Estimating the Effect of Romantic Relationship in Adolescence With IPTW: Using Machine Learning to Compute Propensity Scores
2018 Developmental Methods Conference
Romantic relationships are prevalent during adolescence and engagement in romantic relationships may affect socio-emotional development (Collins, Welsh, & Furman, 2009). For instance, Beckmeyer (2015) found that participation in serious romantic relationships was associated with increased odds of substance use, including alcohol, tobacco, and marijuana. The current study takes a step further to distinguish causation from selection effect of serious romantic relationship participation on adolescents’ socio-emotional development, by computing each adolescent’s probability to participate in romantic relationships (i.e., propensity scores) and adjusting the model using the scores (i.e., Inverse Probability to Treatment Weighting; IPTW). IPTW balances the baseline covariates among treatments, therefore reducing the non-randomness in treatment assignment (i.e., adolescents engage in romantic relationships or not), and creates an experiment-like condition to allow for causal inference (Austin & Stuart, 2015). Estimating propensity scores is more of a prediction problem than an explanatory problem. However, most studies estimate propensity scores with logistic regression (Westreich, Lessler, & Funk, 2011). Very limited attention has been paid to evaluating model assumptions and cross-validation errors. Westreich et al. (2011) suggested considering machine learning methods, and in particular, using boosting algorithms as an alternative to logistic regression. Boosting methods are robust against overfitting issues and perform well in classification problems (Schapire, 2003). Using information about adolescent’s family background, physical development, and socio-emotional development in middle childhood in NICHD Study of Early Child Care and Youth Development (SECCYD), the boosting model correctly predicted whether the adolescent participated in serious romantic relationships in 75% of the cases with test data, with both specificity and sensitivity greater than 70%. Applying the inverse probability weighting with resulting propensity scores, we found that adolescents who participated in serious relationships were less work-oriented and more likely to take risks. However, these adolescents were also less lonely and perceived themselves to be more popular.
Yan, J., & Schoppe-Sullivan, S. J. (September, 2018). Estimating the effect of romantic relationship in adolescence with IPTW: Using Machine Learning to compute propensity scores. Oral Presentation at the 2018 Developmental Methods Conference in Whitefish, MT.