What Actually Predicts Divorce? Sorting the Real Findings From the Overstated Ones

There is a widely repeated claim that psychologists can watch a couple argue for a few minutes and predict divorce with 90-plus percent accuracy. It gets cited constantly, it sounds like real science, and it isn't supportable in the way people think it is. The models behind that number were tested on the same couples used to build them — never on couples the model hadn't seen before. When a proper cross-validation test was finally run, accuracy did not hold up (Heyman & Slep, 2001, *Journal of Marriage and Family*). That's not a minor asterisk. It's the difference between a finding and a party trick, and it's a good place to start this series, because getting it wrong is exactly the kind of thing an unsourced blog does.

This piece is about what survives once that number is set aside: which patterns in how couples relate to each other are genuinely, repeatedly linked to divorce; which demographic factors carry a real but modest signal; and which popular beliefs about why relationships end don't survive contact with the data.

The Real Finding: Four Behaviors, Four Independent Samples

The part of this research that does hold up is John Gottman's observational work on what he called the "Four Horsemen" — criticism, contempt, defensiveness, and stonewalling, observed directly in how couples talk to each other, not self-reported. This isn't a one-off result. It's been seen across several independent longitudinal samples run by the same research group over almost a decade:

• [Buehlman, Gottman & Katz (1992)](https://doi.org/10.1037/0893-3200.5.3-4.295), *Journal of Family Psychology*, 52 couples — how couples narrated their own relationship history predicted, years later, which marriages were still intact.
• [Gottman, Coan, Carrère & Swanson (1998)](https://doi.org/10.2307/353438), *Journal of Marriage and the Family*, 130 couples — newlyweds' conflict-discussion behavior predicted marital stability and satisfaction at follow-up.
• Carrère, Buehlman, Gottman, Coan & Ruckstuhl (2000), *Journal of Family Psychology*, 95 couples — even the first three minutes of a conflict conversation carried signal.
• [Gottman & Levenson (2000)](https://doi.org/10.1111/j.1741-3737.2000.00737.x), *Journal of Marriage and Family*, 79 couples followed for 14 years — the longest run of the group, and the most interesting one. It distinguished two divorce patterns: couples who split relatively early in the marriage tended to show open hostility — escalating criticism and contempt during conflict — while couples who split only after many years together often showed no such fireworks. What predicted their divorce was a flatter emotional register in everyday conversation: little warmth or shared positive affect, even in the absence of visible fighting.

Four separate samples, run over roughly a decade, pointing the same direction: contempt in particular, and the broader pattern around it, tracks with relationship distress and divorce more consistently than most single variable in this literature. That's real, and it's worth knowing.

Where the 90% Number Comes From — And Why It Doesn't Hold Up

Here's the part almost nobody explains, and it's the most useful thing in this article: the "accuracy" figures attached to Gottman's work are in-sample. The model was built using data from a specific group of couples (the 52 to 130 in the studies above), and then its "predictions" were checked against outcomes from that same group — the group whose behavior had already shaped the model's rules.

That's the statistical equivalent of writing a test after seeing the answer key. A model tuned on 130 specific couples will always look impressively accurate when you re-check it against those same 130 couples, because it was built, in part, out of their particular quirks. The real test — the one that tells you whether a pattern generalizes — is running the model against a fresh set of couples it has never seen. That test is called cross-validation, and it's standard practice for a reason: it's the only way to tell a genuine pattern from the model memorizing noise in one specific dataset.

That test was not run before "90%+ accuracy" became a talking point. When something close to it finally was — [Heyman & Slep (2001)](https://doi.org/10.1111/j.1741-3737.2001.00473.x), *Journal of Marriage and Family* — properly cross-validated prediction did not come anywhere near the original figure. The paper is a direct methodological rebuttal, and its title says the quiet part out loud: divorce prediction is hazardous without cross-validation. The concern wasn't confined to academia, either — journalist Laurie Abraham raised the same problem in *Slate* in 2010, and statistician Andrew Gelman has made the identical point publicly: a model graded on the data that trained it will always look better than it performs on anyone else.

None of this erases the finding in the section above. Contempt, criticism, defensiveness and stonewalling really do correlate with distress and divorce across several independent samples — that part replicates. What doesn't hold up is the idea that a therapist, or an app, or a quiz can watch a couple for a few minutes and tell them their odds with 90%-plus confidence, as a validated, generalizable tool. Nobody has shown that it can.

