Date of Award


Degree Type


Degree Name

Master of Science in Statistics


Computer Science and Statistics

First Advisor

Natallia Katenka


People who inject drugs (PWID) are a population with increased HIV risk due in part to sharing drug injection equipment. In networks of people connected through risk behavior, individuals can exert influence on each other. Evidence suggests that receiving medications for opioid use disorder (MOUD) can reduce HIV risk behaviors. However, limited studies have been conducted to determine if there are direct effects (among the participants treated) or disseminated effects (among participants who were not treated themselves, but shared connections with those treated) of MOUD among networks of PWID.

The main goal of this thesis is to evaluate if receiving MOUD causes reductions in HIV risk behavior and if being socially connected to people that receive MOUD also causes reductions in HIV risk behavior. To achieve this goal we analyzed a network of 246 PWID from the Transmission Reduction Intervention Project conducted in Athens, Greece 2013 to 2015. For our analysis we utilized methods for causal inference in the presence of dissemination in combination with methods for missing data imputation. To quantify the causal effects of MOUD on subsequent HIV risk behaviors, we employed a group-level inverse probability weighted approach to adjust for confounding. Specifically, we identified communities of participants in the network using modularity-based community detection. Two community detection methods, a "node-moving" algorithm and a "spectral" algorithm, were used to detect communities for a sensitivity analysis. We employed a finite sample correction for the standard errors of the estimators to account for the relatively small number of communities (20 to 21) in the network.

To impute missing data for covariates, we used a non-parametric random forest method. To impute missing outcome data we first assumed a missing at random (MAR) mechanism. We fit a longitudinal model with a random effect for the community to estimate the posterior distribution of the outcome. Based on this distribution, missing outcomes were imputed and a sensitivity analysis was conducted to account for the more realistic assumption that the outcomes were missing not at random (MNAR).

The results of our analysis showed significant disseminated effects of MOUD on HIV risk behaviors. Although the magnitude of these effects was sensitive at times to the community detection method, the conclusion of the analysis remained unchanged. Analysis of the community structure suggests alternative approaches to identifying interference sets (i.e., communities or clusters) that consider the nearest-neighbors or the full network should be considered. A nearest-neighbors approach would define interference sets uniquely for each individual in the network. Interference sets would consist of the subjects to whom an individual is directly connected. Sensitivity analysis of the missing data mechanism shows strong evidence of significant causal effects under MAR and MNAR assumptions.

Available for download on Saturday, May 08, 2021