The Habits of Couples Who Stay Together: What the Longitudinal Research Measured

What research on couples who last can and cannot measure

Before any list of habits, it is worth knowing what the outcome variable in this literature actually is, because it is cruder than the writing around it suggests.

Most of what follows comes from longitudinal studies: measure something now, come back in a year or six, and see who is still together and how they are doing. Benjamin Karney and Thomas Bradbury reviewed that literature in 1995 across 115 longitudinal studies representing more than 45,000 marriages, and their summary of it is still the field's organising frame: enduring vulnerabilities each person brings, the stressful events the couple meets, and the adaptive processes they use to handle them, with marital quality and stability downstream of all three.

The crude part is that staying together and being happy are different outcomes, and they come apart constantly. Justin Lavner and Bradbury followed 464 newlywed spouses through eight satisfaction reports over four years and then checked divorce rates at four and ten years. Five distinct trajectory groups appeared. Three of them were relatively satisfied with small or no declines, and the most contented group - about 13% of husbands and 20% of wives - stayed exceptionally satisfied throughout. Dissatisfaction was concentrated in two groups making up roughly 19% of the sample.

The divorce numbers are the part worth holding onto. After four years, 3% to 14% of spouses in the three satisfied groups had ended their marriage, against 25% to 54% in the two least satisfied. After ten years the satisfied groups were at 9% to 26%. In other words, a considerable number of couples who negotiated the early years well split up anyway, and a considerable number of visibly unhappy couples were still married a decade later.

So nothing below is a formula for a relationship lasting, and lasting is not the same as good. What the research can offer is a set of behaviours that show up repeatedly on the right side of both outcomes, with the strength of the evidence stated honestly each time. Where the famous version of a finding is stronger than the study behind it, this page says so, and the retractions are logged in our corrections record.

The largest count of what predicts a breakup puts personality near zero

If you wanted one table to summarise this field, it would be the one Benjamin Le, Natalie Dove, Christopher Agnew, Miriam Korn and Amanda Mutso published in 2010. They pooled 137 studies of non-marital relationships conducted between 1973 and 2006, starting with 37,761 participants, and asked which measures taken at the start distinguished the couples who later broke up from those who did not.

The strongest predictors were all about the relationship rather than the people in it. Commitment, measured across 58 studies, produced a standardised difference of d = -0.80: the people who stayed had reported substantially more of it. Love, inclusion of the other in the self and dependence sat in the same range. Perceived quality of alternatives pushed the other way, at d = 0.57. Satisfaction managed d = -0.52. Support from friends and family, the one external factor that held up, came in at d = -0.50 across 11 studies.

Then the individual differences. Neuroticism produced d = 0.014. Extraversion, -0.008. Conscientiousness, 0.017. Agreeableness, -0.046. All four confidence intervals crossed zero. The only personality-adjacent measures with any signal were the relational ones: attachment avoidance at 0.21 and attachment anxiety at 0.14, both small.

That is the same shape Samantha Joel and 85 co-authors found for relationship quality in 2020, pooling 43 longitudinal datasets and 11,196 couples: a person's own relationship-specific reports carried almost all of the predictive weight, and traits measured outside the relationship carried far less. The two projects were asking different questions - who breaks up, and who is satisfied - and they converge on the same answer. It is a finding with a practical edge for anyone auditing themselves or a partner for the right qualities, which is the subject of why intelligent people keep choosing badly, and it is why the attachment style test on this site is careful to say that a result describes a tendency inside relationships rather than a verdict on a person.

One number from that table matters for everything below. Conflict predicted dissolution at d = 0.16 under fixed-effects assumptions across 14 studies, rising to 0.42 under random effects with substantial heterogeneity. It predicts something. It predicts much less than commitment, love or closeness, and far less than the genre's focus on arguing would suggest.

The limits are worth stating plainly, because they apply to most of this page. Participants in those 137 studies were 81% White, 96% heterosexual, about 25 years old on average, and almost all in the United States or Canada.

