Why People Stay in Unhappy Relationships: What the Stay or Leave Research Found

What the search engines answer, and what their own source measured

The question in the title was put to three search engines on 19 October 2026, and the first thing two of them returned was not a page. It was an answer.

Brave's summary gives five reasons: fear and uncertainty, the sunk cost fallacy, interdependence and guilt, practical constraints, and unresolved childhood trauma. DuckDuckGo's gives fear of the unknown, money, a belief that nothing better exists, duty, and occasional moments of joy that keep people invested. Google returned the unusual-traffic challenge that day and so contributed nothing, which is its own kind of answer.

One line in Brave's list is doing something different from the other four. “Research shows people often stay to protect a dependent partner from distress, driven by guilt, responsibility, or a desire to avoid causing harm.” There is a real study behind the first half of that sentence, and it is the best thing in this literature. The second half is the part that study tested and did not find.

Underneath the summaries: Psychology Today three times, Forbes, InsideHook, Elite Daily, Medium twice, Quora three times, four Reddit threads, Mumsnet, CNN, several therapy practices, and the news offices of the University of Toronto and the University of Utah. On DuckDuckGo, Marriage.com twice, The Modest Man twice and Simply Psychology. No Healthline, no Cleveland Clinic, no Verywell Mind — the same split this site has measured repeatedly, where the explanatory questions in this subject are left to lifestyle media and forums. The precedent for what the generated answers do is recent: a week earlier the same engines supplied a causal mechanism for the honeymoon phase ending sourced to a podcast and to TikTok, and the week before that an orphan figure about attachment style change that could not be traced to any study.

The two university pages are there because they are covering the same study Brave is paraphrasing. Almost everything else is a numbered list: three reasons, seven reasons, eleven reasons, fifteen reasons. And a comment in one of the Reddit threads names the investment model of commitment, correctly, with no citation and no figures — a better answer than most of the pages ranking above it.

So the gap is not that nobody knows. It is that the measured answer is more specific than the lists, that in two places it runs the other way, and that not one page carries a number from the studies it is summarising.

“Unhappy” is one measurement. “Staying” is a different one

The question assumes that satisfaction is the variable, and that staying while unsatisfied therefore needs explaining away. The research that addresses it directly starts from a different place: satisfaction and commitment are measured separately, because they come apart.

Caryl Rusbult's investment model, first tested in 1980 and given a standard scale by Rusbult, Martz and Agnew in 1998, holds that commitment to a relationship rests on three things that are not the same thing: how satisfying it is, how the alternatives to it look, and how much has been invested in it. A meta-analysis by Benjamin Le and Christopher Agnew pooled 52 studies, 60 independent samples and 11,582 people and found all three associated with commitment, together accounting for close to two-thirds of its variance, with commitment in turn predicting breakup.

What that framework buys is a plain answer to the title with no pathology in it. A person can be unhappy and highly committed without anything being wrong with their judgement, because two of the three supports of commitment have nothing to do with how good the relationship feels. That is why our relationship satisfaction test scores alternatives and investment as separate numbers instead of folding them into one verdict, and it is set out there and in the section on not leaving in why intelligent people repeat the same relationships. This page does not restate it.

What this page is for is everything the framework does not cover, and the honest starting point is that the framework's own authors kept finding things outside it.

It is worth being clear about what kind of claim any of this is. These are self-report measures, collected mostly from people in Western countries, correlated with whether a relationship later ended. They describe regularities across thousands of people. None of them is a reading of a particular relationship, and none of them tells anybody what to do — a limit the last section comes back to, because on this question it is not a formality.

The study that followed people who were already thinking of leaving

Most research on breakups samples people in relationships and waits to see who ends up single. Samantha Joel, Emily Impett, Stephanie Spielmann and Geoff MacDonald did something narrower and much more useful: they advertised for people who were actively contemplating a breakup, measured them, and came back two months later.

Recruitment ran through Facebook, Twitter, Reddit and Psychology Today. Of 4,106 people who attempted to take part, 1,881 met the criteria and 500 formed the final preregistered sample — 408 women, 85 men, 7 non-binary, mean age 32, mean relationship length 38 months. Eighty-eight per cent said at screening that they were currently thinking about breaking up. By the two-month follow-up, 145 of them (29%) had ended the relationship.

