Abstract
In-person relationship education classes funded by the federal government have problematic attrition rates and do not have the intended effects. Recently, investigators in this area have innovated the curricula to make these programs briefer and web-based, increasing completion rates and improving relationship distress. One area that requires additional attention is the method and intensity of practitioner contact that a couple requires to complete the online program and receive the intended benefit. Using machine learning (i.e., random forests), the current study seeks to:1) accurately predict program completion and changes in relationship satisfaction; 2) identify the most powerful predictors within three different levels of coaching and a waitlist condition; and 3) examine the most powerful predictors of treatment adherence and gains in relationship satisfaction between three levels of coaching and a waitlist condition. Using a randomized clinical trial of 1,246 couples, in the first aim, prediction accuracy was poor but outperformed other machine learning studies in the area. In the second and third aims, the random forest algorithm was able to identify hundreds of non-parametric effects statistically significant in the within- and between-coaching conditions. In contrast, linear methods identified substantially fewer effects. Specifically, greater trust and lower relationship satisfaction at baseline were consistent predictors of linear improvements in relationship satisfaction within most conditions. Between coaching conditions, those with greater income were more likely to complete the program but saw smaller improvements in relationship satisfaction with greater amounts of coaching. Further, those with lower levels of motivation to change the relationship were more likely to complete the program with more coach contact whereas those with greater levels of motivation were more likely to complete the program with less. Strengths, weakness, and implications for future research are discussed.