GRAPH MODELS WITH MISSING DATA TECHNIQUES FOR APPLICATIONS IN DRUG USE NETWORKS

Abstract

Problem: Social context is important in the reduction of engagement in potential risk behaviors, particularly among people who use injection drugs (PWID). People engage in risky behavior together via sharing needles or other drug use equipment and/or though risky sex. If we can identify which attributes are related to increased (or decreased) odds of engaging in risky behavior with others, we can use those results to better target HIV risk reduction interventions among PWID. Methods: Exponential random graph models (ERGMs) were used to model the probability that there is a tie or connection between people in a network based on a set of given attributes. Networks used were from the Social Risk Factors and HIV Risk (SFHR) study and the Transmission Reduction Intervention Project (TRIP). The SFHR study included participants who had injected drugs within the past year and who lived in and around New York city between 1991 and 1993. The TRIP study also recruited participants who were injection drug users and their contacts who lived in Athens, Greece between 2013 and 2015. Two missing data imputation techniques were used (and compared with complete case analysis models): propensity score methods and random forest. Results: Across both data sets, results indicate that people were more likely to engage in risky behaviors with others who were similar to them in some way (e.g., are the same sex or are of the same race/ethnicity). We also found that those who were homeless were more likely to engage in risk behaviors, compared to those who were not homeless, and that they were likely to engage in risky behaviors with other people who were also homeless. Results remained consistent across all models, indicating that the two missing data imputation techniques did not have a strong influence on results.