Then the result that reframes the question. In the model containing the four investment-model components plus one more variable, here is what predicted who actually left:

- Satisfaction: did not predict. b = −.07, p = .50. - Quality of alternatives: did not predict. b = .12, p = .27. - Investment: predicted. b = −.24, p = .008. - Own commitment: predicted. b = −.28, p = .01. - Perceived partner commitment: predicted. b = −.25, p = .007.

Among people already at the point of questioning the relationship, how satisfied they were did not distinguish those who left from those who stayed. Nor did how good their options looked. What distinguished them was what they had tied up in it, what they intended for it, and how committed they believed the other person was.

That is one sample, in one two-month window, of people who had selected themselves into a state of doubt — a group in which satisfaction is compressed to begin with, which is part of why it stops predicting. It is not a claim that satisfaction is irrelevant to relationships in general; Le and Agnew's meta-analysis is clear that it is the strongest single correlate of commitment. It is a claim about the moment the title is asking about, and in that moment the variable everyone names is the one that went quiet.

Staying for the other person, at the size it actually is

The same paper's first study is the one the university press offices covered, and it is where the figure in every summary comes from. It is worth having at full size, because the headline travels without its numbers.

Joel and colleagues recruited 3,952 people in relationships and invited them to answer a weekly email asking one question: are you and your partner still together? A total of 1,348 took part in the follow-up, with 1,281 in the final analysis, and 174 of them (14%) broke up during the study.

Two measures of how much the partner needed the relationship predicted staying. On its own, someone who perceived their partner as committed one standard deviation above the mean was 43% less likely to initiate a breakup at any point (hazard ratio .57). Someone who expected a breakup to distress their partner by one standard deviation above the mean was 39% less likely (hazard ratio .61).

Those are large, and they shrink when the obvious explanations are controlled — which is the interesting part, because they do not vanish. Entered alongside satisfaction, alternatives, investment, commitment and feeling appreciated by the partner, perceived partner commitment still carried 24% and anticipated partner distress 28% per standard deviation. Neither effect was moderated by gender or by relationship length. The headline of this work, which the piece on repeating the same relationships states in one line, is that people stay partly for the other person.

Two limits the authors put in their own paper. People who agreed to the weekly emails were more committed and more satisfied than those who declined (Cohen's d = .33 and .24), so the sample tilts toward better relationships. And in the preregistered second study, a purpose-built scale of partner-focused reasons — eleven items about what a breakup would cost the partner — was only marginal once the investment-model components were in the model. The authors say plainly that it was “not as robust a predictor as expected.” What held up was the simpler thing: the belief that the other person is committed.

It is worth resisting the obvious moral here, in either direction. The broader evidence on giving things up for a partner does not come out one-sided: Francesca Righetti, John Sakaluk, Ruddy Faure and Impett pooled 82 datasets and 9,547 effect sizes from 32,053 people and found the facets of sacrifice pulling apart. Willingness to sacrifice went with better personal and relationship well-being for both people (rs from .09 to .27). The costs of sacrificing went with worse well-being for both (rs from −.10 to −.26), and the act of sacrificing itself was mildly negative for the person doing it (r = −.07). Staying for someone else is neither noble nor pathological by default; what the measurements separate is wanting to from paying for it.

It is not guilt, and it is not fear of what they would do

This section exists because of one clause in a search engine's summary, and because the correction is more interesting than the error.

The folk account of staying for a partner's sake is an account of guilt. You know they would be devastated, you would feel like a monster, so you do not do it. Joel and colleagues anticipated that reading and built three scales for the second study to test it: anticipated guilt (how bad you would feel), anticipated retaliation (that the partner might become angry or aggressive, or damage your reputation) and anticipated negative judgment (that friends and family would think badly of you).

All three were entered alongside perceived partner commitment, in the preregistered model, in the sample of people already contemplating a breakup. None of the three predicted who stayed. Guilt, b = .10, p = .30. Retaliation, b = .13, p = .14. Judgment, b = −.10, p = .29. Meanwhile perceived partner commitment got stronger in that model rather than weaker, b = −.30, p < .001.

The authors' reading is that this is what makes the effect prosocial rather than self-protective. If people were staying to avoid feeling guilty, guilt should have carried the prediction. It did not. What predicted staying was a belief about the other person's stake in the relationship, not a forecast about one's own discomfort.

Two cautions, because this is a null result and nulls are easy to over-read. These are self-report scales built for this study, answered by people in distress, and a scale that does not predict may be measuring the construct badly rather than showing the construct does not matter. And the sample is small enough that a real but modest guilt effect could have been missed.

