The 11,196-Couples Study: What Really Predicts Relationship Quality

Who is Samantha Joel, and what is this study?

Samantha Joel is a Canadian psychologist. She earned her PhD at the University of Toronto (2015), spent time at the University of Texas at Austin and the University of Utah, and has taught at Western University in Ontario since 2018. There she runs a lab devoted to decisions in romantic relationships: why people commit, why they stay, why they leave, and why they so often get those choices wrong.

In 2020, she published a paper in the journal PNAS with Paul Eastwick and 84 other researchers: the largest study ever done on the question “what predicts the quality of a relationship?” Rather than launching a new survey, the team pooled 43 existing longitudinal studies run by labs around the world, 11,196 couples in all, and ran machine learning algorithms on the data.

Why it matters for your relationship: this study answers, with massive data, the question every couples test faces, MyLoveLab included. Should a test measure who you are (your personality, your attachment, your mood) or what you live in the relationship? The answer is clear, and it has direct consequences for what a test can and can’t tell you.

The central idea

What best predicts the quality of a relationship is not the personality of the two people. It is what each person perceives about the relationship itself. How committed I think my partner is, how much I appreciate what they do, our sexual satisfaction, what I believe they feel, how much conflict there is. Once those perceptions are known, knowing whether you are anxious, depressed, or avoidant adds almost nothing. And what your partner answers on their side adds little to what you perceive yourself.

The key concepts

Relationship-specific variables versus individual variables

The study sorts every question asked of couples into two families. Individual variables describe a person independently of their relationship: life satisfaction, negative affect, depression, attachment anxiety and avoidance, personality traits. Relationship-specific variables describe what a person lives and perceives in this particular couple: perceived partner commitment, appreciation, sexual satisfaction, conflict. The whole question was which of the two families weighs the most.

Perceived partner commitment

The number one predictor is what I believe about my partner’s commitment. Not their real commitment, measured on their side, but what I perceive of it. Next come appreciation (how grateful I feel for what they bring), sexual satisfaction, the satisfaction I attribute to them, and the level of conflict. For a couple, the lesson is concrete: making your commitment visible matters as much as feeling it.

The share of variance explained

Researchers don’t predict couples one by one. They measure how much of the differences between couples their variables can explain. Relationship-specific variables explain up to 45% of the variance in relationship quality measured at the same moment; individual variables, 21%. In other words, even with the best data, more than half of what separates a happy couple from an unhappy one stays out of reach of questionnaires.

Predicting the future is much harder

When you try to predict relationship quality several months or years later, the numbers drop: 18% for relationship-specific variables, 12% for individual ones. The quality of a relationship changes for reasons that the questionnaires at the start don’t capture: life events, stress, decisions made along the way. It’s a humble result, and probably the most important one in the study.

What you perceive matters more than what your partner reports

One last lesson: a person’s answers about themselves predict their satisfaction two to four times better than their partner’s answers do. What your partner says about their own commitment matters less for your satisfaction than what you perceive of it. That doesn’t mean your partner doesn’t count. It means the path runs through your perception, so through what they make visible and what you are able to see.

What the research says

Method and data: 43 longitudinal datasets, 11,196 couples, questionnaires only (no lab observation). The algorithm used, random forests, is trained on one part of the data and tested on another, which avoids the classic trap of predictions fitted after the fact. The results are then compared from one study to the next to keep only what replicates.

• At the same moment: relationship-specific variables explain up to 45% of the variance in relationship quality, individual variables 21%.
• Over time: 18% and 12% respectively, for quality measured at the last wave of each study.
• No added value from individual variables or from the partner’s answers once a person’s own relationship perceptions are taken into account.
• Five relationship-specific predictors lead the list: perceived partner commitment, appreciation, sexual satisfaction, perceived partner satisfaction, conflict.
• Five individual predictors lead the list: life satisfaction, negative affect, depression, attachment avoidance, attachment anxiety.

Two other pieces of Samantha Joel’s work round out the picture. In 2017, with Paul Eastwick and Eli Finkel, she showed in Psychological Science that with more than a hundred questionnaires filled out before speed-dating events, machine learning could not predict who would desire whom: attraction to a specific person escapes traits. In 2021, with Geoff MacDonald, she described a progression bias: we are drawn to move forward in a relationship (commit, move in, stay) rather than to leave it, often without asking whether it suits us.

Limits and criticisms

The first limit, which the authors acknowledge: everything rests on self-report questionnaires. The predictor “perceived partner satisfaction” is very close to what researchers are trying to predict, satisfaction itself, and part of the 45% comes from that. Direct observation measures (what couples do during a conflict, what sensors record) were not in the data, even though they bring different information.

The second limit: the 43 studies come mostly from North America and Europe, with mostly heterosexual couples, often students or newlyweds. Whether the findings extend to other cultures and other kinds of couples has not been established. And machine learning tells you which variables matter, not why: there is no mechanism, only robust associations.

Finally, the most important limit for you: this study is about populations, not about one couple. Explaining 45% of the variance across thousands of couples doesn’t let you tell a specific couple what will happen to them, and 18% over time lets you say even less. No test, MyLoveLab included, can predict the future of your relationship. It can describe its present.

How MyLoveLab uses it

This study guided how the test was built. First, in the choice of what to measure: the test mostly asks about what you live and perceive in your relationship (seven relationship dimensions), and very little about your personality. Attachment, the only individual variable that carries weight, is measured separately, with questions inspired by the ECR-R scale, and cross-checked against the rest. Second, in what the test promises: a portrait of the current state of the relationship, never a prediction. And finally in the program, where visible commitment and expressed appreciation come back in almost every dimension, from Communication to Shared project.

Further reading

• Machine learning uncovers the most robust self-report predictors of relationship quality across 43 longitudinal couples studies · Article, 2020The study itself, published in PNAS with 86 authors: the reference for this page.
• Is romantic desire predictable? Machine learning applied to initial romantic attraction · Article, 2017The precedent: more than a hundred questionnaires can’t predict who will desire whom in speed-dating.
• We’re not that choosy: Emerging evidence of a progression bias in romantic relationships · Article, 2021Why we move forward in a relationship by default, without always asking whether it suits us.

Frequently asked questions

References

1. 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, 117(32), 19061-19071.
2. Joel, S., Eastwick, P. W. & Finkel, E. J. (2017). Is romantic desire predictable? Machine learning applied to initial romantic attraction. Psychological Science, 28(10).
3. Joel, S. & MacDonald, G. (2021). We’re not that choosy: Emerging evidence of a progression bias in romantic relationships. Personality and Social Psychology Review, 25(4).
4. Gottman, J. M. & Levenson, R. W. (2000). The timing of divorce: Predicting when a couple will divorce over a 14-year period. Journal of Marriage and Family, 62(3), 737-745.
5. Heyman, R. E. & Slep, A. M. S. (2001). The hazards of predicting divorce without crossvalidation. Journal of Marriage and Family, 63(2), 473-479.
6. Karney, B. R. & Bradbury, T. N. (1995). The longitudinal course of marital quality and stability: A review of theory, method, and research. Psychological Bulletin, 118(1), 3-34.
7. Fraley, R. C., Waller, N. G. & Brennan, K. A. (2000). An item response theory analysis of self-report measures of adult attachment. Journal of Personality and Social Psychology, 78(2), 350-365.

See the full bibliography