What Demographic Factors Actually Predict

Outside the behavioral research, family sociology has a well-replicated set of demographic correlates of divorce. They're real. They're also smaller than people assume, and none of them isolate cause from confound.

Age at first marriage. Sociologist Nicholas Wolfinger's analysis of National Survey of Family Growth data (Institute for Family Studies, 2015) found a U-shaped pattern: elevated divorce risk for people who married very young, risk ticking back up for people who married in their thirties, and a lower-risk middle band in between. The pattern is real in the data. It is not evidence of a scientifically proven best age to marry — the analysis can't separate age itself from what tends to travel with early or late marriage, like financial instability, relationship inexperience, or simply having stayed single by choice for reasons unrelated to age.

Parental divorce. People whose own parents divorced are somewhat more likely to divorce themselves, a pattern researchers call intergenerational transmission (Amato, 1996, *Journal of Marriage and the Family*). It's consistently replicated and modest in size, and it runs through multiple plausible channels that have nothing to do with fate: children of divorced parents tend to marry somewhat younger on average, may have less modeled experience of resolving conflict, or may share the same economic pressures that shaped their parents' marriage.

Education and income. Government survey data has repeatedly found lower divorce probability associated with a bachelor's degree and higher household income (Bramlett & Mosher, 2002, National Center for Health Statistics). Again, this is association, not a mechanism. Financial strain is a well-documented source of marital conflict, but it's tangled up with who marries whom, at what age, and under what pressure — no single study in this literature cleanly separates the cause from the co-occurrence.

The Cohabitation Myth

For years the standard line was that living together before marriage raises your odds of divorcing — the "cohabitation effect." It's still repeated as settled fact. The best available meta-analysis doesn't support it in the simple form it's usually stated.

[Jose, O'Leary & Moyer (2010)](https://doi.org/10.1111/j.1741-3737.2009.00686.x), *Journal of Marriage and Family*, pooled the cohabitation research and found that the raw association is mostly about who cohabits, not what cohabiting does to a relationship. People who lived only with the person they went on to marry showed no meaningful difference in marital stability compared with people who didn't cohabit before marriage at all. The apparent effect concentrates among people who cohabited with more than one partner before marrying — a group that differs from single-partner cohabiters in ways (broader relationship history, different selection into serial relationships) that plausibly explain the added divorce risk on their own. Living with your eventual spouse first isn't the red flag the headlines suggest. For more details, see our guide on view stories anonymously. For more details, see our guide on how long Instagram stories last.

Method

Every research claim above is drawn from a named peer-reviewed journal article or government report, cited inline with author, year, venue, and sample size where one exists — deliberately, since "studies show" is exactly the habit this series exists to break. All are searchable by author and year in Google Scholar.

Separately, we maintain our own dataset on relationship-language search interest, built from the public Wikimedia Pageviews API. It measures interest, not incidence — someone reading Wikipedia's "Divorce" article is not necessarily getting divorced — so we're not using it to argue causation, only to show a real, checkable pattern. English Wikipedia's "Divorce" article (3,745,559 views across the 11 years in our sample) runs 11.8% above its own seasonal baseline in November and 11.3% below it in June; "Breakup" peaks earlier, in March (+10.9%), and troughs in July (−13.0%). We control for each wiki's overall seasonal rhythm before comparing months, because without that control almost every topic on Wikipedia looks like it peaks in January, simply because total traffic does. Full data, all twelve months, and seven language editions: relationship-seasonality.json and relationship-seasonality.csv. A companion file tracks search interest in related vocabulary — ghosting, love bombing, limerence, situationship — over the past decade: relationship-vocabulary.json. Anyone can reproduce or extend either file by querying the same Wikimedia REST API for a different article or period.

What This Actually Means

If you're looking for a test that can tell you whether a relationship will make it, the honest answer is that one doesn't exist — not because no one has looked, but because the tool that got closest never held up once someone checked it properly. What does hold up, across four independent samples run over roughly a decade, is that certain ways of fighting — contempt especially — track with distress and separation more reliably than any single demographic factor in this literature. That's worth paying attention to in your own relationship. It was never a stopwatch you could point at someone else's.