The influence of jittering DHS cluster locations on geostatistical model-based estimates of malaria risk in Cameroon.

Abstract

Background: In low-and-middle income countries, national representative household surveys such as the Demographic and Health Surveys (DHS) and the Malaria Indicator Surveys (MIS) are routinely carried out to assess the malaria risk and the coverage of related interventions. A two-stage sampling design was used to identify clusters and households within each cluster. To ensure confidentiality, DHS made the data available after jittering (displacement) of the geographical coordinates of the clusters, shifting their original locations within a radius of 10 km. Our study assessed the influence of jittering on the estimates of the geographical distribution of malaria risk and on the effects of malaria control interventions using data from the latest MIS in Cameroon.

Methods: We generated one hundred datasets by jittering the original MIS data. For each dataset, climatic factors were extracted at the jittered locations and Bayesian geostatistical variable selection was applied to identify the most important climatic predictors and malaria intervention coverage indicators. The models were adjusted for potential confounding effects of socio-economic factors. Bayesian kriging based on the selected models was used to estimate the geographical distribution of malaria risk. The influence of jittering was analysed using results of the variable selection and the Bayesian credible intervals of the regression coefficients.

Results: Geostatistical variable selection was sensitive to jittering. Among the important predictors identified in the true data, distance to water bodies and presence of forest were mostly influenced by the jittering. Altitude and vegetation index were the least affected predictors. The various sets of selected environmental factors were able to capture the main spatial patterns of the disease risk, but the jittering increased the prediction error. The parameter estimates of the effects of socio-economic factors and intervention indicators were relatively stable in the simulated data.

Conclusion: In Cameroon, the malaria risk estimates obtained from the jittered data were comparable to the ones generated using the true locations; however, jittering modified our interpretation of the relationship between environmental predictors and malaria transmission.