The Famous Divorce Prediction Does Not Predict Divorce
A researcher watches fifteen minutes of a couple arguing and predicts divorce with 93 percent accuracy. Cross-validated against real divorce rates, that becomes about one in five.
Checked against primary sources by the Editorial Team. Every figure in this article was traced to the study it came from, including the ones that contradict what we previously published. How we review.
This article contains affiliate links. If you buy something through one of them, we may earn a commission at no extra cost to you. It never changes what we recommend or what we say about it. Advertiser Disclosure.
The short version
- You have probably heard that a researcher can watch a couple argue for fifteen minutes and predict divorce with over 90 percent accuracy.
- Those models were built and tested on the same small group of couples, which is not prediction.
- When someone ran the test properly, the odds that a couple flagged as heading for divorce actually divorced fell to about one in five.
- Parts of the underlying work do hold up. The famous number is not one of them.
It is one of the most repeated claims in popular psychology. A researcher watches fifteen minutes of a couple arguing, counts four specific behaviours, and predicts with better than ninety percent accuracy whether they will still be together years later.
It is a wonderful story. It has been in books, documentaries, TED talks and, until today, on this website. And the arithmetic behind it does not work.
Where the number comes from
The headline result was published in 2000. Researchers followed 79 couples for fourteen years, 22 of whom divorced, and reported accuracy of 93 percent in identifying who would split and roughly when.
The problem is not the data collection, which was careful. The problem is what "prediction" means here. The statistical model was built using the outcomes it then went on to "predict." There was no separate group of couples held back to test it against.
This matters more than it sounds. Given any small dataset, you can construct a model that describes it almost perfectly. That tells you the model fits those 79 couples. It tells you nothing about whether it works on anyone else.
What happened when someone checked
In 2001 Richard Heyman and Amy Smith Slep did the obvious thing. They took a model of the same type, built it on part of a sample, then tested it on the part it had never seen. That is called crossvalidation, and it is the standard test of whether a prediction is real.
Overall accuracy fell from 90 percent to 69 percent. Sensitivity fell from 92 percent to 46 percent. And the number that matters most to an actual couple, the chance that a prediction of divorce turns out to be right, fell from 65 percent to 29 percent.
Then they adjusted for how common divorce actually is in the population, which any real-world prediction has to do. At a realistic base rate, the figure drops to about 21 percent.
Roughly one time in five. A method that tells four couples out of five they are heading for divorce when they are not is not a diagnostic tool. It is a source of unnecessary fear.
An independent team tried to replicate it
In 2007 a separate group followed 85 community couples for about two and a half years and tested the same affective process models. Their summary sentence is blunt: the major findings failed to replicate.
Specific patterns that the popular account treats as decisive, including a husband's rejection of his wife's influence and a wife's harsh start-up, were not predictive in their sample. Of twenty-two affective processes examined, two predicted separation.
Two claims with no study behind them at all
While checking this article we tried to source two things we had previously published, and could not.
The first is that contempt is the strongest of the four behaviours, more powerful than the other three combined. We could find no primary study that ranks the four against one another. The claim appears in popular writing and in training materials, not in the peer-reviewed record.
The second is that a heart rate above 100 beats per minute makes productive conversation impossible. The underlying idea, that physiological arousal impairs discussion, is real and has laboratory support. The specific number is a clinical rule of thumb from popular books, not a validated threshold.
What does hold up
The ratio
Stable couples in these studies showed roughly five positive interactions for every negative one during conflict, while couples heading for trouble ran much closer to parity. That finding is real, though it comes from the same small samples, between 73 and 79 couples, and should be held loosely.
Demand and withdraw
The pattern in which one partner pushes for discussion while the other retreats is one of the better replicated findings in relationship research, and it does tend to be gendered in the direction the popular account describes. This is the part of the story that survived scrutiny.
What actually predicts relationship quality
The strongest evidence available now comes from a 2020 analysis that pooled 43 longitudinal studies covering more than 43,000 participants and used machine learning to find what genuinely predicts relationship satisfaction.
The answer was not your partner's traits. What predicted satisfaction was how the person felt about that specific relationship: perceived partner commitment, appreciation, sexual satisfaction, conflict frequency. Your own relationship-specific judgments carried far more weight than anything measurable about who your partner is.
That is less cinematic than a laboratory predicting your divorce from fifteen minutes of tape. It is also built on five hundred times as many people.
The Seven Principles for Making Marriage Work
John Gottman, PhD and Nan Silver · Harmony, revised edition
Evidence rating: ModerateWe are recommending this with the prediction claims set aside. The practical material, on repair attempts, on turning toward your partner in small moments, on distinguishing solvable problems from perpetual ones, is genuinely useful and reflects decades of close observation of couples. Read it as an experienced clinician's practice manual, which it is, rather than as a predictive science, which it is not. Skip the accuracy statistics when you reach them.
What to do this week
- If you have ever been told you display one of the four behaviours and worried about what it meant, you can put that down. The forecast attached to it was never validated.
- Notice repair attempts instead. When one of you tries to break the tension and the other accepts it, that is the mechanism the practical literature actually supports.
- If one of you pursues and the other withdraws during conflict, name the pattern out loud when you are not in it. That one is well established and responds to being named.
Sources
- Gottman, J. M., and Levenson, R. W. (2000). The timing of divorce. Journal of Marriage and the Family, 62(3), 737 to 745.
- Heyman, R. E., and Smith Slep, A. M. (2001). The hazards of predicting divorce without crossvalidation. Journal of Marriage and Family, 63(2), 473 to 479. doi.org/10.1111/j.1741-3737.2001.00473.x
- Kim, H. K., Capaldi, D. M., and Crosby, L. (2007). Generalizability of Gottman and colleagues' affective process models. Journal of Marriage and Family, 69(1), 55 to 72. pmc.ncbi.nlm.nih.gov/articles/PMC1828692
- Gottman, J. M., and Levenson, R. W. (1992). Marital processes predictive of later dissolution. Journal of Personality and Social Psychology, 63(2), 221 to 233. doi.org/10.1037/0022-3514.63.2.221
- Christensen, A., and Heavey, C. L. (1990). Gender and social structure in the demand/withdraw pattern. Journal of Personality and Social Psychology, 59(1), 73 to 81. doi.org/10.1037/0022-3514.59.1.73
- Joel, S., et al. (2020). Machine learning uncovers the most robust self-report predictors of relationship quality. PNAS, 117(32), 19061 to 19071. doi.org/10.1073/pnas.1917036117
This article is for educational purposes only and is not a substitute for professional care. If your relationship is in difficulty, a licensed couples therapist is a reasonable next step regardless of what any prediction model says.
Correction, September 2026: an earlier version of this article stated that these four behaviours predict divorce with over 90 percent accuracy, and that contempt is the strongest predictor of the four. Neither claim survives scrutiny. This article has been rewritten.
