Causal Inference Methods for Infectious Disease Prevention Trials

Causal inference is concerned with using data from a sample to draw conclusions about cause-and-effect relationships in a population, an important task in infectious disease research. Data from infectious disease prevention trials often include complicated features, including measurement error, network dependence, and missing data, each of which poses challenges to existing causal inference methods. This dissertation presents three projects developing new casual inference methods to address these challenges. The first project develops methods to estimate causal effects while addressing both confounding and measurement error in a continuous exposure. Three estimators are proposed based on the g-formula, inverse probability weighting, and doubly robust estimation techniques. The proposed estimators are theoretically shown to be consistent and asymptotically normal, empirically shown to have good finite sample performance, and are applied to investigate causal effects of biomarkers on HIV infection using data from an HIV preventative vaccine trial. The second project tests whether an encouragement-to-isolate intervention reduced transmission of influenza-like-illness on a college campus. While the intervention was intended to alter students' social network, the scientific question ultimately concerns the effect of the intervention on disease transmission. To test the null hypothesis of no effect on transmission without assuming no effect on the network, the social network is modeled as a function of treatment and infection status. The resulting hypothesis testing procedure is shown to asymptotically control the type I error rate and have power against deviations from the null of no transmission effect. The third project aims to estimate the effect of a cluster-level intervention using data from a cluster-randomized trial with individual-level nonresponse. Motivated by a study of an intervention to prevent HIV in communities in Zambia and South Africa, this work allows for intervention assignment to affect an individual's outcome and their propensity to respond. Proposed g-formula and inverse probability weighted estimators correct for bias due to nonresponse, and an augmented inverse probability weighted estimator additionally is doubly robust to model specification for the outcome regression and propensity score models.