Still, the asymmetry is the point. The one part of this mechanism that the search summaries assert — driven by guilt — is the part that was measured and came back null, and the part they omit is the part that replicated across two studies and survived every control the authors could think of.

What people say when you let them answer in their own words

Scales constrain the answer. Denise Previti and Paul Amato did the other thing: they asked married people, in open-ended questions, why they stayed, and then followed them for fourteen years to see who was still married.

Two kinds of answer dominated, and they map onto a distinction this literature has carried since the 1970s. Rewards are the things a person gets from the relationship. Barriers are the things that make leaving costly. Most people described their marriage in terms of one, the other, or both.

The first finding is a correction to the first page of search results. Few people referred to a lack of good alternatives. The reason that sits near the top of every list — I would never find anyone else — was rare when people were allowed to say whatever they liked. That does not make the alternatives variable wrong; it makes it something people do not spontaneously narrate, which is a different fact and worth knowing before trusting anybody's stated reasons, including your own.

The second finding runs against the folk theory outright. People who named only barriers were, as expected, less happy and more likely to be thinking about divorce. They were also more likely to have divorced over the following fourteen years — and that association held after controlling for how happy the marriage was and for how divorce-prone the person was.

So barriers are not glue. In this dataset, being able to say only what it would cost you to leave was a leading indicator of leaving, not a guarantee of staying. The common image — the person held in place by the mortgage and the in-laws — has the causal arrow pointing the wrong way. Barriers are better read as a symptom of a relationship that has stopped supplying rewards, and the symptom appears years before the exit does.

This is one national sample of married Americans, coded from open-ended text, and the coding is an interpretive act. It is also, as far as we can find, the only study that asked the question in the title directly and then waited fourteen years.

Three kinds of constraint, and one of them runs the other way

“Constraint” gets used as a single thing. Galena Rhoades, Scott Stanley and Howard Markman split it into three in a national sample of 1,184 unmarried adults, followed over eight months, and the three behave differently enough that collapsing them loses the finding.

Dedication is wanting the relationship — interpersonal commitment, a shared future, treating the pair as a unit. Perceived constraints are the social and procedural costs of ending it: pressure from others, the difficulty of the termination itself. Material constraints are the concrete entanglements: a lease signed, a pet, a car. And felt constraint is the subjective part: feeling trapped.

All four uniquely predicted whether the couple was still together eight months later. But not in the same direction. More dedication predicted staying. More perceived and more material constraints predicted staying. And less felt constraint predicted staying — that is, the people who felt trapped were the ones more likely to be gone.

That pair of results is the most practically useful thing in this article, and it is almost never stated. The objective entanglements and the feeling of being stuck are not two names for the same condition. A shared lease keeps people together. The sense of being held by one does not. Two people can have identical constraints and differ entirely on whether those constraints register as a life or as a trap, and it is the registering that predicts the exit.

It also explains why the advice implied by most of the lists is so weak. Telling someone they are only staying because of the apartment is not news to them, and the research says the apartment is doing real work. The variable that moves with leaving is the one nobody can see from outside.

Eight months is a short window and these were unmarried couples, so the entanglements are smaller than a marriage's. Note also that this is the same research group whose work on drifting into commitments rather than deciding on them underlies what the research says about situationships, where constraints accumulate ahead of dedication rather than behind it.

Where the vocabulary came from, and why it still helps

The rewards-and-barriers language is older than the investment model, and the older version has a distinction worth borrowing.

George Levinger set out the frame in 1976: whether a marriage holds depends on the attractions inside it, the barriers around it, and the alternatives outside it. It is a theoretical paper rather than a study, and it is the ancestor of nearly everything above.

Michael Johnson, John Caughlin and Ted Huston then made a sharper cut. Commitment to a marriage, they argued and then measured, is not one thing but three, and people experience them as three different sentences.

Personal commitment is I want to. It is attraction to the partner, attraction to the relationship, and the degree to which the relationship has become part of one's identity.

Moral commitment is I ought to. It is a sense of obligation — to the person, to the marriage as an institution, to the idea that one finishes what one starts.

Structural commitment is I have to. It is the alternatives, the social pressure, the procedural difficulty of ending it, and what has been irretrievably put in.

The value of that split is diagnostic rather than moral. All three keep people in relationships. They are not interchangeable, they do not feel the same from the inside, and they point at different questions. I want to is about the relationship. I ought to is about a commitment made to a version of a person who may no longer exist in that form. I have to is about logistics, and logistics change.

