poverty has often been linked with the risk of malaria, urban catchment areas were not usually priorities for the core malaria interventions such as LLIN and IRS by the NMCP and MOH. As a result those areas have not received much attention and therefore increasing the rate HFV due malaria in those locations. Another probable reason for the negative relationship of the proportion of persons within low quintile level to HFVR for malaria association of the urban catchment areas could be due to better schooling leading to more awareness of the consequences of not assessing healthcare on time, easy accessibility to health facilities and availability of funds to pay for health bills which enable most people to visit the health facilities whenever they become infected with malaria as compared to persons in the rural catchment areas. The above results mean a lot in relation to any future malaria intervention programme to be carried out in GAR. The GWR model demonstrated that it is not appropriate to use “one jacket fit all” approach in determining the risk factors of a diseases like malaria with progressively heterogeneous nature of transmission. Hence, spatial heterogeneity of the underlying risk factors of diseases in any epidemiological investigations must not be ignored as it holds the keys to a more effective disease control and eradication programme. This study provides insight into the spatial heterogeneity of malaria incidences using geospatial tools and readily available health facility and PHC data. 204 University of Ghana http://ugspace.ug.edu.gh 5.3. Data system and methodological challenges associated with analysis of outpatient data 5.3.1. Overview This study was affected by several data system and methodological challenges. Table 5.1 shows an overview of the challenges and the actions taken to address them. The follow up sections provide thorough discussion of the challenges and the analytical solutions used. Table 5.1: Data system and methodological challenges and actions taken Data system challenges Actions taken Completeness of facility reporting • Estimate completeness of facility reporting rate • Identify non-reporting facilities • Impute for the missing HFV Completeness of indicator data (zero / • Examine report for other diseses from same missing value) facility at same period Missing and ambiguous geographic • Coordinates of facility close by were collected coordinates of health facilities • Facility with missing coordinates dropped • Ambiguous coordinates were verified Effect of attractiveness among facility types • Facility types included in models to control for effects Census data quality and gaps • Projection of census data Challenges with small area boundaries • Using Thiessen polygon catchment areas • A clump of EAs consisting of between 50 to 260 individual EAs dropped • Catchment areas dropped from analyses GIS-based small area analysis tools and • Care must be taken when interpreting accompanying challenges (MAUP, ecological associations to avoid making inappropriate fallacy, GWR challenges) inferences 205 University of Ghana http://ugspace.ug.edu.gh 5.3.2. Data system challenges 5.3.2.1.Completeness of facility reporting challenges Completeness of reporting is a data quality problem and it is usually systemic and not specific to any one dataset (WHO, 2014). It is therefore important to know the reliability of facility-based data prior to its use for analysis so as to provide valuable information to support planning and monitoring. Two issues arose here which include completeness of facility reporting and missing data. Completeness of facility reporting rate was determined for GAR and for all the 16 districts. This was to determine the level of incompleteness of DHIMS2 data by districts and its effect on the region. The problem of missing data was as a result of non-reporting challenges confronting DHIMS2 data in GAR. A good understanding of coverage is required to facilitate the comparison of data (OECD, 2013). A situation where about half of health facilities are not reporting their data into the national database is a cause for concern. This study chose to impute for non-reporting facilities to enable comparison with GDHS data. Analysis of both observed and imputed data were conducted at some points to examine their single effects on the outcome variable. 5.3.2.2.Completeness of indicator data (zero / missing value) Completeness of missing data is an indicator which is used to measure the extent to which facilities include all reportable data elements (WHO, 2014). In a report missing data should be clearly differentiated from true zero values. For a true zero value, no reportable data element is present in the specified reporting period while a missing value specifies that reportable element is present but were not actually reported. The difficulty about DHIMS2 data is that missing entries are assigned a value of 0 which makes it difficult to distinguish between a true zero value (no data element) from a missing value (data elements present but 206 University of Ghana http://ugspace.ug.edu.gh were not reported). To address this challenge where a zero value was reported for a month for malaria or diarrhoea, the number of reported outpatient visits for other diseases were examined. If there were records for other diseases it was assumed that the facility has reported data for the month and the zero is considered as no HFV due to malaria or diarrhoea and if otherwise it was considered as missing data. In the dataset used the number of months for each of the diseases where zeros were considered as missing data were less than 1% and therefore were ignored. In an ideal situation, a true zero value is differentiate from a true missing value, and the missing values can be imputed. 5.3.2.3.Missing and ambiguous geographic coordinates of health facilities Geographic coordinates of 39 health facilities were missing from the original data obtained from different sources. A global Positioning System device was used to capture 19 of those facilities which are located nearby. Eventually, 20 health facilities were dropped from the study for not having geographic coordinates. For some health facilities, their locations as captured by the data sources were ambiguous and this phenomenon warranted a targeted fieldwork to validate those coordinates and the necessary corrections were made. More time and resources were needed to visit the location of the rest of the facilities to capture the coordinates. Accessibility to facility location was also one of the reasons why some of geographic coordinates of some facilities were not captured. 5.3.2.4.Effect of attractiveness among facility types Ghana’s health delivery structure is hierarchical, so services are integrated as one goes down the hierarchy of health structure from the national to the sub-district level. This structure reflected the establishment of regional and district hospitals, polyclinics in the districts, clinic/health centres in the sub-districts and CHPS in the communities. Because of this, 207 University of Ghana http://ugspace.ug.edu.gh different facilities of the hierarchy have varying levels of attractiveness. This phenomenon had significantly influenced the HFV due to malaria and diarrhoea and consequently the HFVR used for the models (see Figure 4.19). To control for the effect of facility attractiveness on HFVR in both OLS and GWR models, four dummies were created, each for a facility type and added to the models with CHPS as the reference facility. 5.3.2.5.Challenges with the use of census data in outpatient data analysis Census data constitutes one of the main sources of small area data (Smith, 2003). As mentioned before, in most countries including Ghana, censuses constitute the most comprehensive source of small-area data, covering a variety of population information (e.g., age, sex, marital status, race, income, education) and housing characteristics (e.g., dwelling type, type of wall, number of rooms, toilet facilities). Ghana chose to conduct a census of the entire population at regular intervals of every 10 years. However, like any developing country, due to financial and logistical constraint this plan gets distorted sometimes (Leete, 2001; Bair & Torrey, 1985). Users often question the reliability of census data from developing countries. Though censuses generally provide accurate and comprehensive data, they are most often not frequent and therefore create data gaps. Administrative records collected and kept by local governments provide small area inter-censured data. These records provide information on a host of variables such as births, deaths, building permits, drivers’ licenses, voter registration, income tax returns etc. These variables may be very useful for providing projections and estimates for tracking the performance of very important demographic and public health indicators (Smith et al., 2004). For example in UK, these indicators are pooled together to create a domain-level score, which can be used as a measure of the levels deprivation in small areas such as Super Output Area (Noble et al., 2006; Anttila & Wright, 2004). Some 208 University of Ghana http://ugspace.ug.edu.gh developed countries with very accurate administrative data systems even produce neighbourhood statistics based exclusively on administrative records (Longva et al., 1998) . In many developing countries like Ghana administrative records are inundated with quality issues such as incomplete coverage, credibility and availability (Cleland, 1996). What Ghana needs is a national spatial infra-structure and address or dwelling referencing system to underpin such data sets as done by some other countries (Committee, 2013; Nations, 2000). In small area analysis, availability of such data are very crucial for estimating population- based indicators such as morbidity, mortality and other population-based rates for public health decision making (Smith et al., 2004). 5.3.2.6.Challenges with small area boundaries When analysing data spatially, say at the regional or district level, much of the detail is lost across the larger boundary area. With the small areas data, analysis in certain areas can be viewed at community levels, but in most developing countries small area boundaries below district level do not exist in a digital form. This is a major challenge for GIS-based small area analysis. In Ghana, administrative boundaries were created for political convenience and for that matter could not match boundaries that would be ideal for health planning purposes. The smallest administrative unit for census data collection is a geographic area covered by one census representative known as Enumeration Area (EA). It is a standard geographical area for which census data is reported. Many major surveys such as DHS, MICS, Ghana Living Standard Survey (GLSS) and a host of others have used EAs for their design and as a unit of analysis (GDHS, 2015; GSS, 2014q, 2011). The problem of clearly defined area on the ground with many different data sets and at the resolution required is a challenge not only limited to developing countries (Thurstain‐Goodwin & Unwin, 2000). Currently, boundaries of majority of the EAs were not delineated into individual polygons which is one of the major 209 University of Ghana http://ugspace.ug.edu.gh challenges hindering small area health information analysis in Ghana. It becomes very difficult working with cluster of EAs which were not delineated into individual EAs. Out of the 16 districts in GAR only three (3) of them have a complete delineated districts which include AMA, LEKMA and La Dade-Kotopon. It was possible for Engstrom et al. (2013) to use the delineated EAs of AMA to define neighbourhood boundaries for urban health research in Accra. Population-based indicators depend on population count information for which EAs serve as reference points. One major challenge encountered by this study was the clumping of EAs as a result of non- delineation of EAs. Affected catchment areas were excluded from the final model as a result. District boundaries which are the next higher administrative unit are well delineated and very suitable for GIS-based small area analysis. Districts are the first level of most public health data records (e.g., morbidity, mortality, hospital discharge, etc.) aggregation. Data at district level are routinely gathered and can easily be incorporated in survey and other data collection efforts and population denominators and other demographic characteristics are often available. Ideally, other boundaries such as EA boundaries, town boundaries, census cluster boundaries that signify neighbourhood identities should have been incorporated into the district boundaries and finally into regional boundaries. The absence of this arrangement makes zoning of areas very problematic in GIS-based small area analysis (Weiss et al., 2007; Eagleson et al., 2002). Incompatible geographies (facility catchments and EAs in this case) are problematic for GIS analysis. With this study, there was the need to create small area boundaries – health facility catchment areas (using Voronoi/Thiessen polygon technique) independent of the administrative boundaries. This resulted in mismatch of relevant boundaries such as EA boundaries not matching catchment area boundaries and catchment area boundaries overlapping district boundaries. A point-in-polygon technique (Haines, 1994), using the EA 210 University of Ghana http://ugspace.ug.edu.gh centroids on top of Thiessen polygons, was used to determine population denominators for each catchment areas. This may result in dilution of some of the rates calculated which would eventually affect the plausibility of the outcome of the study. An alternative approach is to use an areal interpolation technique to perform polygon-to-polygon prediction to predict population values for catchment polygons (Kruase, n.d; ESRI, 2017b). 