Read on Twitter
New paper out today, in which 86 relationship scholars combined forces to find out how much of relationship quality is currently predictable and what predicts it best. https://www.pnas.org/content/early/2020/07/21/1917036117
Altogether, we had 43 longitudinal dyadic datasets, comprised of 11,196 romantic couples recruited from six countries (the US, Canada, Switzerland, New Zealand, the Netherlands, and Israel). The couples were tracked for an average of one year.
We first took all variables collected at the beginning of each study, and categorized them into either relationship-specific variables (judgments about the relationship or the partner), or individual differences (reports about the self).
Then, we inputted these sets of variables into machine learning models (Random Forests) to try to predict relationship satisfaction at the beginning of the study (first time point), at the end of the study (last time point), and over time (change scores).
People’s own judgments about the relationship explained 45% of their satisfaction at the 1st time point, and 16% by the last time point. The best rel-specific predictors were perceptions of the partner’s commitment and satisfaction, appreciation, sexual satisfaction and conflict.
People’s self-reported individual differences explained 19% of their satisfaction at baseline, and 11% by the end of the study. The five best individual difference predictors were satisfaction with life, negative affect, depression, attachment anxiety and attachment avoidance.
We also examined partner reports. One partner’s relationship judgments predicted 15% of the other partner’s satisfaction at the 1st time point, and 10% by the last time point. One partner’s traits predicted only 5% of the other partner’s satisfaction at either time point.
We also tried combinations of these sets of predictors in search of cumulative effects. E.g., if relationship quality is about having the right combination of traits from each partner, then including both partner’s traits in one model should capture those interactions.
But combining both partner’s traits added no predictive power beyond just one person’s traits. And, combining both partner’s relationship reports added no predictive power beyond just one partner’s reports. Nothing added to the models beyond own judgments about the relationship.
We also had difficulty predicting change in relationship quality over time, with any set of the variables. Less than 5% of the variance in change in satisfaction was explained in any given model.
In addition to relationship satisfaction, we conducted all of these same models predicting relationship commitment, which produced highly similar results.
Limitations: These data are definitely WEIRD, and included primarily self-report variables. Future work should examine whether these effects generalize beyond the Western context, and to non-self-report data such as observational or physiological measures.
In sum: the results of our multilab project suggest that relationship happiness may be less about choosing the right partner than it is about building the right relationship. More work is still needed on the fledgling relationship stage to learn exactly how that process unfolds.
Biggest thanks to @PaulEastwick and to all our coauthors. Thank you for taking a chance on this project, for trusting us with your data, and for replying to our endless emails! Our field has so many talented and kind scholars and it's been a joy to get to know you all better.
