Date of Award
2026
Degree Type
Thesis
Degree Name
Master of Science in Statistics
Department
Computer Science and Statistics
First Advisor
Natallia Katenka
Second Advisor
Ashley Buchanan
Abstract
Standard causal inference methods assume that one person’s treatment does not affect another person’s outcome, an assumption that is fundamentally implausible for infectious disease. When an HIV-positive individual receives antiretroviral therapy, their viral load decreases and their capacity to transmit HIV to others is reduced. This dependence between individuals is called interference, and ignoring it produces biased estimates while missing the most important population health benefits of an intervention. Capturing interference requires defining an interference set, the group within which spillover is assumed to operate, yet existing studies adopt a single definition without empirical justification or systematic comparison across alternatives, leaving the spatial level at which spillover operates unexamined.
This thesis applies the IPW2 estimator, a factorized inverse probability weighting approach, to the Botswana Combination Prevention Project, a community-randomized trial conducted in 30 villages from 2013 to 2018. Direct and spillover effects of four intervention components (HIV testing, linkage to care, antiretroviral therapy, and voluntary medical male circumcision) on sexual risk behavior, HIV incidence, and viral suppression are estimated simultaneously under household, plot, and village interference set definitions. Cluster-level summaries of the remaining intervention components are included as confounders to isolate each component’s contribution, and covariate selection used a hybrid change-in-estimate and backward elimination procedure. All propensity score and causal parameters are estimated jointly through a stacked system of M-estimating equations, ensuring that uncertainty in the propensity score step is fully propagated into the causal estimates via the sandwich variance estimator.
Individual propensity score models achieved perfect covariate balance across all 36 treatment-outcome-level combinations. Direct treatment effects were largely non-significant, with two exceptions: HIV testing produced small positive individual effects on HIV incidence at higher coverage levels, most plausibly attributable to detection effects, and voluntary medical male circumcision produced significant individual and total effects on sexual risk behavior at the village level. All other significant findings operated through spillover. Linkage to care, antiretroviral therapy, and voluntary medical male circumcision each produced large significant spillover and overall effect reductions in sexual risk behavior at the household and plot levels, but the linkage to care spillover collapsed to near zero at the village level while voluntary medical male circumcision remained significant at all three levels. Linkage to care and antiretroviral therapy each produced significant spillover reductions in HIV incidence through overlapping treatment-as-prevention mechanisms. The antiretroviral therapy spillover effect on sexual risk behavior was robustly negative at the household and plot levels but non-significant and reversed in sign at the village level. These patterns demonstrate that comparing interference set definitions is a necessary inferential step rather than a sensitivity analysis: the spatial level at which spillover operates varies by intervention component, cannot be assumed in advance, and is only recoverable when multiple definitions are examined simultaneously.
Recommended Citation
Ofori, Francisca A., "ASSESSING CAUSAL EFFECTS OF HIV PREVENTION AND TREATMENT PACKAGE INTERVENTIONS UNDER DIFFERENT DEFINITIONS OF SPILLOVER" (2026). Open Access Master's Theses. Paper 2706.
https://digitalcommons.uri.edu/theses/2706