Reliable and timely information on disease-specific treatment burdens within a health system is critical for the planning and monitoring of service provision. Health Management Information Systems (HMIS) exist to address this need at national scales across Africa but are failing to deliver adequate data due to widespread under-reporting by health facilities. Faced with this inadequacy, vital public health decisions often rely on crudely adjusted regional and national
estimates of treatment burdens. This study has taken the example of presumed malaria in outpatients within the largely incomplete Kenyan HMIS database and has developed geostatistical modelling frameworks for the prediction of the monthly tally of treatments for malaria (MC) at all facilities and months where this value is missing. Three different kriging methodologies were compared to test the effect on prediction accuracy of (a) the extension of a spatial-only to a spacetime prediction approach, and (b) the replacement of a globally-stationary with a locally-varying random function model. Space-time kriging was found to produce predictions with 98.4% less mean bias and 14.8% smaller mean imprecision than conventional spatial-only kriging. A modification of space-time kriging that allowed space-time variograms to be recalculated for every prediction location within a spatially-local neighbourhood resulted in a larger decrease in mean imprecision over ordinary kriging (18.3%) although mean bias was reduced less (87.5%). Because the MC variable included non-spatial variation caused by differences between individual facilities and their catchment populations, a series of studies were conducted to model catchment population size. These predictions require refined models that incorporated rich local data that were not available at the national level so directly estimated catchment population values were not available. An alternative approach was developed that incorporated data on the total number of outpatients seen at facilities each month as a proxy measure of catchment size. Two modelling frameworks were developed to implement this approach and the most accurate model was
identified in a cross-validation exercise. A model-based and an empirical method were developed to measure the uncertainty of predictions of MC and how this changed as sets of predictions were aggregated in space and time. The final set of predictions enabled the national treatment burden for presumed malaria in the government health sector to be defined during the 1996-2002 period. During this time, the national annual treatment burden was predicted as 6.8 million cases, with an expected margin of error of 1.3%. The modelling framework presented here provides for the first time reliable information from imperfect HMIS data to support evidence-based decision making at national and sub-national levels.