This is also where a lot of unnecessary self-blame gets generated. Someone who notices that their answer is mostly I ought to often reads it as evidence of dishonesty or weakness. In this framework it is one of three ordinary components of commitment, present in most long relationships, and the research treats it as a measurement rather than a verdict. What makes it worth noticing is simply that it answers a question people often have not asked: not is this good, but which of the three is holding it up.

The alternatives you can see are shaped by the relationship you are in

“I could not do better” sounds like an observation. In this literature it behaves more like an output.

Dennis Johnson and Rusbult ran three studies on what happens to a person's view of other potential partners as commitment rises. The longitudinal one is the one that matters here. Over time, perceived quality of alternatives decreased among the people who stayed and increased among the people who left. In a second study, using a simulated dating service, the most committed people devalued alternatives most sharply exactly when the alternative was extremely attractive — that is, when it was most threatening. A third, a simulation experiment, found the same pattern. Across all three, devaluation tracked commitment more closely than satisfaction.

The mechanism has a perfectly good function. A relationship in which every attractive stranger is fully appraised is not a stable one, and derogating alternatives is one of the ordinary ways people protect something they want to keep. The problem is that the process does not check whether the relationship deserves protecting.

There is a modern wrinkle worth naming, because it points the other way. What the evidence on dating apps actually shows is that the sense of abundant alternatives is itself a poor guide to how many there are, and the research on how matches distribute is not reassuring in either direction. Perceived alternatives are not a count.

Run it forward in a relationship that has been unsatisfying for years and you get the sentence people arrive with. The belief that there is nothing better out there is partly a product of having stayed, built by the same mental process that keeps good relationships intact. It is not a neutral survey of the world, and the longer the staying has gone on, the less neutral it is.

Which is also why the finding in the previous section — that few people spontaneously cite a lack of alternatives — and this one are not in conflict. Alternatives are a poor thing to introspect about. They are shaped by commitment, they shift as commitment shifts, and they are one of the two numbers our relationship satisfaction test deliberately prints separately, with a note on the result that the reading is a perception rather than a census.

A caution on vintage and scope: these are late-1980s student and community samples, and a modern replication at scale does not exist that we could find. The longitudinal direction is the part that has aged best, and it is the part being leaned on here.

Reasons to stay are not the mirror image of reasons to leave

There is a hidden assumption in every list of reasons people stay: that the reasons to go are the same list with the signs flipped. Laura Machia and Brian Ogolsky tested that assumption across three studies totalling 993 people, including experimental work and longitudinal in-the-moment sampling, and it is not quite true.

Some things are symmetrical. Satisfaction is the most strongly endorsed reason to stay, and dissatisfaction is the most strongly endorsed reason to leave; the same variable is doing both jobs, which is what you would expect.

But other constructs sit lopsidedly on one side. Quality of alternatives showed up far more strongly as a reason to leave than as a reason to stay. Love showed up far more strongly as a reason to stay than its absence did as a reason to leave. The two motivations are not a single dial with a midpoint.

The authors note that this confirms something interdependence theory had asserted for decades without testing: that some outcomes supply motivation to stay and others supply motivation to go, and that they are not simply the positive and negative ends of one scale.

For anyone trying to read their own situation, that matters more than it sounds. The question what is keeping me here and the question what would make me go do not have complementary answers, and answering one does not answer the other. A person can have a long, honest list of reasons to stay and a separate, equally honest list of reasons to go, and neither list cancels the other out — which is a description of ambivalence that does not require anyone to be confused or in denial.

It also makes the genre's favourite move look worse. Setting out seven reasons people stay, implicitly as seven errors to correct, treats the stay column as the whole ledger. In this data it is one of two columns, measured separately, with different contents.

The caveats are the usual ones for this kind of work: these are endorsements of researcher-supplied reasons in mostly young samples, and what people endorse is not necessarily what moves them.

Fear of being single, measured rather than asserted

Fear of being alone is the first item on nearly every list on the first page of results, always unsourced. It is also, unusually for this subject, a construct with a validated scale and seven studies behind it.

Stephanie Spielmann, MacDonald, Jessica Maxwell, Joel, Diana Peragine, Amy Muise and Impett built the Fear of Being Single Scale and tested what it predicts. Three of their results are relevant here.