5.3.3. Methodological Challenges MAUP and ecological fallacy are both problems that particularly affect ecological studies (Rezaeian et al., 2007). Ecological fallacies and MAUP go hand in hand. The approaches to aggregating the data tend to project certain trends that are detrimental to true understanding (MAUP), and the conclusions drawn from the aggregated data can be erroneous (ecological fallacies). Diamond (2013) says that “every map is lying to you all the time”. According to (Openshaw, 1977) MAUP is defined as “a problem arising from the imposition of artificial units of spatial reporting on continuous geographic phenomena which results in the generation of artificial spatial patterns”. That is, same data producing different spatial patterns when aggregated in diverse ways. MAUP is both scale and zonation issues (Amrhein, 1995). Scale effect - showing major analytical variations depending on the size of units used - more pronounced for bigger units. Zonation/aggregation effect - showing major variations depending on how the study area is divided up, even at the same scale. Choropleth maps are based on areal data and the maps show colour scales, where each scale represents a discrete value or a range of values for groups of individuals located in the same region (Cressie, 2015; Bailey & Gatrell, 1995). The choropleth map in Figures 4.10 and 4.16 showed HFVR for malaria and diarrhoea by districts and catchment areas. The two figures were supposed to represent the same data but showed different spatial patterns and therefore 211 University of Ghana http://ugspace.ug.edu.gh were subjected to different interpretations. The methodological difference between the two figures was that Figure 4.10 was constructed using district polygons which were larger than polygons used to construct Figure 4.16 which were smaller catchment areas. Health facility data were collected at facility levels (higher resolution) but the data were aggregated at district level (lower resolution) giving rise to this problem. Because of the bias introduced, care was taken in the interpretation of the map outputs. The difficulty has always been that even if data were available for very small areas, it might not be clear how to aggregate them up. Hence, knowledge of the distribution of the health variables is needed to aid the interpretation of the result. Openshaw (1984) found that the optimal-zone design approach could also provide a general solution to the MAUP in the analysis of spatially aggregated data. The MAUP is usually related with “ecological fallacy”- a situation that occurs when it is inferred that data for areas under study can be applied to the individuals within those areas. As in this study, using district data to make inference for all facilities in a district may not be entirely correct. In this example, (Figure 4.15), dot density maps were produced showing the distribution of HFV due to malaria and diarrhoea relative to the catchment areas to identify spatial point patterns. The dot map of HFV due to malaria suggests that visits were concentrated in AMA, Ashaiman, southern part of Adenta and LANKMA. These geographical clusters were based only on procedures other than an explanatory theory and therefore do not provide any insight into why the clusters were occurring. In some other cases, associations found between variables at the district level have disappeared and some have reversed when recomputed with same data at the catchment level. This might have occurred because of the spatial boundaries used. All the above may lead to what is known as ecological fallacy. It is a well-documented problem with the interpretation of data from small 212 University of Ghana http://ugspace.ug.edu.gh area analyses (Schwartz, 1994; Piantadosi et al., 1988). Care must be taken when interpreting associations to avoid making inappropriate inferences. The GWR used in this study has its limitations, as a result of the statistical methods used (Páez et al., 2011). The statistical methods used by GWR is known to lack robust integrated statistical framework. The method depends on a collection of local spatial regressions, and hence unable to give precise inference. The varying coefficients were interpreted as an exploration and not as exact inference (Wheeler, 2014). One other problem, with regards to GWR is that the model produces local effects that can be inflated because of residual spatial autocorrelation and multi-collinearity . When regression residuals cluster because of spatial autocorrelation, the assumption of independence in a linear regression model is violated leading to inflation of coefficients. Usually, random error term for observations account for this. Recently, inflation of local coefficients has also been associated with local multi- collinearity (Wheeler & Tiefelsdorf, 2005). The effect on outcome variable can be overestimated if predictor variables locally show similar patterns. In this study GWR residuals have been tested for clustering and the result showed that spatial autocorrelation was absent. 5.4. Limitations of the study Though many of the challenges encountered in the course of this study have been dealt with in one way or the other, it will be inappropriate to ignore some very important limitations, otherwise the result of this study would have been more representative. Data system challenges posed the greatest limitations to this study. There were possible misreporting (as opposed to non-reporting which has been taken care of somehow) of HFV, 213 University of Ghana http://ugspace.ug.edu.gh misdiagnosis and zeros versus missing data issues which could not be resolved in this study. These challenges have the potential to inflate or under estimate the number of HFV recorded. Furthermore, multiple imputation by chain rule method used in this study to account for missing data by non-reporting health facilities is known to be an effective and robust method for dealing with missing data and it is expected to reduce any likely bias that may arise (Lodder, 2013; White et al., 2011) . However, imputation of missing data using any means is likely to introduce some error since imputed data cannot be the same as the original data. Nevertheless imputation is the optimal choice to have the full complement of the dataset for comparison to other data sources. Also, health facility catchments defined using Thiessen polygon technique assume that all persons are equally likely to use the health facility within the catchment– this is incorrect. In reality, health facility services utilization within catchments by the resident is not same, but is characterized by shades of use. Resident at the outer edge of a catchment, for example, have an increased probability to use services in adjacent catchment. However, under the circumstance it was the best choice for catchment determination. This is an ecological study (i.e. a study of population groups living in different areas) therefore the findings are affected by many of the generic limitations of an ecological study design which include MAUP (observed aggregated values will vary according to how we draw our area boundaries) and ecological fallacy ( occur in that factors that are associated with national disease rates may not be associated with disease in individuals). Concerning the distance matrix used for most of operations, though the best approach is to experiment with a variety of distance functions and see what works best for the data used, other distance metrics were not considered in this study except Euclidean distance because of its conformity to the physical concept of distance and also non-availability of data for use of other metrics. Result might be different when other distance matrix were used. 214 University of Ghana http://ugspace.ug.edu.gh Population under or over-enumeration in the census and the representation of population in the catchment areas was another limitation. EAs were not necessarily nested in the catchment areas leading to some of the EA boundaries overlapping the catchment boundaries resulting in over estimation or under estimation of population sizes of some of the catchment areas. Though the polygons were converted to centroids and laid on the catchment to determine the population of a catchment that was not enough to eliminate the impact of the overlapping. However, there was a likelihood of the problem to be minimized as a result of both negative and positive deviations for each catchment area. In some parts of GAR, ‘clumps’ of neighbouring EAs have being represented as single polygons which caused those polygons to be excluded from the analysis. 5.5. Review of study objectives This study sought to explore the usage of geospatial and statistical techniques to address issues associated with limited use of facility-based data to enhance its value for small area public health decision-making. The above objective was achieved by achieving the following specific objectives: 1. To determine the level of completeness in the coverage of health facility visits (outpatient data) in DHIMS2 for Greater Accra Region. This objective was a quality assessment of DHIMS2 data. It aims to know the reliability of outpatient data coverage estimates which was needed to understand how much confidence one can put in the health data presented. The study has found out that about 49% of the health facilities in GAR in 2014 did not report outpatient data to DHIMS2 and majority of these facilities were privately owned and quasi-government. Completeness of facility reporting rate for the region was 53%. Half of the districts in 215 University of Ghana http://ugspace.ug.edu.gh GAR recorded completeness of facility reporting rate between 23.08% and 72.73%; this is below 80% limit by WHO, suggesting poor facility reporting for the region and those districts. For the purpose of further analysis, the missing HFV for non-reporting health facilities could not be ignored and therefore were imputed for. The impute HFV for non-reporting facilities were 385, 538 for malaria and 101,435 for diarrhoea which constituted 44% and 43% respectively of the total HFV due to malaria (880,735) and diarrhoea (233,003) used in this study. 2. To determine the comparability between DHIMS2 under-5years old health facility visits (outpatient cases) and that of Ghana Demographic and Health Survey for Greater Accra Region. This objective was equally a quality assessment one. The purpose of this objective was to compare DHIMS2 data with different sources of population estimates (values are calculated differently) to see the level of similarity between the two sources. The higher the level of consistency between denominators from different sources, the more confidence can be instilled in the precision of the population estimates. This study has found out that far fewer HFV in total featured in DHIMS2 than one would estimate based on the GDHS, and that the difference (for under-5s as a whole) was highly unlikely to be due to sampling error. 