Collections > Electronic Theses and Dissertations > Causal Inference for Binary Data with Interference

Developing methods to quantify the effects of interventions to prevent infectious diseases in the presence of interference is the overall objective of this research. Interference is present when an individual's outcome is affected by the treatment of any other individuals under study. First, two methods are developed for constructing randomization based confidence intervals for the average effect of a treatment on a binary outcome without interference. The methods are nonparametric and require no assumptions about random sampling from a larger population. Both of the resulting 1 - α confidence intervals are exact and guaranteed to have width no greater than one. In contrast, previously proposed asymptotic confidence intervals are not exact and may have width greater than one. The first approach combines Bonferroni adjusted prediction intervals for the attributable effects in the treated and untreated. The second method entails inverting a permutation test. While simulations show that the permutation based confidence intervals have smaller average width, the attributable effects based confidence intervals are more computationally feasible as sample size increases. Extensions that allow for stratifying on categorical baseline covariates are also discussed. Secondly, for a two-stage randomized experiment assuming stratified interference, methods are developed for constructing exact confidence intervals for the direct, indirect, total, and overall effect of a treatment on a binary outcome. The methods are nonparametric and require no assumptions about random sampling from a larger population. The new exact confidence intervals are compared via simulation with previously proposed exact and asymptotic confidence intervals. While the asymptotic intervals do not maintain nominal coverage for certain simulation setups, the new exact confidence intervals maintain nominal coverage for all setups and have narrower width than the previously proposed exact confidence interval. Thirdly, we consider a Bayesian approach to causal inference with interference in an observational study under the assumption that the treatment assignment mechanism is ignorable. We compare the methods via a simulation study to previously proposed IPW estimators. The methods are applied to data from the 2007 Demographic and Health Survey in the Democratic Republic of the Congo, examining the impact of individual and community bed net use on malaria.