Bayesian Methods for Global Health Monitoring
Efforts to improve global health depend on monitoring, and successful monitoring depends on rigorous and creative statistical methodology. In particular, the combination of constrained resources and complex data requires novel models that leverage the full extent of available information while honestly reporting uncertainty. The Bayesian paradigm provides a powerful and flexible framework for pursuing these goals.
In the first chapter, we propose a new surveillance system for estimating the prevalence of primary HIV drug resistance. By pooling sera samples, we increase the accuracy of prevalence estimates while simultaneously decreasing surveillance costs. We present a Bayesian model for pooled-testing data that accounts for uncertainty about the inter-subject heterogeneity of resistance levels as well as for the measurement error of resistance assays. By bringing virologic prior information to bear, our model renders the prevalence parameter identifiable in instances when existing non-model-based estimators fail.
In the second chapter, we turn our attention to methods for monitoring cardiometabolic risk factors. We present a model that estimates population-level risk factor trends over the course of the last three decades for 199 countries and territories. The model takes as input sample average risk factor levels, accounting for the fact that not all samples are nationally representative. We allow for time and age nonlinearity, and we borrow strength in time, age, covariates, and within and across regional country clusters to make estimates where data are sparse. The model outputs predictions of various functionals of interest along with their uncertainty.
In the last chapter, we extend this model to make semiparametric estimates of the full distributions of markers of undernutrition in 138 low- and middle-income countries in 2005. Since tail probabilities are of clinical importance for these markers, estimating the distributions' shapes is of particular substantive significance. In addition, this structure allows coherent posterior inference based on both individual-level data when available, as well as a variety of aggregated summary statistics from studies whose individual-level data could not be obtained.
HIV drug resistance, cardiometabolic disease, and undernutrition each contributes substantially to global morbidity and mortality. These three Bayesian models provide practical methodological innovations for monitoring these conditions.