The largest study of relationship quality found that who your partner is barely matters

Pooling studies is not the same as running a big study, and the distinction is what gives this one its force.

Every dataset here was longitudinal, meaning couples were measured at baseline and then followed. That allows a harder question than the usual one: not what correlates with satisfaction today, but what measured at the start predicts satisfaction later.

The machine learning served a specific purpose that is easy to misread. It was not there to find hidden magic. Random forests can capture interactions and nonlinear effects that a standard regression misses, so if compatibility works the way people assume, as a matter of the right combination of two people, this is the method most likely to detect it.

Crucially, models were trained on some datasets and tested on others they had never seen. A predictor that only worked within its own sample did not survive. That is the step that separates a robust finding from a well-fitted one, and it is why the surviving list is short.

The variables split into two families, and they did not perform remotely alike.

Relationship-specific variables, meaning how a person sees this particular relationship, accounted for close to half the variance in relationship quality at baseline, and up to about 18 percent of variance at the end of each study. The strongest were perceived partner commitment, appreciation, sexual satisfaction, perceived partner satisfaction, and conflict.

Individual-difference variables, meaning stable facts about the person, explained around 21 percent. The strongest among them were life satisfaction, negative affect, depression, attachment avoidance, and attachment anxiety. So a person's own general wellbeing does carry real signal about how they will rate their relationship, which is worth keeping in view.

Relationship-specific variables were roughly two to three times as predictive as individual differences. Once you know how someone experiences their relationship, knowing who they are adds comparatively little.

Here is the number worth carrying around. One partner's own self-reported characteristics predicted about five percent of the variance in the other partner's relationship satisfaction.

Five percent. Across 11,196 couples, everything a person reports about themselves, their personality, their attachment style, their values, their mental health, their satisfaction with life, tells you almost nothing about how happy their partner is with the relationship.

This is the opposite of the model most people use. Compatibility is usually imagined as a property of two people, which is why dating platforms match on traits and why so much advice is about identifying the right sort of person. On this evidence, that model has the arithmetic backwards.

Joel's own summary of it is hard to improve on: who I am does not really matter once I know who I am when I am with you.

One result got much less attention than it deserved, and it is arguably the most useful part of the paper for a reader.

The flexible models did not meaningfully outperform straightforward ones. If relationship quality were driven by intricate interactions between partners' traits, the kind of thing where trait A works only in combination with trait B, the random forests would have found it and beaten the linear models. They largely did not.

That is evidence against a specific and widely held idea: that there is a matching formula waiting to be discovered, and that the reason nobody has found it is insufficient data or insufficient computing power. This study had unusual amounts of both. It found that the useful information was not in the combination of two people's traits.

What it was in, instead, was how the relationship is already going.

It is a study of self-reports, which is its central limitation. Everything measured is what people said about themselves and their relationships, and people are not always accurate about either. If something matters that nobody thought to ask about, it is not in the 2,413 variables.

There is also a direction problem that longitudinal design reduces without eliminating. Perceived partner commitment predicting later satisfaction is consistent with commitment causing satisfaction, and equally consistent with both flowing from something else earlier. The design rules out the crudest reverse-causation story, not every version of it.

The datasets also skew toward the populations that academic couples research usually reaches, and the samples are not representative of the world. And the study measured relationship quality, not whether couples stayed together, which are related but not identical outcomes.

None of that undoes the central finding. It does mean the sensible reading is that partner traits explain very little of the variance measured here, rather than that partners do not matter.

The practical implication is not that partner choice is irrelevant. It is that the choosing is a smaller lever than it feels like, and the thing it is usually weighed against is a much larger one.

Every top predictor on that list is a description of what happens between two people rather than a property either of them brought. Whether your partner seems committed. Whether you feel appreciated. Whether the sex is good. How conflict goes. These are not fixed attributes to be screened for; they are what a relationship is doing, and they can move.

That cuts in both directions, which is why it is not simply an encouraging finding. It means a relationship with the right person on paper can still score badly, and it means the work of maintaining one is not decorative. But it does relocate the question. Less about whether you picked correctly, more about what the two of you are currently building.