They answer when the other person reaches for them

The smallest unit of a relationship is not the conversation. It is the moment one person makes a low-stakes attempt to get the other's attention - reading something out loud, pointing at a bird, saying huh at their phone - and the other person either acknowledges it or does not.

Gottman's lab named these bids, and built a coding system for them from 600 hours of videotape of couples living in an apartment laboratory for 24 hours. What the peer-reviewed papers using that system actually report is narrower than the version you have read.

Janice Driver and John Gottman, in 2004, coded a 10-minute dinnertime segment for 49 newlywed couples plus a 15-minute conflict discussion. Playful bids and enthusiastic responses at dinner were associated with humour and affection during the argument, and the path model fitted better running from the ordinary moments to the conflict than the other way round. The authors are explicit that the work is observational and correlational and cannot establish direction.

The companion paper, published by Gottman and Driver in 2005 with 50 couples, is the more interesting one. It pitted a personality account - a person who turns away at dinner is the same person who withdraws in an argument - against the idea that conflict grows out of what went unsaid earlier. The personality account failed: none of the four within-person correlations reached significance. What did hold up was cross-partner. The husband's rate of turning away at dinner correlated .44 with the wife's attack-defend behaviour during the conflict discussion. Ignoring small bids did not predict his own conflict style. It predicted hers.

Now the number. The claim in circulation is that couples still together six years later had turned toward bids 86% of the time while those who divorced managed 33%. It is everywhere, and we could not find it in the peer-reviewed papers that report this coding system - those papers report correlations between behaviour in two settings, with no divorce follow-up and samples of 49 and 50 couples. The earliest widely read source for the figures is a 2014 Atlantic article by Emily Esfahani Smith, which gives the pair as 87% and 33%, describing a six-year follow-up. Treat the mechanism as reasonably supported and the percentages as a popular claim, not a measurement you can check.

What the evidence does support is smaller than the slogan and still usable: acknowledging costs almost nothing, the ordinary failure mode is distraction rather than coldness, and the person paying the price for a missed bid may be the one who missed it. The structural version of the same problem - what a phone does to the number of bids that even get made - is in what phones do to a conversation.

Good news gets a real response

Almost all popular relationship advice concerns the hard moments. The measured evidence says the ordinary good ones do at least as much work.

The process has a name, capitalization: sharing something good with someone who responds in a way that amplifies it rather than flattening it. Shelly Gable, Gian Gonzaga and Amy Strachman had 79 dating couples take turns discussing recent positive and negative events on video, with the discloser rating how understood and cared for they felt and outside observers coding the responder's behaviour. Two months later, responses to the positive-event discussions were more closely related to relationship well-being and to break-up than responses to the negative-event discussions were.

That result is easy to dismiss as a study of undergraduates with little at stake, so the second one matters. Amy Otto and colleagues studied 99 couples in which the woman had early-stage breast cancer, with both partners keeping electronic daily diaries for seven to ten days. On days when someone shared the best event of the day, the sharer's feelings of intimacy were higher - for patients and for partners. When the listener was perceived as responsive to that sharing, the sharer's intimacy rose further and their negative affect fell. All of it held above and beyond the effect of sharing negative events, in couples with no shortage of negative events to share.

The failure mode here is not cruelty. It is the flat acknowledgement - that's nice, good - which receives the news and does nothing with it. What was coded as the good version is active and interested: asking about it, being visibly pleased, returning to it later.

It is also the cheapest thing on this page, because it costs about fifteen seconds and the opportunity arrives several times a day. The longer-run version of the same mechanism, including how much of what one partner does is never noticed at all, is in what actually builds attraction over time.

Appreciation has to arrive, not just exist

Gratitude research in relationships has grown quickly and unevenly, and the useful distinction inside it is between gratitude that is felt, gratitude that is expressed, and gratitude that is received.

Sara Algoe, Shelly Gable and Natalya Maisel sampled the daily experience of both members of cohabiting couples and found that gratitude arising from an interaction predicted increases in relationship connection and satisfaction the following day, for the person who received the benefit and for the person who provided it. Indebtedness - the sense of owing something back - did not do the same work.

