学术文献库

Modern disease mapping draws upon a wealth of high resolution spatial data products reflecting environmental and/or socioeconomic factors as covariates, or `features', within a geostatistical framework to improve predictions of disease risk. Feature selection is an important step in building these models, helping to reduce overfitting and computational complexity, and to improve model interpretability. Selecting only features that have a causal relationship with the response variable could potentially improve predictions and generalisability, but identifying these causal features from non-interventional, spatiotemporal data is a challenging problem. Here we examine the performance of a causal feature selection procedure with regard to estimating malaria incidence in Madagascar. The studied procedure designed for this task combines the PC algorithm with spatiotemporal prewhitening and kernel-based independence tests extended to accommodate aggregated data. This case study reveals a clear advantage for causal feature selection in terms of the out-of-sample predictive accuracy in a forward temporal estimation task, but not in a spatiotemporal interpolation task, in comparison with thresholded spike-and-slab, for both linear and non-linear regression models. Compared to no feature selection, causal feature selection was most beneficial in settings wherein the volume of available data was low relative to the model complexity.