Addressing bias in causal effects estimated under misspecified interference sets, with application to HIV prevention trials
Date of Award
Spring 2024
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Public Health
First Advisor
Forastiere, Laura
Abstract
Interference, or spillover, is a common phenomenon that occurs in studies of biomedical or behavioral interventions, where a participant's outcomes may depend on not only their own exposure to the intervention, but also the intervention received by other participants. For example, it has been shown in various network-based HIV behavioral studies, in which some participants were trained on and encouraged to disseminate HIV knowledge and safe practices to members of their peer networks, that both those who did and did not receive the training demonstrated reduced risk behaviors. Methods for estimating causal effects under interference typically rely on the specification of an interference set for each participant, that is, a set of individuals among whom spillover is possible. In existing literature, interference sets are conventionally assumed to be correctly specified so that the exposure of each participant to the intervention of others is correctly measured, and potential outcomes are in turn well-defined. However, oftentimes investigators may not have accurate information to identify true social connections, and interference sets may be misspecified as a consequence. In this dissertation, I showed that causal effects estimated with usual methods are biased when interference sets are misspecified, and proposed methods to correct for this bias under various settings. In HIV prevention studies where social and sexual networks play a critical role in disease transmission, bias-correcting causal effects is crucial for accurately evaluating of the full impact of interventions and designing future intervention strategies. This dissertation is a collection of three papers, where I developed bias-corrected estimators under various study designs and estimation approaches. In Chapter 2, I considered an egocentric-network randomized trial, a study design commonly used to assess peer-based strategies, where participants who receive the intervention are encouraged to disseminate knowledge in their peer networks. Under this design, networks collected at study baseline may not represent the participants' true network ties, as participants may become friends with those in other networks, or fall out of touch with the members with whom they enrolled. In Chapter 3, I considered a two-stage cluster randomized trial, a popular study design for measuring spillover effects, in which clusters are first randomized to an intervention allocation strategy, then participants within clusters are randomized to the intervention accordingly. In this setting, interference is typically assumed to be contained within randomization clusters, yet social interactions may extend across them but are falsely assumed away. In Chapters 3 and 4, I also considered general cluster randomized trials with intervention non-compliance, as well as observational settings, where intervention uptake is no longer randomized but rather depends on individual characteristics. Causal effects in Chapters 2-4 were estimated using sample average, regression-based, and inverse probability weighting estimators, respectively. In each chapter, I proposed corresponding methods to bias-correct causal effects estimated under misspecified interference sets, by leveraging a validation study in which true interference sets are ascertained alongside the observed and error-prone ones for a subset of the study sample, such that the measurement error process can be empirically estimated. I investigated finite sample properties of our methods using extensive simulation studies, and illustrated our methods in the HIV Prevention Trials Network 037 study in Chapter 2, and the Botswana Combination Prevention Project in Chapters 3 and 4.
Recommended Citation
Chao, Ariel, "Addressing bias in causal effects estimated under misspecified interference sets, with application to HIV prevention trials" (2024). Yale Graduate School of Arts and Sciences Dissertations. 1320.
https://elischolar.library.yale.edu/gsas_dissertations/1320
