Suicide Risk Prediction Among Youth Experiencing Homelessness: A Machine Learning-Based Decision-Tree Analysis

Background/ Purpose: Suicide is a major societal problem but decidedly more so among youth experiencing homelessness (YEH). Unfortunately, suicide research has overwhelmingly focused on the attributes of individual youth as risk factors and that work has not meaningfully improved our ability to predict suicide in the last half century. Decision-tree (DT) methods represent well-established machine learning algorithms that show promise for suicide risk “prediction.” However, no known DT models have incorporated aspects of the social network environment or been developed for youth experiencing homelessness. The present study used a theory-driven machine learning decision-tree analysis to identify individual and social network features that “predict” suicidal ideation and attempts among YEH.

Methods: Data were collected from 940 YEH at two drop-in centers in Los Angeles. Participants completed a survey about their individual characteristics and a social network interview about their relationships. Classification and Regression Tree models were used to predict suicidal ideation and suicide attempts. These DT models explore data by recursively partitioning it into sub-groups to yield the most accurate predictions. To choose the “best” predictor at each partitioning step, we used the Gini impurity measure. In each analysis, we divided the dataset into training data (75% of the sample) and testing data (25% of the sample). The training dataset was used to create each tree model whereas the testing dataset was used to evaluate each tree model’s prediction performance. Area Under the Curve (AUC) metrics were computed to evaluate the prediction results (0.5 = chance; 1.0 = perfect prediction).

Results: Half of the factors identified as being important in our DT analyses were related to the social network environment. Ten variables comprised the suicidal ideation decision tree. The top five predictors (in order of importance) were sum of traumatic experiences, lifetime hard drug use, depression, proportion of street friends providing emotional support, proportion of network members engaging in hard drug use. Seven variables comprised the suicide attempt decision tree. The top five predictors (in order of importance) were depression, proportion of friends objecting to bad behavior, homelessness age, home-based friend presence, and sum of traumatic experiences. Based on the AUC values, our models were found to be 79% accurate in differentiating between YEH with and without suicidal ideation and 86% accurate in differentiating YEH who did and did not attempt suicide.

Conclusion/Implications: Suicide risk evaluation is a daunting endeavor that involves gathering and synthesizing information on many factors to inform clinical decisions. By using DT analysis with a more inclusive set of individual and social network factors, we were able to generate classification trees with promising sensitivity and utility among YEH. Our AUC values are equal to or exceed those observed in most suicide-related studies using DT or similar classification analyses. Notably, we showed that norms operating in the environments of YEH deserve more attention in the context of suicide prevention programming. Therefore, the overarching message is that targeting specific subgroups for intervention may be a promising way to augment traditional individual-level strategies.