Amie Gordon, Emily Impett, Aleksandr Kogan, Christopher Oveis and Dacher Keltner followed this to a stability outcome. In their second study, 99 people kept daily diaries and were recontacted nine months later; 51 responded and 14 of those had broken up. People who felt more appreciative of their partners at baseline were 3.58 times more likely, per scale point, to still be in the relationship nine months on, and they were less likely to report having thought about or discussed a breakup. The sample is small, young and self-reported, and the effect is a correlation - but it survived controlling for satisfaction, which satisfaction alone did not manage in the other direction.

The study that tests the distinction most directly is Allen Barton and colleagues' 2022 work with 316 African American couples across three waves over roughly 16 months. They separated perceived gratitude - feeling appreciated - from expressed gratitude - being appreciative - and asked which one buffered the damage from ineffective arguing and financial strain. Only perceived gratitude did. Expressed gratitude, measured as the giver's own report, did not.

The practical shape of that is specificity rather than frequency, and it belongs to the same family of findings as what makes people say the real thing rather than the polite one. Thanks is a courtesy. You took the call with my mother so I could finish is a description of something the other person actually did, and it is much harder to hear as automatic. Gratitude that arrives on a schedule stops carrying information, which is the same reason the relationship satisfaction test reports appreciation and commitment as separate scores instead of collapsing them into one feeling.

One correction belongs here. An earlier version of this page claimed that relationship power attenuates the effects of gratitude, and attributed it, with the wrong lead author, to the study above. That finding is not in it. The claim has been retired; the record is in our corrections.

Most of their arguments are never going to be solved

One of the more freeing observations in this literature is also one of the least rigorously established, so it is worth reporting exactly.

Gottman and Robert Levenson, reporting in 1999 on how stable marital interaction was across a four-year gap, noticed that couples were largely having the same conversations at both time points. In a post hoc coding of the content of those discussions, couples were arguing about the same type of issue 69% of the time, and about new issues 31% of the time. The authors call these perpetual issues and describe them as rooted in fundamental differences between the two people - one is much more social than the other, and no amount of discussion changes that.

What they noticed next is the part that matters. Around the same unsolvable problem, some couples were almost amused - laughing, listening, affectionate - while others were hurt, angry, belligerent and contemptuous. Their own conclusion was that the importance of problem-solving may have been overemphasised in marital research, and that what varies is the emotional texture of a disagreement that is never going to end.

Hold that loosely. It was a post hoc content coding presented alongside a clinical impression, in a small sample over four years, not a designed test of how many problems are solvable. But it is consistent with the much larger finding from the meta-analytic table above: conflict was among the weakest relationship-level predictors of who broke up.

If you have been reading the recurrence of an argument as evidence of failure, the recurrence is the normal case. It is also an argument for finding out early what the recurring disagreements are going to be, which is what the questions that actually reveal compatibility are aimed at. The thing to watch is the register on the way in, which is the subject of what the repair research supports doing after a fight, and which our conflict style test scores as its own dimension precisely because how a couple argues appears to matter more than how often.

The five-to-one ratio, and what it actually counted

This is the most quoted number in the field and the most misunderstood, and the original paper is clear enough to settle it.

Gottman and Levenson studied 73 married couples in 1983 and again in 1987. Each couple had a 15-minute conflict conversation coded turn by turn with a checklist system. Couples were sorted by the cumulative slope of positive minus negative speaker codes into regulated - both partners' slopes significantly positive - and non-regulated. Over the four years, 18 of the 73 couples separated and 9 divorced.

The ratio appears in the discussion: stability appeared to rest on the ability to maintain a fairly high balance of positive to negative behaviour, positive to negative ratios of approximately 5.0 in those data, rather than on the absence of negativity. That last clause is the important one. Every couple in the study, happy or not, produced negative behaviour.

Now the part that almost never travels with the number. The five positive speaker codes were neutral or positive problem description, task-oriented relationship information, assent, humour or laughter, and other positive. Describing the problem in a neutral tone counted as positive. Saying mm-hm counted as positive. This was never a ratio of compliments to criticisms, and it was never a ratio across a week of ordinary life - it is a count of speaking turns inside a fifteen-minute argument in a laboratory.