3. To determine variability in health facility visits (outpatient cases) and how this variability is compared between district and community levels. The purpose of this objective was to evaluate the capability of facility-based data to achieve the point of small area analysis, that is, to focus on specific areas to reveal differences among populations or small areas within a larger statistical pattern. This study has revealed significant difference between percentage of males and females HFV due to malaria for the age groups 50 – 59 years and 70+ years, and a borderline 216 University of Ghana http://ugspace.ug.edu.gh significant difference was observed in the percentage HFV between males and females for ages 70 years and above for diarrhoea. The study also showed significant spatial variations in HFVR for malaria and diarrhoea at the district and catchment levels. District level variations in HFVR was between 90 HFV per 1,000 population to 577 HFV per 1,000 population, and 16 HFV per 1,000 population to 118 HFV per 1,000 population for malaria and diarrhoea respectively. It was found that spatial variations in the HFVR for malaria and diarrhoea increased with further disaggregation of the data from the district levels to the catchment levels. Catchment level variations in HFVR was between 1 HFV per 1,000 population to 900 HFV per 1,000 population, and 1 HFV per 1,000 population to 636 HFV per 1,000 population for malaria and diarrhoea respectively. Disaggregation of facility-based data revealed detailed small areas with higher HFVR for malaria and diarrhoea which were obscured at the district level. The catchment analysis has revealed several catchments areas (small areas) with statistically significant hotspots of HFVR for malaria and diarrhoea across GAR. However, it is only possible to do this for a sub-set of Greater Accra at the moment because of data system problems like EA clumping etc. 4. To determine plausible spatial relationships between facility visits (outpatient data) and environmental / household factors extracted from census. This objective sought to compare the statistics estimated from facility-based data and environmental / household situations of a number of small areas. This was to enable one identify causes or contributing factors to a condition which is one of the prime objectives of small area analysis. The study has revealed that both HFV due to malaria and diarrhoea have usually been transmitted during high rainfall and for malaria especially during elevated temperature which were consistent with what is known in literature. Type of health facility was found to influence HFV by patients in 217 University of Ghana http://ugspace.ug.edu.gh GAR. Hospitals and polyclinics were more attractive to both malaria and diarrhoea patients as compared to clinics and CHPS. At the global level, the study has established a statistically significant association between HFVR for malaria and household risk factors such as proportion of persons living in walls prone to malaria and proportion of persons within the lowest wealth quintile. No statistically significant relationship was found between HFVR for malaria and any of the household risk factors extracted from the census data. Majority of the spatial relationships established between HFVR for malaria and diarrhoea agreed with literature and support the evidence on ground of the plausibility of the relationships. 5. To establish geographical heterogeneities of the relationship between health facility visits (outpatient data) and environmental/household risk factors. This objective sought to investigate extent of using facility-based data to show diversity in HFVR / household factors relationship among catchment areas. Demonstration of spatial heterogeneity in data set can be seen as characteristic of reliable data for small area analysis. The GWR models run to test model performance and spatial heterogeneity in HFV for malaria in relation to household risk factors. The study revealed that there has been significant improvement in the GWR model performance over the OLS model. By comparing the fit of the OLS and GWR models, the Adjusted R-square for OLS was approximately 12% and the GWR Adjusted R- squared was 22%. The study has also shown significant geographical heterogeneity in the relationship between HFVR for malaria and household factors among catchment areas. Locally weighted 𝑅2 value (goodness of fit statistics) indicate how well the GWR model replicates the HFVR for malaria around proportion of persons living in walls prone to malaria and proportion of persons within the lowest wealth quintile. R2 value by catchment area vary between 10% to 46% with a mean value of 30%. Both 218 University of Ghana http://ugspace.ug.edu.gh the OLS and GWR tests have suggested that some important malaria risk factors were missing from the models which were expected since major determinants such as use of LLIN and IRS were not considered. However, this heterogeneity might not be attributed to only the risk factors though – it could be data system issues, e.g. spatial variation in facility attendance patterns or encoding of outpatient data by staff. 219 University of Ghana http://ugspace.ug.edu.gh CHAPTER SIX SUMMARY, CONCLUSION AND RECOMMENDATION 6.1. SUMMARY OF KEY FINDINGS 1) A total of 533 health facilities (both public and private) have been identified in GAR in 2014, offering outpatient services and out of which 271(51%) reported their data into DHIMS2 and the remaining 262 (49%) did not. The majority of non-reporting health facilities were private, quasi-government and autonomous health institutions. GAR recorded completeness of facility reporting rate of 53% and half of the districts in the region recorded completeness of facility reporting rate below the WHO limit of 85%, suggesting poor reporting. 2) The total number of HFV due to malaria and diarrhoea used in this study were 880, 735 and 233,003 respectively. Out of these numbers, the imputed malaria visits were 385,538 (44%) and that of diarrhoea were 101,435(43%). 3) Data consistency checks revealed a significant discrepancies between facility-based (DHIMS2) and Ghana Demographic & Health Survey (GDHS) data. However, both datasets were consistent with patterns shown by children <1 year and 1-4 years age groups HFV due to malaria and diarrhoea distributions. 4) Throughout the analyses, the type of health facility (i.e. hospitals, polyclinic, clinic/health centre and CHPS) was found to influence HFV due to malaria and diarrhoea. 5) Climatic conditions were found to be associated with HFV due to malaria and diarrhoea. HFV due to malaria and diarrhoea were found to vary among sex and age groupings. 6) HFVR for malaria and diarrhoea showed spatial variations at both the district and catchment area levels. Catchment analysis has revealed several small areas 220 University of Ghana http://ugspace.ug.edu.gh (catchments areas ) with statistically significant hotspots of HFVR for malaria and diarrhoea across GAR 7) GWR results indicated geographical heterogeneity in the relationship of HFVR for malaria with proportion of persons living in walls prone to malaria and proportion of persons within the lowest wealth quintile. 8) Geospatial and statistical tools and techiniques such as MICE, choropleth maps, Getis-Ord Gi* statistics, spatial autocorrelation (Moran’s I), OLS and GWR were used to generate plausible indicators/ information from limited use facility-based data for based small area evidence-based decision-making. 6.2. CONCLUSION The main objective of this study was to explore the usage of geospatial and statistical techniques to address issues associated with the limited use of facility-based data and to evaluate its value for small area evidence-based decision-making. Findings from this study have revealed a poor completeness of facility reporting rate in GAR and huge data gap in the facility-based data due to non-reporting by mostly private and quasi- government health facilities. Data inconsistency was observed between facility-based and GDHS data. However, determination of the cause of the inconsistency was beyond this study. With the incomplete data challenges, nonetheless, with the right methodological adaptations (for example imputation), routine facility-based data was able to generate plausible indicators /information comparable to those found in literature. The capability of the routine facility- based data has been enhanced further when it was linked with PHC data. This study revealed variations in HFV due to malaria and diarrhoea by age groups and by sex. Relationship between HFV due to malaria and diarrhoea, and climate variables (rainfall 221 University of Ghana http://ugspace.ug.edu.gh and temperature) were observed to generally agree with what is known in literature. Spatial variations were also observed at district and catchment levels in HFVR for malaria and diarrhoea. Disaggregation of the data to catchment levels have shown further variations in HFV which were obscured at district levels. At the catchment level, the study demonstrated how facility-based data linked with census data, could be used to identify variations in health outcomes, expose issues you wouldn’t otherwise see, identify populations most at need and provide evidence for targeted health interventions. This study modelled locally-varying risk factors for malaria and explored the spatial heterogeneity of the household risk-factors using GWR. GWR results indicated geographical heterogeneity in the relationship of HFVR for malaria with proportion of persons living in walls prone to malaria and proportion of persons within the lowest wealth quintile. The OLS model, by assuming stationarity of risk-factors fails to explore spatial components that can yield useful information and improve model fit. The GWR models showed considerable improvements in model performance over the OLS models. It offers public health managers and researchers a powerful tool through which to better understand local patterns of diseases epidemiology, population at risk, ultimately leading to the formulation of more effective and efficient public health policy. These varying effects can be used to focus interventions at the population most at need. This study demonstrates the feasibility of conducting GIS-based small area analysis using the most available datasets. It was possible to use a set of geospatial tools to link facility-based data to census data, show variations in HFVR rates for malaria and diarrhoea at district and health facility catchment levels (small areas), and explored the spatial heterogeneity between HFVR for malaria and household predictors at the local levels. 222 University of Ghana http://ugspace.ug.edu.gh The consistency of majority of the findings in this study with those reported in literature is a confirmation that DHIMS2 data can generate plausible indicators that can be used for evidence-based decision making. It is hoped that the findings of this study will serve as a basis for better data reporting by data managers of GHS. It will also encourage public health practitioners and other researchers to have more confidence in using facility-based and census data for effective health sector decision-making. 6.3. RECOMMENDATIONS 6.3.1. For policy 1. The MOH and GHS should persuade non-reporting health facilities, most especially the private and quasi-public health facilities to submit their monthly data to DHIMS2 to improve the coverage and completeness of DHIMS2 data 2. Health Facilities Regulatory Agency (HEFRA) should be empowered to renew licenses for only health facilities (most especially private ones) that report their data to the national database 3. MOH and GHS should ensure all national systems use routinely collected (DHIMS2) data for planning and managing their health services in order to improve the quality and use of DHIMS2 data. 4. Authorities of tertiary health institution should include Geographic Information System Application in Health in their curriculum to afford health information officers the opportunity to learn geospatial techniques to improve health care decision- making. 223 University of Ghana http://ugspace.ug.edu.gh 6.3.2. For future research 1. The research community in Ghana should conduct quality assessment studies into the inconsistencies observed between facility-based data (DHIMS2) and the Survey data (GDHS) even at national level to unravel the cause of the large discrepancies observed. In such a study, facility registers and folders could be used as the “gold standards” for the consistency check. 2. It is recommended that this study be repeated in other regions or for the entire country or in other developing countries to enhance the value of facility-based data for public health decision making 3. Further studies in urban malaria are needed to confirm or otherwise the inverse relationship between malaria and socio-economic status in urban GAR. 4. Public health researchers and practitioners at the facility, district and national levels must be sponsored to enhance their capacity in GIS through trainings and research. 6.4. CONTRIBUTION TO KNOWLEDGE 1. Application of geospatial techniques made it possible to use the facility-based data to calculate plausible population-based indicators to inform evidence based decision- making in smaller geographic areas. 2. Study demonstrated the possibility of integrating the most widely collected datasets (i.e. facility-based, census, DHS and environmental datasets) for evidence-based public health decision-making at any geographic level. 3. Study showed the feasibility of comparing estimates from facility-based data and GDHS data which hitherto has never been explored. ` 4. The methodological procedures adopted in this study are universal and can be applied to all datasets in the RHIS for decision making. 