Model complexity improves the prediction of nonsuicidal self-injury

Efforts to predict nonsuicidal self-injury (NSSI; intentional self-injury enacted without suicidal intent) to date have resulted in near-chance accuracy. Incongruence between theoretical understanding of NSSI and the traditional statistical methods to predict these behaviors may explain this poor prediction. Whereas theoretical models of NSSI assume that the decision to engage in NSSI is relatively complex, statistical models used in NSSI prediction tend to involve simple models with only a few theoretically informed variables. The present study tested whether more complex statistical models would improve NSSI prediction.

Within a sample of 1,021 high-risk self-injurious and/or suicidal individuals, we examined the accuracy of three different model types, of increasing complexity, in predicting NSSI across 3, 14, and 28 days. Univariate logistic regressions of each predictor and multiple logistic regression with all predictors were conducted for each timepoint and compared with machine learning algorithms derived from all predictors.

Results demonstrated that model complexity was associated with predictive accuracy. Multiple logistic regression models (AUCs 0.70-0.72) outperformed univariate logistic models (average AUCs 0.56). Machine learning models that produced algorithms modeling complex associations across variables produced the strongest NSSI prediction across all time points (AUCs 0.87-0.90). These models outperformed all multiple logistic regression models, including those involving identical study variables. Machine learning algorithm performance remained strong even after the most important factor across algorithms was removed.

Results parallel recent findings in suicide research and highlight the complexity that underlies NSSI. (PsycINFO Database Record (c) 2019 APA, all rights reserved).