So the honest reading is that couples heading for stability kept a running balance in which ordinary, neutral, cooperative talk outweighed the attacking kind during a disagreement, in one American sample of 73 couples, decades ago. There is no evidence that a couple can improve anything by counting to five, and treating it as a budget - five nice things earn one cutting one - is a misreading in both directions.

The accuracy claims that grew out of this line of work have a separate problem. Richard Heyman and Amy Smith Slep showed in 2001 that divorce-prediction percentages derived and tested on the same sample are inflated by construction, which is covered in more detail in the article on repair after conflict. The descriptive coding is the durable contribution here. The percentages are not.

Contempt is the one to act on first

Of everything on this page, this is the item to take seriously first if it applies to you.

Contempt is not anger and it is not criticism. It is the register that positions one person above the other: the eye roll, the mockery, the sarcasm that is not affectionate, the explanation delivered slowly. In the 1992 comparison above, the non-regulated couples were more angry, more whining, more withdrawn as listeners, less affectionate, less interested in their partners and less joyful than the regulated ones - a cluster rather than a single behaviour, with contempt the most distinctive member of it.

The honest footnote is that the affective process models built on this work have had a rough time in replication. When Hyoun Kim, Deborah Capaldi and Lynn Crosby tested them on 85 married or cohabiting couples from a community sample with at-risk backgrounds, the major findings did not replicate: men's rejection of their partners' influence, the absence of de-escalation and women's negative start-up did not predict relationship status two and a half years on. The full account of what survived and what did not is in the evidence on emotional maturity.

What that changes is the confidence, not the direction. Nobody has produced evidence that contempt is harmless, and it has the additional property of arriving without a decision: it grows out of accumulated unaddressed resentment and feels justified from the inside by the time it is visible from the outside. Our conflict style test scores it separately for that reason, and the research on jealousy covers the escalation pattern it usually travels with.

And a line this page draws clearly. A pattern in which one person is systematically belittled, monitored, isolated from friends and family, controlled financially, or made responsible for the other's moods is not a communication problem and is not on this list. The United Kingdom's statutory guidance on controlling or coercive behaviour describes exactly those acts - enforcing rules with punishments, monitoring time, isolating someone from their support network, controlling money - and treats them as a pattern of abuse rather than a bad phase. None of the habits above are the right response to it. In the United States the National Domestic Violence Hotline is on 1-800-799-7233, or text START to 88788; Find A Helpline lists free, confidential services by country.

They read each other slightly more generously than the facts warrant

Here is a finding that cuts against the instinct to see a partner clearly, and a second finding that stops it from being advice.

Sandra Murray, John Holmes and Dale Griffin followed 121 dating couples over a year, measuring at three points how each person saw their partner and how their partner saw themselves. Forty-three couples broke up during that year, about 35% of the sample. Relationships were most likely to persist when the two people idealised each other most, and those who idealised more at the start reported greater increases in satisfaction and decreases in conflict and doubt across the year. The measure that did not predict stability at all was accuracy - understanding a partner's actual view of themselves had no relation to whether the relationship survived. When men idealised their partners most, early conflicts and doubts did not trigger breakups.

In the 137-study meta-analysis, positive illusions produced the single largest effect in the whole table, d = -0.99 under fixed effects, though from only five studies, which is why it is reported here after the more heavily replicated findings rather than before them.

Now the limit, and it is a serious one. James McNulty, Michelle O'Mara and Benjamin Karney followed 251 new marriages across the first four years and found that the benefit of benevolent cognitions - making charitable attributions, not letting a bad experience colour the whole relationship - depended on what was being minimised. In healthier marriages it helped. Among spouses facing more frequent negative behaviour and more severe problems, the same strategy predicted steeper declines in satisfaction, apparently by allowing the problems to go unaddressed and worsen.

So the defensible version is narrow: generosity about the small stuff, accuracy about the big stuff. It is also the reason this site is sceptical of early-stage checklists that ask you to grade a person: the markers with any evidence behind them are about what a relationship is doing over time rather than about how favourably you can describe someone. Reading a partner's lateness charitably is the thing with evidence behind it. Reading a pattern that is damaging you charitably is the thing that made troubled marriages worse in a four-year study. The same contingency shows up in how couples argue - blunt, direct negativity predicted better outcomes when problems were severe - and is covered in the research on emotional maturity.

