Machine learning uncovers the most robust self-report predictors of relationship quality across 43 longitudinal couples studies.
Level 3 - non-randomized controlled study
Machine learning analysis of 43 longitudinal dyadic cohort datasets
PubMed 32719123 · doi:10.1073/pnas.1917036117
What was done
Researchers applied machine learning (Random Forests) to 43 dyadic longitudinal datasets from 29 laboratories to determine how well relationship quality can be predicted and to identify the strongest self-report predictors. The analysis evaluated both baseline relationship quality and changes over time, comparing relationship-specific variables, individual-difference variables, and partner-reported ratings.
What was found
Relationship-specific variables predicted up to 45% of variance in relationship quality at baseline and up to 18% at study follow-up. Top relationship-specific predictors included perceived-partner commitment, appreciation, sexual satisfaction, perceived-partner satisfaction, and conflict. Individual differences predicted 21% of variance at baseline and 12% at follow-up (top predictors: life satisfaction, negative affect, depression, attachment avoidance, and attachment anxiety). Actor-reported variables predicted two to four times more variance than partner-reported variables. Individual differences and partner reports added no predictive power beyond actor-reported relationship variables alone. Longitudinal change in relationship quality over time was largely unpredictable from any combination of self-reported baseline variables.
Why it matters
This study synthesizes data across dozens of relationship cohorts to demonstrate that subjective perceptions of the relationship matter far more than individual personality traits or partner-reported variables. It also highlights that baseline self-reports fail to predict whether a relationship will improve or decline over time.
Limits
The total number of individual participants is not reported in the abstract. The findings are based entirely on self-report measures and observational data, precluding causal claims and omitting unmeasured behavioral or physiological interactions.