224 University of Ghana http://ugspace.ug.edu.gh APPENDICES Appendix 1: Result of Significance test for the observed male and female differences in HFV among age groups in GAR in 2014 % HFV due to malaria % HFV due to diarrhoea Age Group mean (95% CI) p-value mean (95% CI) p-value 0 - 4 years 0.7836 0.6147 Male 5.2 (3.4, 6.9) 1.8 (1.2, 2.5) Female 4.8 (3.1, 6.6) 1.6 (1.0, 2.2) 5 - 9 years 0.7851 0.9149 Male 3.0 (2.2, 3.9) 0.6 (0.4, 0.8) Female 3.2 (2.2, 4.2) 0.6 (0.4, 0.8) 10 - 14 years 0.3987 0.6641 Male 2.2 (1.7, 2.8) 0.5 (0.3, 0.6) Female 2.6 (1.8, 3.5) 0.5 (0.3, 0.6) 15 - 19 years 0.2013 0.1653 Male 2.1 (1.6, 2.7) 0.4 (0.3, 0.6) Female 2.9 (1.8, 3.9) 0.6 (0.4, 0.8) 20 - 34 years 0.0804 0.1043 Male 3.1 (1.9, 4.2) 0.8 (0.5, 1.0) Female 5.2 (2.9, 7.4) 1.3 (0.7, 1.9) 35 - 49 years 0.0844 0.1154 Male 2.1 (1.3, 2.8) 0.5 (0.3, 0.6) Female 3.4 (2.0, 4.7) 0.7 (0.4, 0.9) 50 - 59 years 0.047 * 0.0815 Male 1.1 (0.8, 1.5 0.2 (0.2, 0.3) Female 1.8 (1.2, 2.4) 0.3 (0.2, 0.4) 60 - 69 years 0.1302 0.2115 Male 0.8 (0.6, 1.1) 0.2 (0.1, 0.2) Female 1.2 (0.8, 1.5) 0.2 (0.1, 0.3) ≥ 70 years 0.0254* 0.0558 Male 0.7 (0.4, 0.9) 0.1 (0.1, 0.2) Female 1.1 (0.8, 1.4) 0.2 (0.1, 0.3) 225 University of Ghana http://ugspace.ug.edu.gh Appendix 2: Result of Signicance test for the observed male and female differences in HFV among districts in GAR in 2014 Malaria Diarrhoea District Sex Mean 95% CI P - value Mean 95% CI P - value Accra Metro male 1.9 (1.0 , 2.8) 0.6 (0.2 , 1.0) female 1.8 (1.3 , 2.4) 0.858 0.6 (0.3 , 0.8) 0.896 A da East male 1.7 (0.7 , 2.7) 0.75 (0.1 , 1.3) female 2.2 (1.4 , 3.0) 0.370 0.79 (0.2 , 1.3) 0 .91 A da West male 1.1 (0.3 , 2.0) 0.75 (0.1 , 1.3) female 1.3 (0.6 , 1.9) 0 .787 0.79 (0.2 , 1.3) 0.904 Adentan male 4.4 (2.0 , 6.9) 0.75 (0.1 , 1.3) female 5.9 (3.1 , 8.7) 0 .388 0.79 (0.2 , 1.3) 0 .591 A shaiman male 4.9 (2.2 , 7.6) 0.75 (0.1 , 1.3) female 7.0 (3.3 , 10.7) 0 .318 0.79 (0.2 , 1.3) 0 .388 Ga Central male 2.1 (1.4 , 2.9) 0.75 (0.1 , 1.3) female 2.4 (1.7 , 3.2) 0.515 0.79 (0.2 , 1.3) 0.879 Ga East male 0.8 (0.3 , 1.2) 0.75 (0.1 , 1.3) female 1.0 (0.5 , 1.5) 0.426 0.79 (0.2 , 1.3) 0 .714 G a West male 1.1 (0.6 , 1.6) 0.75 (0.1 , 1.3) female 1.3 (0.8 , 1.7) 0 .586 0.79 (0.2 , 1.3) 0 .589 Ga south male 2.1 (1.2 , 3.0) 0.75 (0.1 , 1.3) female 2.6 (1.6 , 3.5) 0 .398 0.79 (0.2 , 1.3) 0.939 Kpone- Katamanso male 2.8 (1.1 , 4.6) 0.75 (0.1 , 1.3) female 5.0 (2.8 , 7.2) 0.097 0.79 (0.2 , 1.3) 0.793 LANKMA male 2.3 (1.3 , 3.4) 0.75 (0.1 , 1.3) female 3.1 (1.7 , 4.6) 0.300 0.79 (0.2 , 1.3) 0 .403 L EKMA male 1.2 (0.7 , 1.7) 0.75 (0.1 , 1.3) female 1.5 (1.0 , 2.0 0 .284 0.79 (0.2 , 1.3) 0 .783 La Dade- Kotopon male 1.1 (0.8 , 1.5) 0.75 (0.1 , 1.3) female 1.4 (1.0 , 1.8) 0 .336 0.79 (0.2 , 1.3) 0.936 N ingo- Prampram male 1.1 (0.3 , 1.9) 0.75 (0.1 , 1.3) female 1.3 (0.7 , 1.9) 0.680 0.79 (0.2 , 1.3) 0.859 S hai-Osudoku male 3.6 (1.6 , 5.6) 0.75 (0.1 , 1.3) female 4.6 (2.7 , 6.5) 0 .402 0.79 (0.2 , 1.3) 0 .613 T ema Metro male 3.0 (2.0 , 4.0) 0.75 (0.1 , 1.3) female 3.4 (2.443 , 4.411) 0 .581 0.79 (0.2 , 1.3) 0.99 P-values in parentheses;*p<0.05,**p<0.01,***p<0.001 values were base d on t-test for comparing the mean percentage disease among males and females. CI : confidence interval. Appendix 3: Result of Significance test for the observed differences in HFV among districts in GAR in 2014 rate of OMC rate of ODC District Mean p-value Mean p-value Accra Met 60.6 200.0 Ada East 65.0 < 0.0001*** 177.1 < 0.0001*** Ada west 39.2 108.9 Adentan 66.4 515.0 Ashaiman 107.0 576.7 Ga Central 33.3 228.0 Ga East 14.9 90.4 Ga West 36.1 118.2 Ga south 24.2 193.4 Kpone-Kat 44.4 412.2 LANKMA 108.4 276.0 LEKMA 43.3 135.8 La-Dade-K 42.9 125.7 Ningo Par 21.2 112.1 226 University of Ghana http://ugspace.ug.edu.gh Shai-Osun 110.4 373.4 Tema Metro 53.3 325.1 P-values in parentheses; *p<0.05,**p<0.01,*** p<0.001 values were based on one-way ANOVA test Appendix 4: List of HFVR hotspots for malaria and diarrhoea at various confidence levels in GAR, 2014 No. Name of Catchment area Facility District Town Confidence Type Level (%) (a) Malaria 1 Pantang Hospital Hospital LANKMA Pantang 99 2 Madina Polyclinic Polyclinic LANKMA Madina 99 (Rawlings C) 3 Leonado Hospital Hospital LANKMA Madina 99 4 Adenta Clinic Clinic/HC Adenta Adenta Estate 99 5 Pentecoast hospital Hospital LANKMA Madina 99 6 Twumasiwa memorial Clinic/HC Adentan Twumasiwa 99 clinic 7 Nyaho Medical Centre Clinic/HC Accra Metro Airport Res. Area 99 8 Bob Freeman Clinic Clinic/HC Accra Metro Adabraka 99 9 C&J Medicare hospital Hospital Accra Metro Accra 99 10 Family Clinic Clinic/HC Accra Metro Accra 95 11 Kpone Health center Clinic/HC Kpone Katamanso Kpone 95 12 New Crystal Clinic Clinic/HC Kpone Katamanso Kakasunaka 95 13 Shai-Osudoku District Hospital Shai-Osudoku Dodowa 95 Hosital 14 Beata Memorial Clinic Clinic/HC Accra Metro Dzorwulu 95 15 Asamoah Clinic Clinic/HC Accra Metro Abelenkpe 95 16 Jesus Saves Clinic Clinic/HC Accra Metro Okaishie 95 17 St Andrews Catholic Clinic/HC Shai-Osudoku Kordiabe 95 Clinic 18 Clecom Memorial Hospital Accra Metro Dzorwulu 95 227 University of Ghana http://ugspace.ug.edu.gh hospital 19 Lebanon Community Hospital Kpone Katamanso Zenu/Lebanon 95 hospital 20 Achimota Hospital Hospital Accra Metro Achimota 95 21 Holy Cross Clinic & Mat. Clinic/HC Accra Metro Laterbiokorshie 95 Home 22 Dordoekope Health Clinic/HC Ada East Dordoekope 90 Centre 23 New Crystal Clinic Clinic/HC Kpone-Katamanso Kakasunaka 90 24 Tema General Hospital Hospital Tema Metro Tema 90 25 Tema Polyclinic Polyclinic Tema Metro Tema Comm 2 90 26 Mother Love Clinic Clinic/HC LANKMA Gbentanaa 90 27 The Thrust Clinic Clinic/HC Adentan Gbentanaa 90 28 Mayera Faase Health Clinic/HC Ga West Pokuase 90 Centre 29 Pokuase CHPS CHPS Ga West Pokuase 90 30 Akai House Clinic Clinic/HC Accra Metro Roman Ridge 90 31 The Thrust Hospital Hospital Accra metro Osu 90 32 Abora Clinic Clinic/HC Accra Metro Asylum Down 90 33 Eden Specialist hospital Hospital Accra Metro North Kanaeshie 90 (b) Diarrhoea 1 Princess Marie Louise Hospital Accra Metro Paladium 99 Hospital 2 Bob Freemen Clinic Clinic/HC Accra Metro Adabraka 99 3 Civil Service Clinic Clinic/HC Accra Metro Osu 99 4 Lekma Hospital Hospital Lekma Teshie 99 5 Madina Polyclinic Polyclinic LANKMA Madina 99 (Rawlings C) 6 Police Hospital Hospital La Dade-Kotopon Cantoment 99 7 Pentecoast hospital Hospital LANKMA Madina 99 8 Royal MMR Clinic Clinic/HC Ga East Dome 99 9 Duffor Health Center Clinic/HC Shai-Osudoku 95 10 St Andrews Catholic Clinic/HC Shai-Osudoku Kordiabe 95 Clinic 11 Lebanon Community Hospital Kpone Katamanso Zenu/Lebanon 95 hospital 12 Ga West Municipal Hospital Ga West Amasaman 95 Hospital 13 Mamprobi Polyclinic Polyclinic Accra Metro Mamprobi 90 14 Kpone Health center Clinic/HC Kpone Katamanso Kpone 90 15 Twumasiwa memorial Clinic/HC Adentan Twumasiwa 90 clinic 16 Achimota Hospital Hospital Accra Metro Achimota 90 17 Dam Side clinic Clinic/HC Ga Central Ablekuma 90 18 GPHA Clinic Clinic/HC Tema Metro Tema Comm 2 90 19 Ghana Post Clinic Clinic/HC Accra Metro Accra 90 20 Stadium Clinic Clinic/HC Accra Metro Osu 90 21 VRA Hospital Hospital Accra Metro Accra 90 22 Adabraka Polyclinic Clinic/HC Accra Metro Adabraka 90 228 University of Ghana http://ugspace.ug.edu.gh REFERENCES Abdullah, S., Adazu, K., Masanja, H., Diallo, D., Hodgson, A., Ilboudo-Sanogo, E., . . . Smith, T. (2007). Patterns of age-specific mortality in children in endemic areas of sub-Saharan Africa. AbouZahr, C., & Boerma, T. (2005). Health information systems: the foundations of public health. Bull World Health Organ, 83. Adaletey, D., Poppe, O., & Braa, J. (2013). Cloud computing for development — Improving the health information system in Ghana. Adams, I., Darko, D., & Accorsi, S. (2004). The burden of diseases: Breif overview of disease profile, service utilization patterns and health outcomes Bulletin of Health Information, 1. Adiamah, J., Koram, K., Thomson, M., Lindsay, S., Todd, J., & Greenwood, B. (1993). Entomological risk factors for severe malaria in a peri-urban area of The Gambia. Annals of Tropical Medicine & Parasitology, 87(5), 491-500. Adjei, E. (2003). Health sector reforms and health information in Ghana. Information development, 19(4), 256-264. Adjuik, M., Bagogoko. M., Binka, F., Coetzee, M., Cox, J., Craig. M., . . . Touré, Y. (1998). MARA/ARMA - Towards an Atlas of Malaria Risk in Africa: First Technical Repot of the MARA/ARMA Collaboration. Durbon. Agustin, A. L., JaehyunShin, Z., Curtis, A., Baezconde-Garbanati, L., & Cockburn, M. G. (n.a). Mapping late-stage breast cancer rates for community-based cancer screening interventions. Alegana, V. A., Wright, J., Bosco, C., Okiro, E. A., Atkinson, P. M., Snow, R. W., . . . Noor, A. M. (2017). Malaria prevalence metrics in low-and middle-income countries: an assessment of precision in nationally-representative surveys. Malaria Journal, 16(1), 475. Alexandrescu, R., O'Brien, S. J., Lyons, R. A., & Lecky, F. E. (2008). A proposed approach in defining population-based rates of major injury from a trauma registry dataset: delineation of hospital catchment areas BMC Health Services Research, 8(1), 80. Alho, J. (2001). Demographic techniques: small-area estimates and projections. Alonso, P., Brown, G., Arevalo-Herrera, M., Binka, F., Chitnis, C., Collins, F., . . . Levine, M. (2011). A research agenda to underpin malaria eradication. PLoS medicine, 8(1), e1000406. Amoakoh-Coleman, M., Kayode, G., Brown-Davies, C., Agyepong, I., Grobbee, D., Klipstein-Grobusch, K., & Ansah, E. (2015). Completeness and accuracy of data transfer of routine maternal health services data in the greater Accra region. BMC research notes, 8(1), 1. Amrhein, C. (1995). Searching for the elusive aggregation effect: evidence from statistical simulations. Environment and Planning A, 27(1), 105-119. Anselin, L. (2010). Thirty years of spatial econometrics. Papers in regional science, 89(1), 3-25. Anteneh, A., & Kumie, A. (2010). Assessment of the impact of latrine utilization on diarrhoeal diseases in the rural community of Hulet Ejju Enessie Woreda, East Gojjam Zone, Amhara Region. Ethiopian Journal of Health Development, 24(2). Antonio-Nkondjio, C., Fossog, B. T., Ndo, C., Djantio, B. M., Togouet, S. Z., Awono-Ambene, P., . . . Ranson, H. (2011). Anopheles gambiae distribution and insecticide resistance in the cities of Douala and Yaounde (Cameroon): influence of urban agriculture and pollution. Malaria Journal, 10(1), 154. Anttila, C., & Wright, G. (2004). Multiple Deprivation in New Deal for Communities Areas: Applying the Index of Multiple Deprivation 2004 to NDC Areas. Social Disadvantage Research Centre. Aqil, A., Lippeveld, T., & Hozumi, D. (2009). PRISM framework: a paradigm shift for designing, strengthening and evaluating routine health information systems. Health Policy and Planning, 24(3), 217-228. Arab, A., Jackson, M. C., & Kongoli, C. (2014). Modelling the effects of weather and climate on malaria distributions in West Africa. Malaria Journal, 13(1), 126. ArcGIS_Pro. (2017). Modeling spatial relationships. from http://pro.arcgis.com/en/pro-app/tool- reference/spatial-statistics/modeling-spatial-relationships.htm 229 University of Ghana http://ugspace.ug.edu.gh Asare, E. O., & Amekudzi, L. K. (2017). Assessing Climate Driven Malaria Variability in Ghana Using a Regional Scale Dynamical Model. Climate, 5(1), 20. Ashton, T. (2015). Measuring health system performance: A new approach to accountability and quality improvement in New Zealand. Health Policy, 119(8), 999-1004. Asiimwe, J. B., Jehopio, P., Atuhaire, L. K., & Mbonye, A. K. (2011). Examining small area estimation techniques for public health intervention: lessons from application to under-5 mortality data in Uganda. Journal of Public Health Policy, 32(1), 1-15. Askham, N., Cook, D., Doyle, M., Fereday, H., Gibson, M., Landbeck, U., . . . Schwarzenbach, J. (2013). The six primary dimensions for data quality assessment. DAMA UK Working Group, 432-435. Awoonor-Williams, J. K., Schmitt, M. L., Tiah, J., Ndago, J., Asuru, R., Bawah, A. A., & Phillips, J. F. (2016). A qualitative appraisal of stakeholder reactions to a tool for burden of disease–based health system budgeting in Ghana. Global health action, 9. Bailey, T., & Gatrell, A. (1995). Interactive spatial data analysis (Vol. 413): Longman Scientific & Technical Essex. Bair, R. R., & Torrey, B. B. (1985). The challenge of census taking in developing countries. Government information quarterly, 2(4), 433-452. Baker, L. C. (2001). Measuring competition in health care markets. Health services research, 36(1 Pt 2), 223. Basu, S., Andrews, J., Kishore, S., Panjabi, R., & Stuckler, D. (2012). Comparative performance of private and public healthcare systems in low-and middle-income countries: a systematic review. PLoS medicine, 9(6), e1001244. Baume, C. A., & Franca-Koh, A. C. (2011). Predictors of mosquito net use in Ghana. Malaria Journal, 10(1), 265. Belay, H., & Lippeveld, T. (2013). Inventory of PRISM Framework and Tools: Application of PRISM Tools and Interventions for Strengthening Routine Health Information System Performance. Measure Evaluation. Bellamy, C. (2001). The state of the world's children 2002: Unicef. Berke, O. (2004). Exploratory disease mapping: kriging the spatial risk function from regional count data. International Journal of Health Geographics, 3(1), 1. Bhavnani, D., Goldstick, J. E., Cevallos, W., Trueba, G., & Eisenberg, J. N. S. (2014). Impact of rainfall on diarrhoeal disease risk associated with unimproved water and sanitation. The American journal of tropical medicine and hygiene, 90(4), 705-711. Bhunia, G. S., Kesari, S., Chatterjee, N., Kumar, V., & Das, P. (2013). Spatial and temporal variation and hotspot detection of kala-azar disease in Vaishali district (Bihar), India. BMC infectious diseases, 13(1), 1. Bithell, J. (2000). A classification of disease mapping methods. Statistics in medicine, 19(17‐18), 2203-2215. Black, M., Ebener, S., Aguilar, P. N., Vidaurre, M., & El Morjani, Z. (2004). Using GIS to measure physical accessibility to health care. Paper presented at the International Health Users Conference. Black, R. E., Morris, S. S., & Bryce, J. (2003). Where and why are 10 million children dying every year? The Lancet, 361(9376), 2226-2234. Bødker, R., Akida, J., Shayo, D., Kisinza, W., Msangeni, H., Pedersen, E., & Lindsay, S. (2003). Relationship between altitude and intensity of malaria transmission in the Usambara Mountains, Tanzania. Journal of Medical Entomology, 40(5), 706-717. Bourne, L., Hendricks, M., Marais, D., & Eley, B. (2007). Addressing malnutrition in young children in South Africa. Setting the national context for paediatric food‐based dietary guidelines. Maternal & child nutrition, 3(4), 230-238. Bousema, T., Drakeley, C., Gesase, S., Hashim, R., Magesa, S., Mosha, F., . . . Msuya, E. (2010). Identification of hot spots of malaria transmission for targeted malaria control. The Journal of infectious diseases, 201(11), 1764-1774. 