This study provides evidence in support of the hypothesis that the inclusion of covariates in a geostatistical model can improve the predictive performance of the model with respect to the classification into ranges of prevalence, both at pixel and sub-county levels. Direct evidence of the importance of covariates is also given by the reduction in the variance of the residual spatial random effects, as well as in the standard errors of the predicted prevalence.
Another finding of our study is that, despite the low overall prevalence of STH in Kenya, the model incorporating covariates was able to classify a larger proportion of sub-counties into specific prevalence classes with a higher degree of confidence compared to the model without covariates. This clearly demonstrates the added value of incorporating these covariates into predictive models for STH prevalence, particularly in regions where the prevalence is low. Both empirical and simulated data exhibited this result.
To the best of our knowledge, this is the first study to report the impact and usefulness of spatially referenced covariates to classify areas into different classes of STH transmission using geostatistical methods. This allows policy makers to make more informed decisions on the administration of MDA programs by locally tailoring the interventions and surveillance, potentially improving the effectiveness of these interventions.
The Western and Eastern regions of Kenya in this study were analysed separately due to their ecological differences. This strategy allowed for a more accurate reflection of the unique environmental and socioeconomic conditions that affect STH prevalence in different parts of the country. The prevalence map developed in our study indicates the heterogeneities associated with STH and that the requirement for MDA is not homogenous throughout the study region, which is important for disease control. Using STH prediction maps to identify sub-regions in need can improve the efficiency of intervention and resource allocation.
The use of the WHO’s STH decision tree17 provides a structured and globally recognized framework for the classification of STH prevalence classes. The maps of probability for different prevalence thresholds generated from this research provide critical information for policymaking and implementation of MDA strategies. The predictive probabilities reduce the possibility of misclassifying sub-counties and pixels. This ensures that the endemicity levels of STH are accurately determined so that resources are not squandered on interventions and populations that do not need them.
In this study, the spatial predictors for STH that were included in the final model were aridity index, enhanced vegetation index, tasselled cap wetness, availability of improved drinking water source from WASH, rainfall, and elevation. This is in line with findings from previous studies23–27. In addition to the covariates considered in this paper, the performance of geostatistical models could be further enhanced by socio-economic variables28,29. When comparing the predictive inferences between the model with covariates against the model without covariates, as expected, this tended to differ the most in areas with fewer or not sampled locations. In addition, predictions from the model with covariates also exhibited a higher spatial heterogeneity than the model without covariates. These differences are explained by the fact that the model without covariates only relies on the Gaussian process for the interpolation of prevalence which reverts to its zero mean in areas that are further away from the data. The validity of the inferences of the model with covariates may also be biased due to the potential misspecifications of the relationship between prevalence and the covariates. However, the use of a linear relationship on the log-odds scale may still adequately capture monotone relationships and, we would expect that, in this case, the negative effects of misspecification would be negligible.
For covariate selection we used a simple, objective method. This is justified by our focus on spatial prediction rather than on the estimation of covariate effects per se. In the presence of residual spatial correlation, estimated covariate effects in observational studies of the kind considered here can be biased because of spatial confounding, which potentially occurs when a covariate d(x) is spatially correlated with the Gaussian process S(x). When comparing the estimates of the regression coefficients before after the inclusion of the Gaussian process S(x), we did find that this changed substantially, thus suggesting the presence of a non-negligible correlation between d(x) and S(x).
To measure socio-economic status (SES), studies have used variables such as land and farm animal ownership, the type of housing (rented or owner-occupied), level of education of the household head, demographic characteristics, and other proxy indicators for economy such as household head’s occupation30. In household and individual level surveys, such as the Demographic Health Survey, data on long-term asset ownership, access to utilities and infrastructure, as well as characteristics of housing are collected. One can use these factors to construct an SES index with principal components analysis30. However, the data on SES is only available at the sampled locations, and thus the use of an SES index as a predictor for STH prevalence would require the development of a bivariate geostatistical model. This could be used to infer the SES at unsampled locations whilst allowing for the propagation of the uncertainty into predictive inferences for prevalence. However, one of the issues of this approach is that the variation in SES can occur at fine spatial scale, making its use as a predictor for prevalence more difficult, especially when data are sparsely sampled over space. In this paper, we have considered proxies for SES that are available as raster files, such as poverty and WASH information. However, these proxies are modelled indicators and, as such, may not adequately capture the spatial heterogeneity inherent to SES.
Another important finding of this study is that the use of sparsely sampled data can lead to high uncertainty levels in the classification of areas into endemicity levels. As already argued, covariates can help to reduce this uncertainty but that may not be enough if a very low level of uncertainty is required to make a decision regarding continuation or cessation of MDA. Our analysis is based on site-level data from 23 counties from Kenya, with no data available from other areas, especially in the northern regions. Our results showed that a relatively large number of sub-counties cannot be classified if uncertainty levels below 80% are required. In this case, two approaches can be followed to reduce the number of unclassified sub-counties. If resources are available, more data should be collected from the unclassified regions to generate sharper classification probabilities. If this is not feasible, then one should consider reducing the probability threshold for classification at the cost of higher uncertainty.