Practicable Characterization of Systematic Heterogeneity

Abstract

In public health, personalized medicine is the ideal. For example, an effective strategy for improving the health of a population is to measure the health of constituent subpopulations and intervene where the treatment is most needed. Alternatively, a member of a subpopulation presents with an ailment and relevant covariates are used to determine the appropriate treatment. Such strategies reasonably assume that there is heterogeneity in the effect of the treatment on health across subpopulations. However identification of heterogeneity tends to be expensive, understandably so due to the demands we are making of our data. Costs appear when having to make strong a priori assumptions about the number and identifying characteristics of the subpopulations across which the treatment effect differs, in the increased sample size required for the data to fill in gaps left by the absence of assumptions, and/or in the manual evaluation of large numbers of covariates. This dissertation discusses different approaches to reducing this cost. Chapter 1 compares a Lot Quality Assurance Sampling (LQAS) survey conducted in southwestern Uganda to an unaffiliated but coincident Demographic Health Survey (DHS) and shows that if we redefine our goal in terms of the programmatic decisions we need to make, we can come to the same conclusions at a fraction of the cost. In Chapter 2 I consider just how expensive it is to identify heterogeneity in the absence of a priori assumptions, and draw some general conclusions about the capabilities and limitations of extant modern methods of causal inference. I conclude with Chapter 3, where I leverage the insights of Chapter 2 to build a visualization application that facilitates the exploratory, hypothesis-generating analysis of treatment effect heterogeneity (TEH), particularly for large datasets where a manual evaluation of covariates is not practicable.