230 University of Ghana http://ugspace.ug.edu.gh Bousema, T., Griffin, J. T., Sauerwein, R. W., Smith, D. L., Churcher, T. S., Takken, W., . . . Gosling, R. (2012). Hitting hotspots: spatial targeting of malaria for control and elimination. PLoS medicine, 9(1), e1001165. Briggs, D. J., & Elliott, P. (1994). The use of geographical information systems in studies on environment and health. World health statistics quarterly. Rapport trimestriel de statistiques sanitaires mondiales, 48(2), 85-94. Brockerhoff, M. (1995). Child survival in big cities: the disadvantages of migrants. Social Science & Medicine, 40(10), 1371-1383. Brody, H., Rip, M. R., Vinten-Johansen, P., Paneth, N., & Rachman, S. (2000). Map-making and myth- making in Broad Street: the London cholera epidemic, 1854. The Lancet, 356(9223), 64-68. Brunsdon, C., Fotheringham, A., & Charlton, M. (2002). Geographically weighted summary statistics—a framework for localised exploratory data analysis. Computers, Environment and Urban Systems, 26(6), 501-524. Buescher, P. (1997). Problems with rates based on small numbers,[w:] Statistical Primer: North Carolina State Center for Health Statistics. Burgert, C. (2014). Spatial interpolation with Demographic and Health Survey data: Key considerations DHS Spatial Analysis Reports No. 9. Rockville, Maryland, USA: ICF International. Buuren, S., & Groothuis-Oudshoorn, K. (2011). mice: Multivariate imputation by chained equations in R. Journal of statistical software, 45(3). Carlson, B. L., Armitage, P., & Colton, T. (1998). Software for statistical analysis of sample survey data. Encyclopedia of biostatistics, 4160-4167. Carlton, E. J., Eisenberg, J. N. S., Goldstick, J., Cevallos, W., Trostle, J., & Levy, K. (2013). Heavy rainfall events and diarrhoea incidence: the role of social and environmental factors. American journal of epidemiology, kwt279. Carter, R., Mendis, K., & Roberts, D. (2000). Spatial targeting of interventions against malaria. Bulletin of the World Health Organization, 78(12), 1401-1411. Castro, M. C., Kanamori, S., Kannady, K., Mkude, S., Killeen, G. F., & Fillinger, U. (2010). The importance of drains for the larval development of lymphatic filariasis and malaria vectors in Dar es Salaam, United Republic of Tanzania. PLoS Negl Trop Dis, 4(5), e693. CDC. (2013). CDC in Ghana - Factsheet, 2013 CDC in Ghana, Factsheet Atlanta - USA: Centers for Disease Control and Prevention. Chaccour, C., Kobylinski, K., Bassat, Q., Bousema, T., Drakeley, C., Alonso, P., & Foy, B. (2013). Ivermectin to reduce malaria transmission: a research agenda for a promising new tool for elimination. Malaria Journal, 12(1), 153. Chae, Y. M., Kim, S. I., Lee, B. H., Choi, S. H., & Kim, I. S. (1994). Implementing health management information systems: measuring success in Korea's health centers. The International journal of health planning and management, 9(4), 341-348. Chaki, P. P., Govella, N. J., Shoo, B., Hemed, A., Tanner, M., Fillinger, U., & Killeen, G. F. (2009). Achieving high coverage of larval-stage mosquito surveillance: challenges for a community- based mosquito control programme in urban Dar es Salaam, Tanzania. Malaria Journal, 8(1), 311. Chambers, R., & Tzavidis, N. (2006). M-quantile models for small area estimation. Biometrika, 93(2), 255-268. Chandra, H. (2003). Overview of Small Area Estimation Techniques. Techniques of Estimation and Forecasting of Crop Production in India, 75-88. Chang, K.-T. (2015). Introduction to geographic information systems: McGraw-Hill Science/Engineering/Math. Citterio, C., Luzzago, C., Sala, M., Sironi, G., Gatti, P., Gaffuri, A., & Lanfranchi, P. (2003). Serological study of a population of alpine chamois (Rupicapra r rupicapra) affected by an outbreak of respiratory disease. The Veterinary Record, 153(19), 592-596. 231 University of Ghana http://ugspace.ug.edu.gh Cleland, J. (1996). Demographic data collection in less developed countries 1946–1996. Population Studies, 50(3), 433-450. Clements, A. C., Reid, H. L., Kelly, G. C., & Hay, S. I. (2013). Further shrinking the malaria map: how can geospatial science help to achieve malaria elimination? The Lancet infectious diseases, 13(8), 709-718. Cohen, M., Naylor, C., Basinski, A., Ferris, L., Llewellynthomas, H., & Williams, J. (1992). Small-area variations-what are they and what do they mean. Canadian Medical Association J., 146, 467- 470. Coleman, M., Coleman, M., Mabaso, M., Mabuza, A., Kok, G., Coetzee, M., & Durrheim, D. (2010). Household and microeconomic factors associated with malaria in Mpumalanga, South Africa. Transactions of the Royal Society of Tropical Medicine and Hygiene, 104(2), 143-147. Committee, F. G. D. (2013). National Spatial Data Infrastructure Strategic Plan 2014–2016. Reston, Virginia. Cressie, N. (1992). Smoothing regional maps using empirical Bayes predictors. Geographical Analysis, 24(1), 75-95. Cressie, N. (2015). Statistics for spatial data: John Wiley & Sons. Cringoli, G., Rinaldi, L., Albonico, M., Bergquist, R., & Utzinger, J. (2013). Geospatial (s) tools: integration of advanced epidemiological sampling and novel diagnostics. Geospatial health, 7(2), 399-404. Cromley, E., & McLafferty, S. (2002). GIS and Public Health. New York: The Guilford Press. Cromley, E. K., & McLafferty, S. L. (2012). GIS and public health (2nd Edition ed.): Guilford Press. Crump, J., Otieno, P., Slutsker, L., Keswick, B., Rosen, D., Hoekstra, R. M., . . . Luby, S. P. (2005). Household based treatment of drinking water with flocculant-disinfectant for preventing diarrhoea in areas with turbid source water in rural western Kenya: cluster randomised controlled trial. Bmj, 331(7515), 478. Da Silva, A., & Laprega, M. (2004). [Critical evaluation of the Primary Care Information System (SIAB) and its implementation in Ribeirao Preto, Sao Paulo, Brazil]. Cadernos de Saude Publica, 21(6), 1821-1828. Darkoh, E. L., Larbi, J. A., & Lawer, E. A. (2017). A Weather-Based Prediction Model of Malaria Prevalence in Amenfi West District, Ghana. Malaria research and treatment, 2017. Dean, T., Creese, A., & Prentice, T. (1999). The double burden: emerging epidemics and persistent problems. The World Health Report 1999, 13-27. Dery, D. B., Brown, C., Asante, K. P., Adams, M., Dosoo, D., Amenga-E.S, . . . Owusu-Agyei, S. (2010). Patterns and seasonality of malaria transmission in the forest-savannah transitional zones of Ghana. Malaria Journal, 9(1), 314. Devillea, P., Linard, C., Martine, S., Gilbert, M., Stevensf, F. R., Gaughan, A. E., . . . Tatemg, A. J. (2014). Dynamic population mapping using mobile phone data. PNAS, 111(45). Diamond, L. (2013). Geographic Data Assumptions: MAUP and Ecological Fallacies. Collective spatial learning from students and professionals. from http://giscollective.org/geographic-data- assumptions-maup-and-ecological-fallacies/ Diggle, P. J. (2000). Overview of statistical methods for disease mapping and its relationship to cluster detection. Spatial Epidemiology: Methods and Applications, 87-103. Dongus, S., Nyika, D., Kannady, K., Mtasiwa, D., Mshinda, H., Gosoniu, L., . . . Killeen, G. F. (2009). Urban agriculture and Anopheles habitats in Dar es Salaam, Tanzania. Geospatial health, 3(2), 189-210. Duffy, P., & Fried, M. (2005). Malaria in the pregnant woman Malaria: Drugs, Disease and Post- Genomic Biology (pp. 169-200): Springer. Eagleson, S., Escobar, F., & Williamson, I. (2002). Hierarchical spatial reasoning theory and GIS technology applied to the automated delineation of administrative boundaries. Computers, Environment and Urban Systems, 26(2), 185-200. 232 University of Ghana http://ugspace.ug.edu.gh Eck, J., Chainey, S., Cameron, J., Leitner, M., & Wilson, R. (2008). The Utility of Hotspot Mapping for Predicting. Security Journal, 4-28. Engstrom, R., Ofiesh, C., Rain, D., Jewell, H., & Weeks, J. (2013). Defining neighbourhood boundaries for urban health research in developing countries: A case study of Accra, Ghana. Journal of maps, 9(1), 36-42. Ernst, K. C., Adoka, S. O., Kowuor, D. O., Wilson, M. L., & John, C. C. (2006). Malaria hotspot areas in a highland Kenya site are consistent in epidemic and non-epidemic years and are associated with ecological factors. Malaria Journal, 5(1), 78. Ernst, K. C., Lindblade, K. A., Koech, D., Sumba, P. O., Kuwuor, D. O., John, C. C., & Wilson, M. L. (2009). Environmental, socio‐demographic and behavioural determinants of malaria risk in the western Kenyan highlands: a case–control study. Tropical Medicine & International Health, 14(10), 1258-1265. Eshete, W. B. (2009). A stepwise regression analysis on under-five diarrhoael morbidity prevalence in Nekemte town, western Ethiopia: maternal care giving and hygiene behavioral determinants. ESRI. (2009). Early Detection and Response to Infectious Disease. GIS Best Practices. ESRI. (2011). Create Thiessen Polygons (Analysis) (ArcInfo only). ArcGIS Desktop Help 9.3. from http://webhelp.esri.com/arcgisdesktop/9.3/index.cfm?id=1349&pid=1347&topicname=Crea te_Thiessen_Polygons_%28Analysis%29 ESRI. (2017a, February, 2017). Create Thiessen Polygons. ArcGIS Pro. from http://pro.arcgis.com/en/pro-app/tool-reference/analysis/create-thiessen-polygons.htm ESRI. (2017b). Using areal interpolation to perform polygon-to-polygon predictions. Geostatistical Analyst. from http://desktop.arcgis.com/en/arcmap/latest/extensions/geostatistical- analyst/using-areal-interpolation-to-predict-to-new-polygons.htm# Fotheringham, A. S., Brunsdon, C., & Charlton, M. (2003). Geographically weighted regression: John Wiley & Sons, Limited. Frenk, J., & Gómez-Dantés, O. (2011). The triple burden: disease in developing nations. Harvard International Review, 33(3), 36. Gachogu, P. W. (2003). Determinants of diarrhoea and Its treatment in Kenya. University of Nairobi. Ganatra, B., & Hirve, S. (1994). Male bias in health care utilization for under-fives in a rural community in western India. Bulletin of the World Health Organization, 72(1), 101. GDHS. (2015). 2014 Ghana Demographic and Health Survey Report. Rockville, Maryland, USA: Ghana Statistical Service (GSS), Ghana Health Service (GHS), and ICF International. Genser, B., Strina, A., Teles, C. A., Prado, M. S., & Barreto, M. L. (2006). Risk factors for childhood diarrhoea incidence: dynamic analysis of a longitudinal study. Epidemiology, 17(6), 658-667. Geraghty, E. M., Balsbaugh, T., Nuovo, J., & Tandon, S. (2010). Using Geographic Information Systems (GIS) to assess outcome disparities in patients with type 2 diabetes and hyperlipidemia. The Journal of the American Board of Family Medicine, 23(1), 88-96. Gething, P., Noor, A., Zurovac, D., Atkinson, P., Hay, S., Nixon, M., & Snow, R. (2004). Empirical modelling of government health service use by children with fevers in Kenya. Acta tropica, 91(3), 227-237. Gething, P. W., Patil, A. P., Smith, D. L., Guerra, C. A., Elyazar, I. R., Johnston, G. L., . . . Hay, S. I. (2011). A new world malaria map: Plasmodium falciparum endemicity in 2010. Malaria Journal, 10(1), 378. Getis, A., & Ord, J. K. (1992). The analysis of spatial association by use of distance statistics. Geographical Analysis, 24(3), 189-206. Ghebreyesus, T., Haile, M., Witten, K., Getachew, A., Yohannes, M., Lindsay, S., & Byass, P. (2000). Household risk factors for malaria among children in the Ethiopian highlands. Transactions of the Royal Society of Tropical Medicine and Hygiene, 94(1), 17-21. Ghosh, M., & Rao, J. (1994). Small area estimation: an appraisal. Statistical science, 55-76. 233 University of Ghana http://ugspace.ug.edu.gh GHS. (2017). Structure of Delivery of Services. 