First, it is not anxious attachment with a different name. The effects held when anxious attachment was controlled — a distinction worth holding onto, since the genre routinely merges the two, and what the evidence says about anxious attachment and love covers the separate question of telling the feeling from the pattern.

Second, and most directly on topic: in a longitudinal study, fear of being single predicted a lower likelihood of initiating the end of a less satisfying relationship. An earlier study in the same paper found the same thing cross-sectionally as greater dependence on unsatisfying relationships. This is the folk reason, measured prospectively, and it survived.

Third, a detail that cuts against the obvious interpretation. Across two samples, fear of being single was unrelated to people's stated standards for a partner — with the single exception of consistently higher standards for parenting. It is not that people lower what they say they want. In the experimental studies, what fear of being single predicted was romantic interest in targets manipulated to be less responsive and less attractive, and in a speed-dating study it predicted being less selective in expressing interest while not predicting whether other daters were interested in return.

The companion finding is worth knowing before anyone treats leaving as the cure. Spielmann, MacDonald, Joel and Impett followed people through a breakup and found that fear of being single rose afterwards, regardless of who ended it, and that on the days it was strongest people longed for their ex-partner more and made more attempts to get back together. Lagged analyses pointed from the fear to the longing, not the other way. That same asymmetry between how an ending feels and what it means is the subject of the signs you have moved on and the signs you have not.

The sunk cost claim, tested on a relationship

Sunk cost is the most confidently repeated mechanism in this genre and the one most often asserted without a study. There is a study. Its result is more interesting than the slogan, because the first experiment failed in exactly the place people invoke it.

Sara Rego, Joana Arantes and Paula Magalhães gave 902 participants a scenario describing an unhappy relationship and a choice: stay or end it. They varied what had already been put in. Money and effort both increased the likelihood of staying. Time did not.

That is the opposite of how the idea is used. Nobody says I have spent too much money on this to leave. What people say is I have given him five years, and five years was the one investment that did nothing in the first experiment.

In a second experiment the authors changed the measure. Instead of a stay-or-go choice, 275 participants said how much more time they would be willing to put into a relationship. With that continuous measure, a sunk time effect appeared: people offered more future time where more past time had been invested.

So the honest summary is in two halves. Sunk cost in relationships is real and has been demonstrated. And the specific popular version — that the years you have already spent make you choose to stay — showed up only when the question stopped being about leaving and started being about how much more to give. Those are different behaviours, and the distinction is not a quibble: continuing to invest is not the same decision as declining to exit.

The design limit applies to both experiments and should be stated loudly: these are hypothetical scenarios about an invented relationship, answered by people who were not in one. Vignette studies are good at isolating a variable and poor at telling you what a person does at their own kitchen table. The investment component of the investment model, which is measured in real relationships and is one of the three numbers the satisfaction test reports, is the better-evidenced version of the same intuition.

“They might change” — what the research on trying says

Hope that the other person will change is on every list, always as an error. The research on deliberately trying to change a partner is small, but it says something more specific and less dismissive.

Nickola Overall, Garth Fletcher, Jeffry Simpson and Chris Sibley videotaped 61 couples while each person tried to produce a change they wanted in the other, coded the strategies along two dimensions — positive or negative, direct or indirect — and then measured actual change in the targeted feature every three months for a year.

The result is a near-perfect inversion of how the conversations felt. Direct strategies, both positive and negative, were rated as relatively unsuccessful immediately afterwards — and predicted more change over the following twelve months, as reported by the partner who had been targeted. Positive-indirect strategies — softening, hinting, encouraging — felt more successful at the time and predicted no later change at all.

So “nothing I say makes any difference” and “that conversation went badly” are not the same statement, and the second is a poor proxy for the first. The approach that produces change is the one that is uncomfortable while it happens, which is the same shape as the finding about repair attempts in what the research on reconnecting after a fight supports.

Three limits, and the third is the important one. Sixty-one couples is small. The study measured change in a specific feature each partner had nominated — an amount of time spent with friends, a way of handling money — not a transformation of character. And the same paper found that increases in problem severity forecast lower relationship quality over time, meaning the strategy is not a universal solvent.

None of that licenses waiting indefinitely, and it says nothing about behaviour that is not a trait to be negotiated. Where what somebody is hoping will change is monitoring, pressure or anger, that is a different subject, it has its own section further down, and what emotional maturity looks like when it is measured rather than described covers how little of the popular version survives testing. What this evidence supports is narrower: asking directly for a specific change is not naive, it is the thing that has been shown to work, and it tends to feel like failure on the day.

