Bayesian spatial analysis of a national urinary schistosomiasis questionnaire to assist geographic targeting of schistosomiasis control in Tanzania, East Africa

Abstract

Spatial modelling was applied to self-reported schistosomiasis data from over 2.5 million school students from 12,399 schools in all regions of mainland Tanzania. The aims were to derive statistically robust prevalence estimates in small geographical units (wards), to identify spatial clusters of high and low prevalence and to quantify uncertainty surrounding prevalence estimates. The objective was to permit informed decision-making for targeting of resources by the Tanzanian national schistosomiasis control programme. Bayesian logistic regression models were constructed to investigate the risk of schistosomiasis in each ward, based on the prevalence of self-reported schistosomiasis and blood in urine. Models contained covariates representing climatic and demographic effects and random effects for spatial clustering. Degree of urbanisation, median elevation of the ward and median normalised difference vegetation index (NDVI) were significantly and negatively associated with schistosomiasis prevalence. Most regions contained wards that had >95% certainty of schistosomiasis prevalence being >10%, the selected threshold for bi-annual mass chemotherapy of school-age children. Wards with >95% certainty of schistosomiasis prevalence being >30%, the selected threshold for annual mass chemotherapy of school-age children, were clustered in north-western, south-western and south-eastern regions. Large sample sizes in most wards meant raw prevalence estimates were robust. However, when uncertainties were investigated, intervention status was equivocal in 6.7-13.0% of wards depending on the criterion used. The resulting maps are being used to plan the distribution of praziquantel to participating districts; they will be applied to prioritising control in those wards where prevalence was unequivocally above thresholds for intervention and might direct decision-makers to obtain more information in wards where intervention status was uncertain.

Fig. 1

Raw prevalence of self-reported schistosomiasis (a) and blood in urine (b) in Tanzanian wards.

Fig. 1

Raw prevalence of self-reported schistosomiasis (a) and blood in urine (b) in Tanzanian wards.

Fig. 2

Frequency histograms of observed and fitted values using Bayesian spatial models for prevalence of self-reported schistosomiasis (a) and blood in urine (b) in Tanzanian wards.

Fig. 3

Bayesian probability maps of the prevalence of schistosomiasis (a) and blood in urine (b) using a posterior median prevalence threshold of 10%. Wards with >95% probability of having prevalence >10% are dark grey, wards with >95% probability of having prevalence <10% are white and wards with <95% probability of having prevalence > or <10% are light grey.

Fig. 3

Bayesian probability maps of the prevalence of schistosomiasis (a) and blood in urine (b) using a posterior median prevalence threshold of 10%. Wards with >95% probability of having prevalence >10% are dark grey, wards with >95% probability of having prevalence <10% are white and wards with <95% probability of having prevalence > or <10% are light grey.

Fig. 4

Bayesian probability maps of the prevalence of schistosomiasis (a) and blood in urine (b) using a posterior median prevalence threshold of 30%. Wards with >95% probability of having prevalence >30% are dark grey, wards with >95% probability of having prevalence <30% are white and wards with <95% probability of having prevalence > or <30% are light grey.

Fig. 4

Bayesian probability maps of the prevalence of schistosomiasis (a) and blood in urine (b) using a posterior median prevalence threshold of 30%. Wards with >95% probability of having prevalence >30% are dark grey, wards with >95% probability of having prevalence <30% are white and wards with <95% probability of having prevalence > or <30% are light grey.

Fig. 5

Probability maps of spatially structured residual components of Bayesian models for self-reported schistosomiasis (a) and blood in urine (b) in Tanzanian wards. Wards with >95% probability of having positive spatially structured residual prevalence (i.e. significant high-prevalence clusters) are dark grey, wards with >95% probability of having negative spatially structured residual prevalence (i.e. significant low-prevalence clusters) are white and wards with <95% probability of having positive or negative spatially structured residual prevalence are light grey.

Fig. 5

Probability maps of spatially structured residual components of Bayesian models for self-reported schistosomiasis (a) and blood in urine (b) in Tanzanian wards. Wards with >95% probability of having positive spatially structured residual prevalence (i.e. significant high-prevalence clusters) are dark grey, wards with >95% probability of having negative spatially structured residual prevalence (i.e. significant low-prevalence clusters) are white and wards with <95% probability of having positive or negative spatially structured residual prevalence are light grey.