They respond to bad behaviour with something other than matching it

Everybody in a long relationship eventually behaves badly. What happens in the next ten seconds has a name in this literature: accommodation, meaning the willingness to inhibit the impulse to react destructively and do something constructive instead.

Caryl Rusbult, Julie Verette, Gregory Whitney, Linda Slovik and Isaac Lipkus laid out the model across six studies in 1991. Accommodation was lower when social concern was reduced and when the two people were less interdependent. It was higher among people reporting more satisfaction, commitment, investment and partner perspective-taking, and poorer alternatives elsewhere. Commitment did much of the mediating work, which fits the meta-analytic table: commitment was the most reliable single predictor of who was still together later. And couple functioning was associated with joint and mutual tendencies to inhibit destructive reactions - not with one person absorbing everything.

The direction of causation is genuinely unclear here, and the 1991 studies cannot settle it. Committed people accommodate more; accommodating probably builds commitment; both are measured at once in most of this work.

What can be said is that the moment is identifiable while it is happening. Someone has just been short with you, or dismissive, or has said the thing they know lands. The destructive options are obvious and immediate, and the constructive ones require about two seconds of delay. Doing it once is not a personality; doing it as a rate over months is what the measure actually tracked.

The complementary finding, from the other side of the exchange, is that how a request is made changes how it is received: in a study of 180 couples coded every 30 seconds, anger and withdrawal fell in the intervals when the person asking for a change softened it - staying sensitive to the other person's autonomy and conveying that they were valued - while still asking. That work is covered in the article on emotional maturity.

They talk like a unit, and they laugh at the same moment

Two markers with real measurement behind them, and a caveat that applies to both.

Benjamin Seider, Gilad Hirschberger, Kristin Nelson and Robert Levenson recorded 154 middle-aged and older couples in a 15-minute conflict conversation with physiology monitored continuously, then coded the transcripts for pronouns. Greater use of we-words was associated with lower cardiovascular arousal, more positive emotional behaviour and less negative behaviour. Greater use of me-and-you words went with more negative behaviour and lower marital satisfaction. Older couples used more we-words than middle-aged ones.

Alexander Karan, Robert Rosenthal and Megan Robbins pooled 30 studies in a meta-analysis and found the association held overall, across relationship outcomes, relationship behaviours and health. The detail worth keeping is that a partner's we-talk was generally more strongly related to a person's functioning than their own was - which is a strange result for anyone hoping to use this as a technique.

And that is the caveat. These are markers of how interdependent two people already feel, not instructions. No study here shows that deciding to say we changes anything, and saying it deliberately in an argument you do not feel that way about was never what was measured.

The same applies to the laughter finding, which is charming and narrow. Laura Kurtz and Sara Algoe coded spontaneous laughs in 71 couples talking about how they first met, and found that the proportion of the conversation spent laughing simultaneously with the partner was uniquely associated with relationship quality, closeness and perceived support - controlling for all the other laughter in the room. One person laughing tells you little. The overlap is what tracked.

It is cross-sectional, so it cannot say whether shared laughter builds anything or simply shows what is already there. What it does do is give a way of noticing, which is more than most of this genre offers, and it matches something that shows up wherever play is measured between people: the playful register does not stop at the beginning of a relationship. The research on flirting between friends and partners covers the same behaviour from the other end.

What to watch is warmth leaving, not conflict arriving

If all of the above has to reduce to one question you can ask yourself in any given month, this is the section that supplies it.

Ted Huston, John Caughlin, Renate Houts, Shanna Smith and Laura George followed newlywed couples and checked back thirteen years later. Changes over the first two years did foreshadow the outcome. But the results gave little support to the idea that distress emerging early - arguments starting, negativity increasing - leads to failure. What distinguished the couples heading for divorce was disillusionment: love abating, overt affection declining, the conviction that a spouse is responsive weakening, ambivalence rising.