What staying costs, at the size the evidence supports

If the lists understate anything, it is this. Staying is treated as the neutral default, the thing that happens when nothing happens. The longitudinal data treat it as an exposure.

Alan Hawkins and Alan Booth used a nationally representative panel to follow marriages over twelve years, with a multi-item quality measure, and compared people who remained in low-quality marriages against people who remained in other continuously married ones. Remaining unhappily married was associated with significantly lower overall happiness, lower life satisfaction, lower self-esteem and worse overall health, with elevated psychological distress. They also found that people in low-quality marriages were less happy than people who had divorced and remarried, and lower in life satisfaction, self-esteem and health than people who had divorced and stayed unmarried. Their own conclusion is unusually direct for a sociology paper: there was no aspect of well-being on which the unhappily married were better off.

The health side has been synthesised. Theodore Robles, Richard Slatcher, Joseph Trombello and Meghan McGinn pooled 126 articles over fifty years and more than 72,000 people. Better marital quality went with better health, with mean effect sizes of r = .07 to .21, including lower mortality risk at r = .11. The authors make two points this site will repeat rather than soften: the effects are small, comparable in size to associations between health behaviours such as diet and health outcomes; and a subset of the clinical outcomes showed evidence of publication bias. The earlier review by Janice Kiecolt-Glaser and Tamara Newton describes the plausible routes — depression, health habits, and direct cardiovascular, endocrine and immune effects — and notes that the pathways differ by gender.

The causal direction is the obvious objection and it is a fair one. Unhealthy, unhappy people may have worse marriages rather than the reverse, and the panel designs narrow that gap without closing it. It is also worth separating this from ordinary decline: satisfaction falls slowly in most couples and that is not what is being described here, as the trajectory research on the honeymoon phase sets out, and the markers of a relationship that is working are specific rather than vibes-based in what the research on couples who last reports.

What the numbers do rule out is the version where staying is free. It is not free, the cost is measurable, and it is also smaller than the genre's language about toxicity implies.

What leaving costs, at the size the evidence supports

The other half of the ledger is measured too, and the lists that treat leaving as liberation are no more accurate than the ones that treat staying as death.

Rhoades, Claire Kamp Dush, David Atkins, Stanley and Markman used a within-subjects design — the same people before and after — on 1,295 unmarried adults aged 18 to 35, of whom 36.5% had at least one breakup over twenty months. Experiencing a breakup was associated with an increase in psychological distress and a decline in life satisfaction. Having cohabited and having had plans to marry both predicted larger declines in life satisfaction; having started seeing someone new predicted smaller ones.

One result from that study is counterintuitive enough to be worth sitting with. Higher relationship quality at the previous wave was associated with smaller declines in life satisfaction after the breakup — smaller, not larger. Leaving a better relationship hurt less, not more, on this measure. That is not what anyone predicts, and it is a reason to distrust intuitions about how a particular ending will feel.

Over a longer horizon, Richard Lucas used an eighteen-year panel of more than 30,000 Germans to track life satisfaction around divorce. Satisfaction drops as divorce approaches and then gradually rebounds — but the return to baseline is not complete. And prospectively, people who would later divorce were already less happy than people who would stay married, before either group married at all. The association between divorce and life satisfaction is therefore made of two things at once: differences that were there in advance, and lasting change after the event.

Put the two sections together and neither one wins. Staying in a bad relationship carries measurable costs. Ending a relationship carries measurable costs. The research describes both and arbitrates between them for nobody, which is why the next two sections are about the only studies that have tried to say which people came out ahead.

People overestimate how bad leaving will feel — and most of all the people least able to leave

There is one study that compared what people predicted a breakup would feel like against what it actually felt like, in the same people, before and after.

Paul Eastwick, Eli Finkel, Tamar Krishnamurti and George Loewenstein recruited 69 Northwestern University freshmen in relationships of at least two months and surveyed them every two weeks for nine months. The analysis rests on the 26 participants whose relationship ended during the first fourteen waves. Before the breakup, each had repeatedly forecast how distressed they would be two, four, eight and twelve weeks after an ending. After it, they reported how distressed they actually were.

They had overestimated. And the paper's contribution is where the error sat. There are two candidate explanations for this kind of mistake: people might misjudge how intense the initial reaction will be, or they might misjudge how fast it fades. The data showed an initial intensity bias and no decay bias. Even at the first report after the breakup, participants were already less distressed than they had predicted two weeks earlier. They were not wrong about the rate of recovery. They were wrong about the starting point.

