Machine Learning for Community-Based Intervention Studies: Predicting Post-Program Levels of and Changes in Relationship Quality Following Couple Relationship Education Participation

Abstract

With the understanding that applying machine learning methods to study predictors of Couple Relationship Education (CRE) program outcomes could be particularly helpful for identifying critical factors for CRE program designs, the current study used a racially and economically diverse sample of CRE participants to investigate the most important predictors of post-program levels of and changes in relationship quality over the time of one year following CRE program start. Using 79 possible predictors that included participant characteristics, relationship history, relationship skills and practices, individual functioning, and class contextual factors, random forest models identified that up to 33% of the variance in CRE program participants’ post-program levels of relationship quality at one year following program-start, and using 158 possible predictors, 17% of the variance in their changes in relationship quality one year after program-start could be predicted by several, key relational processes variables, individual functioning variables, and family environment variables, regardless of individual characteristics and other contextual variables. Models provided information on the most salient 20-26 factors and align well with a recent framework used for guiding key topics for Couple Relationship Education (CRE). The top five predictors for post-program levels of relationship quality were self-reported satisfaction with dyadic coping in the relationship, partner’s dyadic coping, family harmony, partner’s conflict management, and knowledge of partner. For changes in relationship quality following CRE participation, the top five predictors were self-reported satisfaction with dyadic coping in the relationship, partner’s dyadic coping, acceptance of partner, partner’s negative health control messages, and family environment chaos. Self-reported individual mental health was another important predictor, uniquely accounting for additional variance in changes in relationship quality but not post-program levels of relationship quality. This study offers critical information for skill-based CRE program designs. The current study demonstrated the feasibility of applying machine learning methods to community-based CRE studies and further validated the content emphasis on a broad spectrum of relational processes in skill-based CRE programs. Research and practical implications were discussed.