Abstract
BACKGROUND: Little attention has been paid to comprehensively understanding multisystem-level factors that promote father involvement in the context of the U.S. child welfare system. Using novel data science methods, such as machine learning, could further advance child welfare research on father involvement. OBJECTIVE: The current exploratory study applied an ecological systems framework to national child welfare survey data and used random forest, a type of machine learning, to identify the most important predictors of father-child contact across multiple system levels (i.e., individual, family, neighborhood, child welfare systems). PARTICIPANTS AND SETTINGS: Data came from the National Survey of Child and Adolescent Wellbeing-Third Cohort (N = 2380). METHODS: Random forest was used for variable importance analysis using 124 multisystem-level predictors. RESULTS: Fathers' sociodemographic characteristics (e.g., race/ethnicity, education) were among the most important predictors of father-child contact, followed by caseworker characteristics (e.g., other language spoken), service referrals by the child welfare system (e.g., health, legal), family economic condition and composition (e.g., access to food, number of children), and child characteristics (e.g., sex). CONCLUSIONS: Factors at different system levels interact in complex ways to predict father-child contact. Child welfare researchers can consider machine learning as a complementary method to traditional statistics. Methodological advantages machine learning brings include enhancing predictive accuracy, handling large amounts of data, and modeling complex and non-linear relationships with numerous predictors. These advantages allow for uncovering new quantitative insights and patterns to advance child welfare research, practice, and policies.