Then the moderators, which is why this section is placed here. The overestimate was larger for participants who were more in love with their partner, who felt it was less likely they would soon begin a new relationship, and who played less of a role in initiating the breakup.

Read that back as a sentence about staying. The forecast of post-breakup devastation is least accurate for exactly the people whose forecast is doing most of the work — those who love the person, who cannot picture anyone else, and who would not be the one ending it.

The sample is twenty-six American undergraduates on one campus in relationships averaging fourteen months. That is small and it is young, and nothing here extends to a marriage of twenty years or to people whose material security is involved. What it supports is narrow and still useful: a predicted feeling is a measurement of the present, not a report from the future.

The one study that says which half came out better

Everything so far describes costs on both sides. Paul Amato and Bryndl Hohmann-Marriott asked the question people actually want answered: among people who divorced, who was happier afterwards?

Using two waves of the National Survey of Families and Households, they cluster-analysed 509 couples who divorced between the waves — and found the divorcing couples split into roughly two halves. About half had been in high-distress marriages. The rest had been in low-distress ones. The authors checked that this was not an artefact of when people happened to be interviewed or of measurement error, and it was not.

The outcome split along the same line. People in high-distress marriages reported increases in happiness following divorce. People in low-distress marriages reported declines. The authors read the two groups as having two different motivations: poor relationship quality in one case, and weak commitment to marriage in the other.

That is the most decision-relevant result in this article, and it is also the one that resists being made into advice. It says that the answer to would I be better off depends on a thing the person already knows and may be the worst-placed to judge — whether the relationship is genuinely distressed or merely disappointing. It is a population-level description of two groups, offered after the fact. Nobody has shown that it can be applied forward to one person.

Two further cautions. “Low distress” in a survey instrument is not the same as “fine”: it means the marriage did not register high on a distress scale, and quiet unhappiness is exactly the kind that scales miss. And both groups shared many of the same risk characteristics, such as having divorced parents, so they are not two different kinds of people.

What it does do is retire the two confident claims this genre alternates between — that leaving an unhappy relationship reliably makes people happier, and that most people who leave regret it. In the only dataset that split them, both were true of about half the sample each.

Staying for the children, and what was actually measured

This is the reason most likely to be stated as though it settles the matter, and it is one of the few with a direct longitudinal test.

Amato, Laura Spencer Loomis and Booth used twelve years of panel data to ask what happened to offspring whose parents' marriages differed in conflict and in whether they ended. The design is what makes it useful: earlier work could show that adults from divorced or high-conflict families reported lower well-being, but could not separate conflict from divorce.

The result depends entirely on the variable people tend not to assess.

In high-conflict families, offspring had higher well-being as young adults if their parents had divorced than if they had stayed together.

In low-conflict families, offspring had higher well-being if their parents had stayed together than if they had divorced.

And in marriages that did not end, parental conflict was negatively associated with offspring well-being — which is the finding that the phrase staying for the children generally skips. Staying is not a neutral act performed around the children. It is an environment they are in.

So the honest answer to “should we stay for the kids” is that the research has a conditional answer rather than a yes or a no, and the condition is the amount of conflict the children are living inside, not the amount the parents feel. Note what is being measured: conflict, not unhappiness. A quietly disappointed marriage and a loudly hostile one are different exposures, and this study distinguishes them. Conflict itself is also a smaller predictor of dissolution than the genre assumes — d = 0.16 under fixed effects in the 137-study count covered in what the research on couples who last reports — and how a couple conflicts is a separate measurement again, which is what the conflict style test scores.

The usual limits apply and one is specific. These are American families followed from the 1980s, and both divorce and single parenthood have changed in prevalence and in social meaning since. Reported well-being in early adulthood is also a narrow outcome. The structure of the finding — that it depends on conflict rather than on divorce as such — has held up better than any particular number from it.

Money, and why it is not the same constraint for everyone

Financial dependence appears on every list as a barrier. It has been measured, and the measurement shows it is not one barrier but a condition that interacts with how the person feels.

Liana Sayer, Paula England, Paul Allison and Nicole Kangas used longitudinal data and a model that distinguishes which spouse leaves — something most divorce research collapses — on American marriages. Two results.

When husbands were not employed, either spouse was more likely to leave. Men's unemployment destabilised marriages from both sides.

