Evidence
What 43 longitudinal datasets showed about prediction
In 2020, a large collaboration pooled decades of longitudinal couples research and applied machine learning to it. The headline result concerned which variables predicted best. The more instructive result concerned how much remained unpredicted.
The project was led by Samantha Joel, then at Western University, with Paul Eastwick and a large team of contributing researchers, and appeared in the Proceedings of the National Academy of Sciences. Its design addressed a structural problem in the field: individual studies typically test a handful of variables in one sample, so the literature accumulates many small findings that cannot easily be compared or ranked against one another.
The collaboration assembled 43 longitudinal datasets covering more than eleven thousand couples, harmonised the measures across them, and used random-forest models to estimate how well each variable predicted later relationship quality. Because the datasets were independent, models trained on some could be tested on others — the cross-validation that comparable earlier work had often lacked.
Two classes of variable were compared. Individual-difference measures described a person independent of any relationship: personality traits, attachment dimensions, depressive symptoms, life satisfaction. Relationship-specific measures described how a person characterised the particular relationship: perceived commitment from a partner, appreciation, sexual satisfaction, perceived partner satisfaction, reported conflict.
Relationship-specific variables predicted substantially better than individual-difference variables. The strongest single predictors of a person's later relationship quality were their own reports about the relationship — how committed they judged their partner to be, how appreciated they felt, how satisfied they judged their partner to be. Traits described as belonging to the person, taken alone, carried much less predictive weight.
Across the pooled data, the models accounted for roughly 45 percent of the variance in a person's own reported relationship quality, and considerably less — under a fifth — in a partner's. That leaves most of the variance in partner-reported quality, and more than half of it overall, unaccounted for by every measure the field had collected.
The authors also reported that models did poorly at predicting change over time. Predicting a later state from an earlier one is largely a matter of the earlier state persisting; anticipating a shift is a harder problem, and one this apparatus did not solve.
Two limitations apply to the whole exercise. Everything analysed was self-report, so the finding that self-reported relationship variables predict self-reported relationship quality carries the shared-reporter caveat that runs through this literature. And the pooled datasets inherit the sampling of their sources — predominantly North American and Western European, with the demographic skew that follows from university and community recruitment.
Read carefully, the study is a corrective in two directions. It weakens the position that individual traits are the main determinant of how relationships go: the relationship-level measures did better, and did so consistently across independent samples. It also weakens confidence in prediction as such. This was the largest and best-validated attempt the field has made, and most of the variance stayed out of reach.
That is an ordinary result for social science, where prediction at the level of the individual case is rare and effect sizes are modest. It is a notable result for a research area whose findings frequently reach general readers in the form of confident forecasts.
Sources
- Joel, S., Eastwick, P. W., et al. (2020). Machine learning uncovers the most robust self-report predictors of relationship quality across 43 longitudinal couples studies. Proceedings of the National Academy of Sciences.
- Heyman, R. E., & Slep, A. M. S. (2001). The hazards of predicting divorce without crossvalidation. Journal of Marriage and Family.
This article summarises published research. It is not professional advice, does not assess any individual situation, and makes no claim about outcomes. See the disclaimer.