2017, from http://www.ghanahealthservice.org/ghs- subcategory.php?cid=2&scid=44 Gilles, H. M. (1966). Malaria in children. British medical journal, 2(5526), 1375. Gilmour, S. J. (2010). Identification of hospital catchment areas using clustering: an example from the NHS. Health services research, 45(2), 497-513. Gimbel, S., Micek, M., Lambdin, B., Lara, J., Karagianis, M., Cuembelo, F., . . . Sherr, K. (2011). An assessment of routine primary care health information system data quality in Sofala Province, Mozambique. Population Health Metrics, 9, 12-12. doi: 10.1186/1478-7954-9-12 Glaser, S. L. (1990). Spatial clustering of Hodgkin's disease in the San Francisco Bay area. American journal of epidemiology, 132(supp1), 167-177. GNA. (2015a). Diarrhoea, second leading cause of death among children, Ghana News Agency. Retrieved from http://www.ghananewsagency.org/health/diarrhoea-second-leading-cause- of-death-among-children--90840 GNA (Producer). (2015b). Malaria cases in Ghana on the rise. Retrieved from https://www.newsghana.com.gh/ghana-records-11-million-cases-opd-malaria-2013/ Godlee, F., Pakenham-Walsh, N., Ncayiyana, D., Cohen, B., & Packer, A. (2004). Can we achieve health information for all by 2015? The Lancet, 364(9430), 295-300. Gollogly, L. (2009). World health statistics 2009: World Health Organization. Gómez-Rubio, V., & López-Quılez, A. (2006). Empirical and Full Bayes estimators for disease mapping. Paper presented at the International workshop on spatio-temporal modelling (METMA3). Granell, C., Gould, M., Manso, M. A., & Bernabé, M. A. (2009). Spatial data infrastructures Handbook of Research on Geoinformatics (pp. 36-41): IGI Global. Graves, B. A. (2012). Geographical information systems and the analysis of health for rural populations. Online Journal of Rural Nursing and Health Care, 8(1), 11-12. Green, S. T., Small, M. J., & Casman, E. A. (2009). Determinants of national diarrhoeal disease burden: ACS Publications. Griffith, D. A. (1987). Spatial autocorrelation. A Primer (Washington, DC, Association of American Geographers). GSS. (2011). GHANA Multiple Indicator Cluster Survey with an enhanced Malaria Module and Biomarker 2011. Accra: Ghana Statistical Service. GSS. (2012). 2010 POPULATION & HOUSING CENSUS SUMMARY REPORT OF FINAL RESULTS. Accra, Ghana: Ghana Statistical Service. GSS. (2013). 2010 Population and Housing Cencus: Regional Analytical Report, Greater Accra Region. In S. N. Cudjoe, E. Sepah, J. K. Anarfi, S. K. Gaisie & N. Bentsi-Enchill (Eds.). Accra: Ghana Statistical Service. GSS. (2014a). 2010 Population & Housing: District Analytical Report - Accra Metropolotan. Accra: Ghana Statistical Service. GSS. (2014b). 2010 Population & Housing: District Analytical Report - ADA EAST DISTRICT. Accra: Ghana Statistical Service. GSS. (2014c). 2010 Population & Housing: District Analytical Report - ADA WEST DISTRICT. Accra: Ghana Statistical Service. GSS. (2014d). 2010 Population & Housing: District Analytical Report - ADENTAN MUNICIPALITY Accra: Ghana Statistical Service. GSS. (2014e). 2010 Population & Housing: District Analytical Report - ASHAIMAN MUNICIPALITY. Accra: Ghana Statistical Service. GSS. (2014f). 2010 Population & Housing: District Analytical Report - GA CENTRAL MUNICIPALITY. Accra: Ghana Statistical Service. 234 University of Ghana http://ugspace.ug.edu.gh GSS. (2014g). 2010 Population & Housing: District Analytical Report - GA EAST MUNICIPALITY. Accra: Ghana Statistical Service. GSS. (2014h). 2010 Population & Housing: District Analytical Report - GA SOUTH MUNICIPALITY. Accra: Ghana Statistical Service. GSS. (2014i). 2010 Population & Housing: District Analytical Report - GA WEST MUNICIPALITY. Accra: Ghana Statistical Service. GSS. (2014j). 2010 Population & Housing: District Analytical Report - KPONE-KATAMANSO DISTRICT. Accra: Ghana Statistical Service. GSS. (2014k). 2010 Population & Housing: District Analytical Report - LA DADE-KOTOPON MUNICIPALITY. Accra: Ghana Statistical Service. GSS. (2014l). 2010 Population & Housing: District Analytical Report - LA NKWANTANANG-MADINA MUNICIPALITY. Accra: Ghana Statistical Service. GSS. (2014m). 2010 Population & Housing: District Analytical Report - LEDZOKUKU-KROWOR MUNICIPALITY. Accra: Ghana Statistical Service. GSS. (2014n). 2010 Population & Housing: District Analytical Report - NINGO-PRAMPRAM MUNICIPALITY. Accra: Ghana Statistical Service. GSS. (2014o). 2010 Population & Housing: District Analytical Report - SHAI OSUDOKU DISTRICT. Accra: Ghana Statistical Services. GSS. (2014p). 2010 Population & Housing: District Analytical Report - TEMA METROPOLITAN. Accra: Ghana Statistical Service. GSS. (2014q). Ghana Living Standards Survey (GLSS6). Accra: Ghana Statistical Service. Gupta, M. D. (1987). Selective discrimination against female children in rural Punjab, India. Population and development review, 77-100. Guthmann, J.-P., Hall, A., Jaffar, S., Palacios, A., Lines, J., & Llanos-Cuentas, A. (2001a). Environmental risk factors for clinical malaria: a case-control study in the Grau region of Peru. Transactions of the Royal Society of Tropical Medicine and Hygiene, 95(6), 577-583. Guthmann, J.-P., Hall, A. J., Jaffar, S., Palacios, A., Lines, J., & Llanos-Cuentas, A. (2001b). Environmental risk factors for clinical malaria: a case-control study in the Grau region of Peru. Transactions of the Royal Society of Tropical Medicine and Hygiene, 95(6), 577-583. Hackett, L., & Missiroli, A. (1932). Housing as a factor in malaria control. Transactions of the Royal Society of Tropical Medicine and Hygiene, 26(1), 65-72. Haggard, L. M., Shah, G., Stat, M., Rolfs, R. T., Haggard, L., Shah, G., & Rolfs, R. (1998). Assessing community health status: Establishing geographic areas for small area analysis in Utah. Utah’s health: An annual review, 18-35. Haines, E. (1994). Point in polygon strategies. Graphics gems IV, 994, 24-26. Hart, T., & Zandbergen, P. (2014). Kernel density estimation and hotspot mapping: examining the influence of interpolation method, grid cell size, and bandwidth on crime forecasting. Policing: An International Journal of Police Strategies & Management, 37(2), 305-323. Hay, S. I., Guerra, C. A., Tatem, A. J., Atkinson, P. M., & Snow, R. W. (2005). Opinion—tropical infectious diseases: Urbanization, malaria transmission and disease burden in Africa. Nature Reviews Microbiology, 3(1), 81. Heady, P., & Ralphs, M. (2005). EURAREA: an overview of the project and its findings. Paper presented at the SAE2005 Conference. Heeks, R. (2006). Health information systems: Failure, success and improvisation. International journal of medical informatics, 75(2), 125-137. Hiscox, A., Khammanithong, P., Kaul, S., Sananikhom, P., Luthi, R., Hill, N., . . . Lindsay, S. (2013). Risk factors for mosquito house entry in the Lao PDR. PLoS ONE, 8(5), e62769. Homan, T., Maire, N., Hiscox, A., Pasquale, A., Kiche, I., Onoka, K., . . . Smith, T. A. (2016). Spatially variable risk factors for malaria in a geographically heterogeneous landscape, western Kenya: an explorative study. Malaria Journal, 15(1), 1. 235 University of Ghana http://ugspace.ug.edu.gh Hotchkiss, D. R., Diana, M. L., & Fleischman Foreit, K. G. (2012). How can routine health information systems improve health systems functioning in low-and middle-income countries? Assessing the evidence base Health Information Technology in the International Context (pp. 25-58): Emerald Group Publishing Limited. Huttly, S. (1990). The impact of inadequate sanitary conditions on health in developing countries. World health statistics quarterly. Rapport trimestriel de statistiques sanitaires mondiales, 43(3), 118-126. Imbahale, S. S., Fillinger, U., Githeko, A., Mukabana, W. R., & Takken, W. (2010). An exploratory survey of malaria prevalence and people's knowledge, attitudes and practices of mosquito larval source management for malaria control in western Kenya. Acta tropica, 115(3), 248- 256. Joseph, H., Moloney, J., Maiava, F., McClintock, S., Lammie, P., & Melrose, W. (2011). First evidence of spatial clustering of lymphatic filariasis in an Aedes polynesiensis endemic area. Acta tropica, 120, S39-S47. JSI. (2013). Report of the Ghana Urban Malaria Study President's Malaria Initiative. Ghana: JSI Research & Training Institute, Inc. JSI. (2017). USAID | DELIVER PROJEC. from http://deliver.jsi.com/ Judge, A., Welton, N. J., Sandhu, J., & Ben-Shlomo, Y. (2009a). Geographical variation in the provision of elective primary hip and knee replacement: the role of socio-demographic, hospital and distance variables. Journal of Public Health, 061. Judge, A., Welton, N. J., Sandhu, J., & Ben-Shlomo, Y. (2009b). Geographical variation in the provision of elective primary hip and knee replacement: the role of socio-demographic, hospital and distance variables. Journal of Public Health, fdp061. Kadobera, D., Sartorius, B., Masanja, H., Mathew, A., & Waiswa, P. (2012). The effect of distance to formal health facility on childhood mortality in rural Tanzania, 2005–2007. Global health action, 5(1), 19099. Kintu, P., Nayunja, M., Nzabanita, A., & Magoola, R. (2004). Development of HMIS in Least Developed Country Settings: the Case for Uganda. Development of HMIS in Least Developed Country Settings: the Case for Uganda/AHIMA, American Health Information Management Association. Kirby, M., Green, C., Milligan, P., Sismanidis, C., Jasseh, M., Conway, D., & Lindsay, S. (2008). Risk factors for house-entry by malaria vectors in a rural town and satellite villages in The Gambia. Malaria Journal, 7(1), 2. Kogan, M., Lifsey, S., Kirby, R., Kramer, M., & Stopka, T. (2015). DataSpeak - Clusters, Maps and Hotspots: Small Area Analysis in Maternal and Child Health. Paper presented at the DataSpeak series, USA. Kreuels, B., Kobbe, R., Adjei, S., Kreuzberg, C., von Reden, C., Bäter, K., . . . May, J. (2008). Spatial variation of malaria incidence in young children from a geographically homogeneous area with high endemicity. The Journal of infectious diseases, 197(1), 85-93. Kruase, E. (n.d). Performing Polygon-to-Polygon Prediction using areal interpolation. Paper presented at the ESRI Kulldorff, M. (1997). A spatial scan statistic. Communications in Statistics-Theory and methods, 26(6), 1481-1496. Kurz, K. M., & Johnson-Welch, C. (1997). Gender bias in health care among children 0-5 years: opportunities for child survival programs. Leete, R. (2001). Population and housing censuses: A funding crisis. Paper presented at the Symposium on global review of 2000 round of population and housing censuses: Mid-decade assessment and future prospects. Lin, C.-H., & Wen, T.-H. (2011). Using geographically weighted regression (GWR) to explore spatial varying relationships of immature mosquitoes and human densities with the incidence of dengue. International journal of environmental research and public health, 8(7), 2798-2815. 236 University of Ghana http://ugspace.ug.edu.gh Lippeveld, T. (2013). Strengthening routine facility-based health information systems in developing countries: Towards a sustainable data source for measuring the delivery of evidence-based interventions. Paper presented at the 141st APHA Annual Meeting and Exposition (November 2-November 6, 2013). Lippeveld, T. (2015). The Time is Now for Improving Routine Health Information Systems. The Pump. Retrieved 19/07/2016, 2015, from http://thepump.jsi.com/the-time-is-now-for-improving- routine-institution-based-health-information-systems/ Liu, J., Bousema, T., Zelman, B., Gesase, S., Hashim, R., Maxwell, C., . . . Gosling, R. (2014). Is housing quality associated with malaria incidence among young children and mosquito vector numbers? Evidence from Korogwe, Tanzania. PLoS ONE, 9(2), e87358. Lodder, P. (2013). To Impute or not Impute: That’s the Question. Paper Methodological Advice. Longva, S., Thomsen, I., & Severeide, P. I. (1998). Reducing costs of censuses in Norway through use of administrative registers. International Statistical Review, 66(2), 223-234. Luo, L., McLafferty, S., & Wang, F. (2010). Analyzing spatial aggregation error in statistical models of late-stage cancer risk: a Monte Carlo simulation approach. International Journal of Health Geographics, 9(1), 51. Luo, W., & Qi, Y. (2009). An enhanced two-step floating catchment area (E2SFCA) method for measuring spatial accessibility to primary care physicians. Health & place, 15(4), 1100-1107. Lwetoijera, D., Kiware, S., Mageni, Z., Dongus, S., Harris, C., Devine, G., & Majambere, S. (2013). A need for better housing to further reduce indoor malaria transmission in areas with high bed net coverage. Parasites & vectors, 6(1), 57. Macassa, G., Ghilagaber, G., Bernhardt, E., & Burstrom, B. (2004). Contribution of household environment factors to urban childhood mortality in Mozambique. East African medical journal, 81(8), 408-414. Manninen, A., Martin-Guzmán, P., & Bond, D. (2016). A growing demand for small area statistics. How to make demand and supply meet? Marshall, S. (2004). Developing countries face double burden of disease. Bulletin of the World Health Organization, 82(7), 556-556. Martens, P., & Hall, L. (2000). Malaria on the move: human population movement and malaria transmission. Emerging infectious diseases, 6(2), 103. Martin, D., & Williams, H. (1992). Market-area analysis and accessibility to primary health-care centres. Environment and Planning A, 24(7), 1009-1019. Mate, K. S., Bennett, B., Mphatswe, W., Barker, P., & Rollins, N. (2009). Challenges for routine health system data management in a large public programme to prevent mother-to-child HIV transmission in South Africa. PLoS ONE, 4(5), e5483. Mathur, M. (2015). Spatial autocorrelation analysis in plant population: An overview. Journal of Applied and Natural Science, 7(1), 501-513. Matthys, B., Koudou, B., N'Goran, E., Vounatsou, P., Gosoniu, L., Koné, M., . . . Utzinger, J. (2010). Spatial dispersion and characterisation of mosquito breeding habitats in urban vegetable- production areas of Abidjan, Cote d'Ivoire. Annals of Tropical Medicine & Parasitology, 104(8), 649-666. Mavimbe, J. C., Braa, J., & Bjune, G. (2005). Assessing immunization data quality from routine reports in Mozambique. BMC Public Health, 5. doi: 10.1186/1471-2458-5-108 McLafferty, S. L. (2003). GIS and health care. Annual review of public health, 24(1), 25-42. Menard, S. (2002). Applied logistic regression analysis: Sage. Mensah, O. A., & Kumaranayake, L. (2004). Malaria incidence in rural Benin: does economics matter in endemic area? Health Policy, 68(1), 93-102. Metanat, M. (2015). Malaria in children. Int J Infect, 2(1), e20984. Meza, J. L. (2003). Empirical Bayes estimation smoothing of relative risks in disease mapping. Journal of Statistical Planning and Inference, 112(1), 43-62. 