Forecasting Relationship Dissolution: Machine Learning Identifies Key Predictors Across Two Longitudinal Samples

What determines whether a romantic relationship survives? We used machine learning (random forest and elastic net models) to predict relationship dissolution in two longitudinal samples. In Study 1 ( N = 1,281), drawn from individuals in dating relationships in the United States and Canada, both models achieved 64% balanced accuracy, suggesting that dissolution is largely explained by linear combinations of relational predictors. Top predictors included commitment, relational uncertainty, and concerns about the consequences of ending the relationship. Commitment-focused variables outperformed trait-focused ones in predicting breakup (64% vs. 55% balanced accuracy). In Study 2 ( N = 6,947), a nationally representative German sample of young to middle-aged adults tracked over 10 years, the random forest model achieved higher (71%) balanced accuracy among a broader set of contextual predictors. Top predictors included socioeconomic background, division of household labor, and conflict communication. Several nonlinear interactions emerged, highlighting the complexity of predicting dissolution. The findings highlight the joint role of relational, demographic, and contextual factors in predicting relationship stability. (163 words).