The most quoted number in relationship science is that a researcher can watch a couple argue for fifteen minutes and predict with ninety-four percent accuracy whether they will divorce. It has been repeated in bestsellers, in therapists' offices, in wedding toasts and in a great many articles like this one. It is also the weakest thing in the field. The strongest thing in the field is a much quieter finding from a study of more than eleven thousand couples, and it says something considerably more useful about what actually holds two people together.

Where the Famous Number Came From

John Gottman spent decades recording couples in a laboratory apartment in Seattle, coding facial expressions and vocal tone and physiological arousal second by second. Out of that came the framework most people have absorbed whether or not they know his name: the four corrosive patterns he called criticism, contempt, defensiveness and stonewalling, with contempt the most damaging of the four.

As a description of how conflict goes wrong, that framework has been enormously useful and remains widely used by clinicians who find it maps onto what they see in the room. The trouble is not the constructs. The trouble is the accuracy claim.

In 2001 Richard Heyman published a paper with the unimprovable title "The Hazards of Predicting Divorce Without Crossvalidation." His argument was technical and devastating in the ordinary way of statistics. A model built to fit the couples in a particular study will fit those couples very well, because it was built to. That is not prediction. Prediction means the model, fixed in advance, works on couples it has never seen. Gottman's models were generated and evaluated within the same samples, and were not carried forward intact to be tested against independent data. Heyman's word for it was postdiction.

Some of the broader findings have held up in other laboratories. The specific ninety-four percent, and the several other high accuracy figures that have circulated alongside it, should be read as a description of how well a model described the couples it was built from. It is not a claim that anyone can watch you argue and know how your marriage ends.

What the Largest Study Found Instead

In 2020 a team led by Samantha Joel did something the field had not done. Rather than run another study, they gathered forty-three existing longitudinal datasets on couples, more than eleven thousand of them in total, and threw machine learning at the combined pile to ask which self-reported variables actually predicted relationship quality over time.

The answer came in two lists, and the gap between them is the finding.

The strongest predictors were all about the relationship itself. Perceived partner commitment came first. Then appreciation. Then sexual satisfaction, perceived partner satisfaction and conflict. The strongest individual predictors, the things about you as a person rather than about the pair of you, were life satisfaction, negative affect, depression, and the two forms of insecure attachment.

Read that first list again, because there is a word doing a lot of work in it. Perceived. What predicted how someone felt about their relationship was their sense of how committed their partner was, and their sense of being appreciated. Not the partner's own separately reported commitment. Your read on the other person turned out to carry more information than the other person's self-report.

The relationship-specific variables accounted for up to forty-five percent of the variance in relationship quality measured at the same time, and up to eighteen percent at the end of each study's follow-up period.

Eighteen Percent Is the Honest Part

It is tempting to skip past that last figure and most write-ups do. It deserves the attention.

Eighteen percent means that the best available combination of everything couples researchers know how to ask about, applied to the largest dataset ever assembled on the question, leaves the large majority of what happens to a relationship over time unexplained. Not mysterious in principle. Just not captured by any of the questionnaires, in any of the studies, so far.

Some of that is measurement. Self-report is a blunt instrument and people are unreliable narrators of their own marriages. Some of it is that lives contain illness and money and children and grief and work and luck, and no scale asks about the specific thing that will matter most to you in 2029.

The appropriate response to this is not despair. It is a certain relief. The predictive machinery is genuinely not good enough to tell you how it goes. Anyone selling you a diagnostic quiz with a verdict at the end is selling you the ninety-four percent, not the eighteen.

The One That Is Actually Actionable

Of the top predictors, most are not things you can decide to change on a Tuesday. You cannot resolve to find your partner more committed. Sexual satisfaction is an outcome as much as an input. Conflict frequency has causes upstream of anyone's intentions.

Appreciation is different. Appreciation is a top-five predictor of relationship quality and it is also, uniquely on that list, something one person can simply do, unilaterally, starting immediately, without agreement or negotiation or a conversation about the relationship.

And note the direction it runs. The predictor was feeling appreciated. Which means the lever you hold is not your own gratitude, privately felt, but whether the other person knows. Gratitude that stays inside your head has no measured effect on anybody's marriage. It has to be said out loud, and it has to be specific enough to be believed. Thank you for handling the insurance call. I noticed you got up with her both nights. That was a hard thing you did and I saw you do it.

This sounds too small to matter, which is exactly why it goes undone in long relationships. The things a partner does reliably become invisible precisely because they are reliable. Twenty years of someone taking the bins out produces no thanks at all, and the absence is not hostility, it is just habituation working as designed.

What Repair Looks Like

The other thing that survives from the conflict research, once the accuracy claims are set aside, is not really about avoiding fights. Couples who stay together are not couples who argue less. They are couples who come back.

Repair is the move after the rupture. It is unglamorous and often clumsy. It looks like coming back into the room twenty minutes later. Like a hand on a shoulder before either party is ready to concede anything. Like saying that came out worse than I meant it. Like a joke that is not quite funny, offered as a flag rather than a punchline. The specific quality of a repair attempt matters far less than whether one is made, and whether the other person lets it land.

Contempt is the one to watch, and here the clinical observation and the data agree. Contempt is different in kind from anger. Anger says this matters to me. Contempt says you are beneath the conversation, and it is very hard to repair from, because there is no rupture to come back from when one person has quietly left.

The Version Worth Keeping

Strip out the machinery and what the largest study on this subject says is close to what people knew already, which is often how it goes with good research.

Whether your relationship is any good has less to do with who you each are than with what happens between you. It rests substantially on whether each of you believes the other one is in it, and whether each of you feels seen for what you contribute. Both of those are perceptions, which means both are maintained or eroded by the ordinary traffic of a Tuesday rather than by anything anybody says on an anniversary.

And most of it, still, is not predictable from anything anyone knows how to ask. Which leaves the whole thing, mercifully, in your hands rather than in a model's.