237 University of Ghana http://ugspace.ug.edu.gh Millward, R., & Bell, F. N. (1998). Economic factors in the decline of mortality in late nineteenth century Britain. European Review of Economic History, 2(03), 263-288. Mishra, V., & Retherford, R. D. (2006). Does biofuel smoke contribute to anaemia and stunting in early childhood? International journal of epidemiology, 36(1), 117-129. MOH. (1996). Ghana Health Services and Teaching Hospitals Act, 1996 - ACT 525. Accra: Government Printer, Assembly Press. MOH. (2010). Standard Treatment Guidelines. (978-9988-1-2538-7). Accra: Yamens Press. Ltd. MOH. (2012). Private Health Sector Development Policy. Accra: Government Printer, Assembly Press. MOH. (2014). Health Sector Medium Term Development Plan 2014-2017. Assembly Press: Ministry of Health. MOH. (2016). CATEGORY: ROLE & FUNCTIONS OF MOH. from http://www.moh.gov.gh/category/role-functions-of-moh/ Moran, P. (1950). Notes on continuous stochastic phenomena. Biometrika, 37(1/2), 17-23. Muhindo, R., Joloba, E., & Nakanjako, D. (2016). Health management information system (HMIS); whose data is it anyway. Contextual challenges. Rev Public Adm Manag, 4, 190. Müller, I., Smith, T., Mellor, S., Rare, L., & Genton, B. (1998). The effect of distance from home on attendance at a small rural health centre in Papua New Guinea. International journal of epidemiology, 27(5), 878-884. Mungiole, M., Pickle, L. W., Simonson, K. H., & White, A. A. (1999). Application of a weighted head‐ banging algorithm to mortality data maps. Statistics in medicine, 18(23), 3201-3209. Murray, C. J., Laakso, T., Shibuya, K., Hill, K., & Lopez, A. D. (2007). Can we achieve Millennium Development Goal 4? New analysis of country trends and forecasts of under-5 mortality to 2015. The Lancet, 370(9592), 1040-1054. Murray, M., Versteeg, M., & Hugo, J. (2008). The impact of diarrhoea in infants on the quality of life of low-income households. South African Family Practice, 50(2), 62-62c. Musengimana, G., Mukinda, F., Machekano, R., & Mahomed, H. (2016). Temperature variability and occurrence of diarrhoea in children under five-years-old in Cape Town Metropolitan Sub- Districts. International journal of environmental research and public health, 13(9), 859. Mutale, W., Chintu, N., Amoroso, C., Awoonor-Williams, K., Phillips, J., Baynes, C., . . . Sherr, K. (2013). Improving health information systems for decision making across five sub-Saharan African countries: implementation strategies from the African Health Initiative. BMC Health Services Research, 13(2), 1. Mwesigwa, J., Achan, J., Jawara, M., Ditanna, G., Worwui, A., Affara, M., . . . D'Alessandro, U. (2017). SPATIAL-TEMPORAL DYNAMICS IN HETEROGENEITY OF MALARIA INFECTION IN A SETTING WITH SEASONAL TRANSMISSION: A LONGITUDINAL STUDY IN THE GAMBIA. BMJ Global Health, 2(Suppl 2), A26-A26. Naing, N. (2000). Easy way to learn standardization: Direct and indirect methods. The Malaysian journal of medical sciences: MJMS, 7(1), 10. Nations, U. (2000). Handbook on geographic information systems and digital mapping: New York: United Nations Department of Economic and Social Affairs Statistics Division. Ndugwa, R., Ramroth, H., Müller, O., Jasseh, M., Sié, A., Kouyaté, B., . . . Becher, H. (2008). Comparison of all-cause and malaria-specific mortality from two West African countries with different malaria transmission patterns. Malaria Journal, 7(1), 15. doi: 10.1186/1475-2875-7- 15 Neutra, R., Swan, S., & Mack, T. (1992). Clusters galore: insights about environmental clusters from probability theory. Science of the total environment, 127(1-2), 187-200. NMCP. (2010). 2010 ANNUAL REPORT: NATIONAL MALARIA CONTROL PROGRAMME- GHANA HEALTH SERVICE/MINISTRY OF HEALTH. NMCP. (2013). GHANA MALARIA PROGRAMME REVIEW. 238 University of Ghana http://ugspace.ug.edu.gh Noble, M., Wright, G., Smith, G., & Dibben, C. (2006). Measuring multiple deprivation at the small- area level. Environment and Planning A, 38(1), 169-185. Nutley, T., & Reynolds, H. W. (2013). Improving the use of health data for health system strengthening. Global health action, 6. Nyarko, S. H., & Cobblah, A. (2014). Sociodemographic determinants of malaria among under-five children in Ghana. Malaria research and treatment, 2014. Nyonator, F., Ofosu, A., Osei, D., & Atweam, D. K. (2013). DISTRICT HEALTH INFORMATION MANAGEMENT SYSTEM DHIMS II THE EXPERIENCE OF THE GHANA HEALTH SERVICE. Paper presented at the ICT4D Conference. O’brien, R. M. (2007). A caution regarding rules of thumb for variance inflation factors. Quality & Quantity, 41(5), 673-690. Obila, C. O. (2012). Risk factors associated with under-five diarrhoea in Kenya. University of Nairobi, Kenya. OECD. (2013). COVERAGE (OF DATA SOURCES). GLOSSARY OF STATISTICAL TERMS from https://stats.oecd.org/glossary/detail.asp?ID=464 ONS. (2016). Small Area Population Estimates: Summary of methodology review and research update. Methodology. Retrieved December 2, 2017, from https://www.ons.gov.uk/peoplepopulationandcommunity/populationandmigration/populati onestimates/methodologies/smallareapopulationestimatessummaryofmethodologyreviewa ndresearchupdate#summary-of-methodology-review-and-research-update Onwujekwe, O., Uzochukwu, B., Eze, S., Obikeze, E., Okoli, C., & Ochonma, O. (2008). Improving equity in malaria treatment: relationship of socio-economic status with health seeking as well as with perceptions of ease of using the services of different providers for the treatment of malaria in Nigeria. Malaria Journal, 7(1), 5. Openshaw, S. (1977). A geographical solution to scale and aggregation problems in region-building, partitioning and spatial modelling. Transactions of the institute of british geographers, 459- 472. Openshaw, S. (1984). The modifiable areal unit problem. Organization, W. H., & ESPD. (2005). Enhancing health Systems: Malaria's Negative Impact in Africa. Poverty and Social Policy Team. Osei-Kwakye, K., Otupiri, E., Dabo, E. O., Browne, E. N. L., & Adjuik, M. (2010). Determinants of Under-Five Mortality in Builsa District, Upper East Region, Ghana. Journal of Science and Technology (Ghana), 30(1). Osei, D., Ofosu, A., & Nyonator, F. (Producer). (2013). District Health Information Management System DHIMS II: The Data Challenge For Ghana Health Service Provider. Retrieved from http://solutionscenter.nethope.org/case_studies/view/district-health-information- management-system-dhims-ii-the-data-challenge-f Páez, A., Farber, S., & Wheeler, D. (2011). A simulation-based study of geographically weighted regression as a method for investigating spatially varying relationships. Environment and Planning A, 43(12), 2992-3010. Pandey, A., Sengupta, P. G., Mondal, S. K., Gupta, D. N., Manna, B., Ghosh, S., . . . Bhattacharya, S. K. (2002). Gender differences in healthcare-seeking during common illnesses in a rural community of West Bengal, India. Journal of Health, Population and Nutrition, 306-311. Parchman, M. (1995). Small area variation analysis: a tool for primary care research. Family medicine, 27(4), 272-276. Pathak, S., Rege, M., Gogtay, N. J., Aigal, U., Sharma, S. K., Valecha, N., . . . Sharma, S. (2012). Age- dependent sex bias in clinical malarial disease in hypoendemic regions. PLoS ONE, 7(4), e35592. Patz, J., Githeko, A., McCarty, J., Hussein, S., Confalonieri, U., & De Wet, N. (2003). Climate change and infectious diseases. Climate change and human health: risks and responses, 103-132. 239 University of Ghana http://ugspace.ug.edu.gh Peersman, G., Rugg, D., Erkkola, T., Kiwango, E., & Yang, J. (2009). Are the investments in national HIV monitoring and evaluation systems paying off? JAIDS Journal of Acquired Immune Deficiency Syndromes, 52, S87-S96. Peletz, R., Simuyandi, M., Sarenje, K., Baisley, K., Kelly, P., Filteau, S., & Clasen, T. (2011). Drinking water quality, feeding practices, and diarrhoea among children under 2 years of HIV-positive mothers in peri-urban Zambia. The American journal of tropical medicine and hygiene, 85(2), 318-326. Pfeffermann, D. (2002). Small Area Estimation‐New Developments and Directions. International Statistical Review, 70(1), 125-143. Phibbs, C. S., & Robinson, J. C. (1993). A variable-radius measure of local hospital market structure. Health services research, 28(3), 313. Piantadosi, S., Byar, D., & Green, S. (1988). The ecological fallacy. American journal of epidemiology, 127(5), 893-904. Pickle, L. W. (2003). Usability testing of map designs. Paper presented at the Proceedings of Symposium on the Interface of Computing Science and Statistics. Portnoy, B., Lee, S. J. C., Kincheloe, J., Breen, N., Olson, J. L., McCormally, J., & Brown, E. R. (2014). Independent state health surveys: responding to the need for local population health data. Journal of Public Health Management and Practice, 20(5), E21-E33. Pullan, R., Sturrock, H., Magalhaes, R., Clements, A., & Brooker, S. (2012). Spatial parasite ecology and epidemiology: a review of methods and applications. Parasitology, 139(14), 1870-1887. Purcell, D. W., Hall, H. I., Bernstein, K. L., Gift, T. L., McCray, E., & Mermin, J. (2016). The Importance of Population Denominators for High-Impact Public Health for Marginalized Populations. JMIR public health and surveillance, 2(1). Racioppi. (2002). Effects on Household Environmental Health Hazards to Child Survival in selected states in India. Raghunathan, T. E., Lepkowski, J. M., Van Hoewyk, J., & Solenberger, P. (2001). A multivariate technique for multiply imputing missing values using a sequence of regression models. Survey Methodology, 27(1), 85-96. Rahman, A. (2008). A review of small area estimation problems and methodological developments. NATSEM-University of Canberra Discussion paper, 66. Rao, J. (1999). Some recent advances in model-based small area estimation. Survey Methodology, 25(2), 175-186. Rao, J. (2003). Small Area Estimation Wiley Online Library. Rao, J. (2015). Small‐Area Estimation: Wiley Online Library. Rao, J., & Choudhry, G. H. (1995). Small‐Area Estimation: Overview and Empirical Study. Business Survey Methods, 527-542. Rathan, V., Latha, K., & Sulthana, A. (2012). Mapping Disease: Deciphering Geographic Patterns of Breast Cancer risk Areas in the Western region of Tamil Nadu, India. Raymondo, J. C. (1992). Population estimation and projection: Methods for marketing, demographic, and planning personnel: Greenwood Publishing Group. Real, L. A., & Biek, R. (2007). Spatial dynamics and genetics of infectious diseases on heterogeneous landscapes. Journal of the Royal Society Interface, 4(16), 935-948. Reuben, R. (1993). Women and malaria—special risks and appropriate control strategy. Social Science & Medicine, 37(4), 473-480. Rezaeian, M., Dunn, G., St Leger, S., & Appleby, L. (2007). Geographical epidemiology, spatial analysis and geographical information systems: a multidisciplinary glossary. Journal of Epidemiology & Community Health, 61(2), 98-102. Roma, E., & Pugh, I. (2012). Toilets for Health: A Report of the London School of Hygiene and Tropical Medicine in collaboration with Domestos. London: Unilever. Rubin, D. B. (1976). Inference and missing data. Biometrika, 63(3), 581-592. Rubin, R. (1987). Multiple imputation for nonresponse in surveys (J Wiley & Sons, New York, NY). 240 University of Ghana http://ugspace.ug.edu.gh Rutstein, S. O. (2000). Factors associated with trends in infant and child mortality in developing countries during the 1990s/Shea O. Rutstein. Sabel, C. E., Boyle, P., Löytönen, M., Gatrell, A. C., Jokelainen, M., Flowerdew, R., & Maasilta, P. (2003). Spatial clustering of amyotrophic lateral sclerosis in Finland at place of birth and place of death. American journal of epidemiology, 157(10), 898-905. Salvati, N., Chandra, H., Ranalli, M. G., & Chambers, R. (2010). Small area estimation using a nonparametric model-based direct estimator. Computational Statistics & Data Analysis, 54(9), 2159-2171. Salvatore, M., Pozzi, F., Ataman, E., Huddleston, B., & Bloise, M. (2005). Mapping global urban and rural population distributions. Poverty Mapping. Sauro, J. (2012). 10 THINGS TO KNOW ABOUT CONFIDENCE INTERVALS. 