Use of Machine Learning to Extract Predictors of Graduation from Social Networks in Therapeutic Communities

Background and Purpose: Mutual aid based treatment for substance abuse has attracted the attention of social network researchers in recent years, who have added to the already strong evidence of the importance of peer support in recovery. However, analyzing social network structure as a predictor of success in these programs, as opposed to simply counting the number of network connections that individual have with peers, is challenging due to the difficulty of gathering complete networks. Therapeutic communities (TCs) often record a variety of resident interactions. By treating these records as a social network it is possible to gain insight into the relationship between social network structure and program outcomes.

Methods: This study analyzes network structure as a predictor of graduation in a clinical dataset of approximately 300,000 recorded affirmations exchanged between the residents of three minimum security correctional TCs over a period of eight years. Because the TCs were gender segregated men and women were analyzed separately. The machine learning program XGBoost, a decision tree algorithm that yields a test error statistic that indicates the number of cases that were misclassified along with the variables that predict classification, was used in the analysis. Analysis was done for separate units and subsequently for all TCs.

The analysis included variables at three system levels. The individual level variables of age, race, and score on the Level of Service Inventory-Revised (LSI-R) were included in all models. (It was not possible to include gender in the models of individual TCs because the units were gender segregated.) Network level variables were outdegree and indegree centrality (direct connections between residents), eigenvector centrality (the extent to which residents are connected to well-connected peers), closeness centrality (the overall closeness of a resident to all peers), betweenness centrality (the number of paths between peers on which a resident lies) and network clustering (a measure of closed triads in the resident's network). Contextual variables were average peer LSI-R, participant entrance date (a proxy for changes in unit conditions over time) and in the final model a dummy variable that differentiated between TC units.

Results: The XGBoost models correctly predicted graduation between 70% and 80% of the time. Measures of network centrality and clustering predicted graduation in all units, but which variables were the best predictors varied between units. At least two patterns did emerge. First, in all cases at least one social network variable predicted graduation more effectively than any contextual variable. Second, the race of the participants did not predict graduation in any model.

Conclusions and Implications: While the predictive value of particular social network structures varied between units, overall these results establish the importance of network structure as a predictor of TC graduation. Further, in all cases at least one measure of social network structure was a better predictor than any contextual variable, suggesting that social network structure buffers against context. These results suggest that it should be possible to improve program graduation rates by influencing social network structure. Finally, race did not predict graduation in these programs.