Irene Tsapelas, Arthur Aron and Terri Orbuch approached the same thing from the boredom side, following 123 married couples from year seven of marriage to year sixteen. Greater boredom at year seven - a single question about whether the marriage felt like a rut - predicted significantly less satisfaction nine years later, and the association ran through closeness.

Put those beside the meta-analytic table, where conflict was one of the weakest predictors and love, closeness and commitment the strongest, and the same instruction falls out of all three: the early warning sign is not that you have started fighting. It is that the warmth is going quiet. A version of this argument aimed at newer relationships, where people tend to look for milestone checkpoints instead, is in the rules with numbers in them, and the version aimed at working out whether something is already over is in the signs you have moved on.

One related number, because it gets misused in both directions. Amy Muise, Ulrich Schimmack and Emily Impett analysed three datasets totalling 30,645 people and found that the association between sexual frequency and well-being is curvilinear: for people in relationships it rises and then flattens, with no further association above about once a week. Sexual satisfaction was among the strongest predictors in the 11,196-couple project; frequency past that point was not associated with more of anything. The two are not the same variable, and treating frequency as a scoreboard confuses them. Warmth leaving is also the variable that separated the couples who divorced in the longest study of the newlywed years, which is the finding at the centre of why the honeymoon phase ends.

What follows from all of this

Five things, and none of them are techniques.

Acknowledge the small approach. The bids research does not support the percentages attached to it, but the underlying measurement - ordinary moments predicting the emotional weather of an argument weeks later, and predicting it across partners rather than within a person - is real. Looking up costs nothing.

Do something with good news. It is the only item here where a two-month study found responses to positive events mattering more than responses to negative ones, and it is the cheapest to act on.

Make appreciation land. The variable that buffered stress across 316 couples over 16 months was feeling appreciated, not being appreciative. That is a fact about specificity and delivery, not about how grateful you privately are.

Be generous about small things and accurate about large ones. Idealisation predicted couples surviving a year. Benevolent reinterpretation of severe problems predicted marriages getting worse over four. Both are in the evidence and the difference between them is what is being minimised.

Track the direction of warmth, not the count of arguments. Two long follow-ups and one large meta-analysis agree that affection and closeness draining away is the signal, and that conflict on its own is a weak one.

And one thing to resist. None of this works as behaviour performed to produce a result in another person. Nearly every measured effect on this page is an effect on both people at once - gratitude raised the next day for giver and receiver, accommodation associated with joint rather than one-sided tendencies, we-talk predicting a partner's functioning more than one's own. A version of these habits performed unilaterally, indefinitely, while keeping score, is not what any of these studies measured.

What this evidence cannot tell you

Four limits, stated because the alternative is a page that sounds more certain than the literature it rests on.

The samples are narrow. The meta-analysis of breakups was 81% White, 96% heterosexual and almost entirely North American. Most of the observational work is married, middle-class and Western, and where it has been tested outside that slice - as with the 85 at-risk community couples - parts of it have not replicated.

Almost none of it is experimental. You cannot randomly assign people to appreciate each other or to turn away at dinner. Nearly every finding here is a correlation, sometimes with the strong advantage of a lagged or longitudinal design, sometimes not. Where a study can only show that two things co-occur, this page has tried to say so rather than to imply direction.

Group averages do not describe a couple. Even the strongest predictor in the table, commitment at d = -0.80, leaves enormous overlap between the people who stayed and the people who left. A relationship is not a probability, and no number on this page tells you what will happen to yours.

None of it says when to leave. The research is about what distinguishes trajectories, not about what anyone should do. A relationship can be statistically unremarkable and right for the two people in it, and it can look fine on every measure here and still be one somebody needs to end. If what you recognise in yours is the pattern described in the section on contempt - belittling, monitoring, control, isolation - the next step is a conversation with someone qualified or one of the services listed there, not a change of communication technique.

If you want something more structured than a reading list, the relationship satisfaction test is built around the variables that actually carried weight in the 11,196-couple analysis, and it says on the page what a score cannot tell you. More of this under Mindset.