2017, from https://measuringu.com/ci-10things/ Schofield, C., & White, G. (1984). Engineering against insect-borne diseases in the domestic environment. Trans. R. Soc. Trop. Med. Hyg, 78, 285-292. Schwartz, S. (1994). The fallacy of the ecological fallacy: the potential misuse of a concept and the consequences. American journal of public health, 84(5), 819-824. SEDAC. (2017). Gridded Population of the World (GPW), v4. from http://sedac.ciesin.columbia.edu/data/collection/gpw-v4 Sharma, R., Singh, M., Saha, K., Bharti, P., Jain, V., Singh, P., . . . Chand, S. (2015). Socio-economic & household risk factors of malaria in tribal areas of Madhya Pradesh, central India. Indian Journal of Medical Research, 141(5), 567. Sharma, S. K., Chattopadhyay, R., Chakrabarti, K., Pati, S. S., Srivastava, V. K., Tyagi, P. K., . . . Das, B. S. (2004). Epidemiology of malaria transmission and development of natural immunity in a malaria-endemic village, San Dulakudar, in Orissa state, India. The American journal of tropical medicine and hygiene, 71(4), 457-465. Shi, X., Miller, S., Mwenda, K., Onda, A., Rees, J., Onega, T., . . . Moeschler, J. (2013). Mapping disease at an approximated individual level using aggregate data: a case study of mapping New Hampshire birth defects. International journal of environmental research and public health, 10(9), 4161-4174. Simon, A. K., Hollander, G. A., & McMichael, A. (2015). Evolution of the immune system in humans from infancy to old age. Paper presented at the Proc. R. Soc. B. Siziya, S., Muula, A., & Rudatsikira, E. (2009). Diarrhoea and acute respiratory infections prevalence and risk factors among under-five children in Iraq in 2000. Italian journal of pediatrics, 35(1), 8. Skarbinski, J., Winston, C. A., Massaga, J. J., Kachur, S. P., & Rowe, A. K. (2008). Assessing the validity of health facility‐based data on insecticide‐treated bednet possession and use: comparison of data collected via health facility and household surveys–Lindi region and Rufiji district, Tanzania, 2005. Tropical Medicine & International Health, 13(3), 396-405. Smith, G., Noble, M., Anttila, C., Gill, L., Zaidi, A., Wright, G., . . . Barnes, H. (2004). The Value of Linked Administrative Records for Longitudinal Analysis Smith, S. (2003). Small area analysis. Encyclopedia of population, 898-901. Smith, S., & Morrison, P. (2005). Small-Area and Business Demography Handbook of Population. Snow, J. (1855). On the mode of communication of cholera John Churchill: London. Snow, R., & Marsh, K. (1998). New insights into the epidemiology of malaria relevant for disease control. British medical bulletin, 54(2), 293-309. Snow, R. W., Omumbo, J. A., Lowe, B., Molyneux, C. S., Obiero, J.-O., Palmer, A., . . . Obonyo, C. (1997). Relation between severe malaria morbidity in children and level of Plasmodium falciparum transmission in Africa. The Lancet, 349(9066), 1650-1654. Snyman, K., Mwangwa, F., Bigira, V., Kapisi, J., Clark, T. D., Osterbauer, B., . . . Liu, J. (2015). Poor housing construction associated with increased malaria incidence in a cohort of young Ugandan children. The American journal of tropical medicine and hygiene, 92(6), 1207-1213. 241 University of Ghana http://ugspace.ug.edu.gh Song, L., Mercer, L., Wakefield, J., Laurent, A., & Solet, D. (2016). Peer Reviewed: Using Small-Area Estimation to Calculate the Prevalence of Smoking by Subcounty Geographic Areas in King County, Washington, Behavioral Risk Factor Surveillance System, 2009–2013. Preventing chronic disease, 13. Stefan, R. (2015). Demographic Techniques: Small-area Estimates and Projections. Elsevier: ScienceDirect ® - International Encyclopedia of the Social & Behavioral Sciences, 6( 2nd Edition), 162-169. doi: 10.1016/B978-0-08-097086-8.31015-7 Steketee, R., Nahlen, B., Parise, M., & Menendez, C. (2001). The burden of malaria in pregnancy in malaria-endemic areas. The American journal of tropical medicine and hygiene, 64(1_suppl), 28-35. Stock, R. (1983). Distance and the utilization of health facilities in rural Nigeria. Social Science & Medicine, 17(9), 563-570. Tanser, F., & Wilkinson, D. (1999). Spatial implications of the tuberculosis DOTS strategy in rural South Africa: a novel application of geographical information system and global positioning system technologies. Tropical Medicine & International Health, 4(10), 634-638. Tatalovich, Z. (2005). A comparison of Thiessen-polygon, Kriging, and spline models of UV exposure. Proceedings of the University Consortium of Geographical Information Science Summer Assembly. Tatem, A. J., Campiz, N., Gething, P. W., Snow, R. W., & Linard, C. (2011). The effects of spatial population dataset choiceon estimates of population at risk of disease. Population Health Metrics, 9(4). Tatem, A. J., Gething, P. W., Smith, D. L., & Hay, S. I. (2013). Urbanization and the global malaria recession. Malaria Journal, 12(1), 133. Tegabu, D. (2011). Introduction to Routine Health Information Systems. Teklehaimanot, & Paola Mejia, A. (2008). Malaria and poverty. Annals of the New York Academy of Sciences, 1136(1), 32-37. Thurstain‐Goodwin, M., & Unwin, D. (2000). Defining and delineating the central areas of towns for statistical monitoring using continuous surface representations. Transactions in GIS, 4(4), 305-317. Timæus, I. M., & Lush, L. (1995). Intra-urban differentials in child health. Health transition review, 163-190. Tobler, W. R. (1970). A computer movie simulating urban growth in the Detroit region. Economic geography, 46(sup1), 234-240. Tsikriktsis, N. (2005). A review of techniques for treating missing data in OM survey research. Journal of Operations Management, 24(1), 53-62. Tusting, L., Ippolito, M., Willey, B., Kleinschmidt, I., Dorsey, G., Gosling, R., & Lindsay, S. (2015). The evidence for improving housing to reduce malaria: a systematic review and meta-analysis. Malaria Journal, 14(1), 209. UN. (2010). Handbook on Population and Housing Census Editing: Revision 1 (Vol. 82): United Nations Publications. UNFPA. (2009). COSTING GUIDE FOR POPULATION AND HOUSING CENSUSES. Paper presented at the 57th Session of the International Statistical Institute (ISI), Durban (South Africa),. http://unstats.un.org/unsd/censuskb20/KnowledgebaseArticle10649.aspx Unicef. (2010). Diarrhoea: why children are still dying and what can be done. http://www. unicef. org/media/files/Final_Diarrhoea_Report_October_2009_final. pdf. Unicef. (2014). Lack of toilets dangerous for everyone, UNICEF [Press release]. Retrieved from https://www.unicef.org/media/media_77952.html USAID|DeliverProject. (2016). USAID׀Deliver Project Final Country Report : Ghana USAID׀Deliver Project. Arlington: John Snow, Inc. 242 University of Ghana http://ugspace.ug.edu.gh Van Ginkel, J. R., Van der Ark, L. A., & Sijtsma, K. (2007). Multiple imputation of item scores in test and questionnaire data, and influence on psychometric results. Multivariate Behavioral Research, 42(2), 387-414. Vanwambeke, S., Bennett, S., & Kapan, D. (2011). Spatially disaggregated disease transmission risk: land cover, land use and risk of dengue transmission on the island of Oahu. Tropical Medicine & International Health, 16(2), 174-185. Vlassoff, C., & Manderson, L. (1998). Incorporating gender in the anthropology of infectious diseases. Tropical medicine & international health: TM & IH, 3(12), 1011-1019. Voight, R. B. (1967). Recent Efforts To Make Census Results More Useful for Local Purposes. American Statistical Association, Washington, DC, December, 27. Vyas, S., & Kumaranayake, L. (2006). Constructing socio-economic status indices: how to use principal components analysis. Health Policy and Planning, 21(6), 459-468. Wagenaar, B. H., Sherr, K., Fernandes, Q., & Wagenaar, A. C. (2016). Using routine health information systems for well-designed health evaluations in low-and middle-income countries. Health Policy and Planning, 31(1), 129-135. Wang, A., & Wheeler, D. C. (2015). Catchment Area Analysis Using Bayesian Regression Modeling. Cancer informatics(Suppl. 2), 71. Wang, L., & Wu, C. (2010). Population estimation using remote sensing and GIS technologies: Taylor & Francis. Wanzirah, H., Tusting, L. S., Arinaitwe, E., Katureebe, A., Maxwell, K., Rek, J., . . . Dorsey, G. (2015). Mind the gap: house structure and the risk of malaria in Uganda. PLoS ONE, 10(1), e0117396. Weiss, L., Ompad, D., Galea, S., & Vlahov, D. (2007). Defining neighbourhood boundaries for urban health research. American journal of preventive medicine, 32(6), S154-S159. Wheeler, D. (2014). Geographically weighted regression Handbook of regional science (pp. 1435- 1459): Springer. Wheeler, D., & Tiefelsdorf, M. (2005). Multi-collinearity and correlation among local regression coefficients in geographically weighted regression. Journal of Geographical Systems, 7(2), 161-187. White, I., Royston, P., & Wood, A. (2011). Multiple imputation using chained equations: issues and guidance for practice. Statistics in medicine, 30(4), 377-399. Whitworth, A. (2013). Evaluations and improvements in small area estimation methodologies. WHO. (2007). Insecticide-treated mosquito nets: a WHO position statement. Geneva: WHO. WHO. (2008a). Framework and standards for country health information systems. WHO. (2008b). Health information systems, toolkit on monitoring health system strengthening. Geneva: WHO. WHO. (2008c). Toolkit on monitoring health systems strengthening: Health Information Systems. . Health Metrics Network Framework and Standards for Country Health Information Systems. WHO. (2011a). Assessment of health facility data quality: Data quality report card Uganda, 2010– 2011. Geneva: WHO. WHO. (2011b). Country health information systems: a review of the current situation and trends. Suiza: World Health Organization. WHO. (2013). Diarrhoeal disease. WHO Fact sheet No 330. Retrieved 26-03-2017, 2017, from http://www.who.int/mediacentre/factsheets/fs330/en/ WHO. (2014). Guide to the health facility data quality report card. WHO. (2015a). Climate change impact on diarrhoeal diseases and malaria in the health districts of Nara and Commune VI of Bamako District, Mali, 2000-2010: a retrospective study: technical report. WHO. (2015b). Ghana: WHO Statistical Profile. In G. H. Observatory (Ed.), Country statistics and global health estimates. USA: WHO and UN partners. WHO. (2015c). Health in 2015 from MDGs to SDGs. Geneva: World Health Organization. 243 University of Ghana http://ugspace.ug.edu.gh WHO. (2016a). FactSheet. Media centre. from http://www.who.int/mediacentre/factsheets/fs094/en/ WHO. (2016b). World malaria report 2014. Geneva: WHO; 2014. WHO. (2016c). World malaria report 2016. Wilkinson, D., & Tanser, F. (1999). GIS/GPS to document increased access to community-based treatment for tuberculosis in Africa. The Lancet, 354(9176), 394-395. Wolff, C., Schroeder, D., & Young, M. (2001). Effect of improved housing on illness in children under 5 years old in northern Malawi: cross sectional study. Bmj, 322(7296), 1209-1212. WorldBank. (2000). World Development Indicators 2000: Oxford University Press, USA. Worrall, E., Basu, S., & Hanson, K. (2005). Is malaria a disease of poverty? A review of the literature. Tropical Medicine & International Health, 10(10), 1047-1059. Wu, C. (2012). Handbook of Applied Spatial Analysis: Software Tools, Methods and Applications edited by Manfred M. Fischer and Arthur Getis: Wiley Online Library. Wu, S.-s., Qiu, X., & Wang, L. (2005). Population estimation methods in GIS and remote sensing: a review. GIScience & Remote Sensing, 42(1), 80-96. Xu, Z., Liu, Y., Ma, Z., Toloo, G., Hu, W., & Tong, S. (2014). Assessment of the temperature effect on childhood diarrhoea using satellite imagery. Scientific reports, 4. Yassin, K. M. (2000). Indices and sociodemographic determinants of childhood mortality in rural Upper Egypt. Social Science & Medicine, 51(2), 185-197. Yerramilli, S., & Fonseca, D. G. (2014). Assessing geographical inaccessibility to health care: Using GIS network based methods. Public Health Research, 4(5), 145-159. Yeshiwondim, A., Gopal, S., Hailemariam, A., Dengela, D., & Patel, H. (2009). Spatial analysis of malaria incidence at the village level in areas with unstable transmission in Ethiopia. International Journal of Health Geographics, 8(1), 5. Zhou, X., Zhou, Y., Chen, R., Ma, W., Deng, H., & Kan, H. (2013). High temperature as a risk factor for infectious diarrhoea in Shanghai, China. Journal of epidemiology, 23(6), 418-423. Zinszer, K., Charland, K., Kigozi, R., Dorsey, G., Kamya, M. R., & Buckeridge, D. L. (2014). Determining health-care facility catchment areas in Uganda using data on malaria-related visits. Bulletin of the World Health Organization, 92, 178-186. Zinszer, K., Charland K., Kigozi R., Dorsey, G., Kamya, M. R., & Buckeridge, D. l. (2014). Determining health-care facility catchment areas in Uganda using data on malaria-related visits. Bulletin of the World Health Organization 2014, 92, 178-186. Zinszer, K., Kigozi, R., Charland, K., Dorsey, G., Kamya, M., & Buckeridge, D. (2015). A Novel Method for Defining Health Facility Catchment Areas in a Low Income Country. Online journal of public health informatics, 7(1). 244