Wives' employment did something more conditional. When wives reported better-than-average marital satisfaction, their employment affected neither spouse's exit. When wives reported below-average satisfaction, their employment made it more likely that they would leave.

That is the clearest statement in this literature of what a material barrier actually does. A wife's job is not a cause of divorce and does not make happy marriages fragile. It is an enabling condition that converts dissatisfaction into an exit, and only for the people who were dissatisfied. Remove the job and the dissatisfaction does not disappear; the exit does.

Which reframes the “I cannot afford to leave” line. It is usually treated as a rationalisation to be examined. The data suggest it is frequently a correct description of the situation, and that the asymmetry it creates is structural rather than personal. The authors' own conclusion is that theories of divorce need “gendering” — that the predictors are not the same for husbands and wives, and a model that averages them describes neither.

The limits: these are mixed-sex American marriages, the employment measures are coarse, and the period covered predates much of the change in how couples earn. Whether the same asymmetry holds in cohabiting couples, same-sex couples, or in countries with different welfare provision has not, as far as we can find, been tested with a comparable design. It is also the single most obvious place where a population finding and an individual circumstance diverge, and where a page like this one is the wrong tool.

When staying is not a decision

If you are being hurt, controlled or threatened, the sections above are not the right frame, and nothing on this page should be read as a reason to stay. Find A Helpline, in the sources below, lists free and confidential services by country. In the United States the National Domestic Violence Hotline is 1-800-799-7233, or text START to 88788, 24 hours a day.

The reason this is a section and not a footnote is that the investment model was tested here, deliberately, and it worked — which is a disturbing result rather than a reassuring one.

Rusbult and Martz analysed intake interviews at a shelter for women who had experienced abuse. Commitment to the relationship was greater among women who had poorer economic alternatives, who were more heavily invested, and who reported less severe abuse. And commitment predicted behaviour: it distinguished the women who returned to their partners immediately on leaving the shelter from those who did not. The authors call this nonvoluntary dependence, and the phrase is exact. The same three variables that explain why anyone stays anywhere explain this, which means it requires no special psychology of the person and no deficit in her judgement. It is an answer to a question that gets asked in the wrong direction.

The constraint research says something related. In 1,278 unmarried adults, Rhoades, Stanley, Kelmer and Markman found 35.6% reporting physical aggression in their relationship in the past year. Those who had were more likely to have broken up a year later, and they reported lower dedication and higher constraint than those with no history of aggression — the combination described earlier as wanting it less and being held in place more.

Two more things belong here. Control is describable rather than intuitive: deciding who you may see, monitoring your time and requiring an account of it, controlling money or transport, going through your phone, pressure that continues after a no, and punishment by silence or by anger. Where the feeling has to be told apart from the behaviour, the psychology of jealousy works through it at length, with the evidence on why the period around leaving carries elevated risk — which is why safety planning with a service, rather than an article, is the step this page points to.

What this evidence cannot tell you

The limits go in the article rather than in a footnote, because the most common use of a page like this one is to get an answer it cannot give.

“Should I leave” is not answerable from this literature, and nothing here is an attempt. The closest anything comes is Amato and Hohmann-Marriott's split between high- and low-distress divorces, and that is a description of two groups after the fact, not a test anyone has shown works forward. Every study above reports what moved an average. None of them was designed to be applied to one relationship, and the research group that has tried hardest to predict relationship outcomes at scale found that change over time was predictable at about 5% of variance across 11,196 couples, a result set out in what actually builds attraction over time.

Stated reasons are unreliable, including the ones in these studies. Previti and Amato found that the reason everyone lists — no good alternatives — was rarely volunteered. Johnson and Rusbult found that the perception of alternatives is partly manufactured by commitment itself. Most of what is measured here is self-report, and self-report is a person's account of their situation rather than the situation.

The samples are narrow. Overwhelmingly American, overwhelmingly mixed-sex, heavily student and online-panel. The sunk cost work is hypothetical scenarios. The affective forecasting work is twenty-six undergraduates. The constraint work is eight months of unmarried couples. Each is the best available evidence for its specific claim and none is a general law.

Correlation is doing most of the work. Marital quality and health, staying and well-being, constraint and stability: these are associations in panel data. The designs narrow the alternative explanations without eliminating them.

And the one thing worth taking away is smaller than a decision. Staying is not a failure of nerve or a lapse of self-respect; it is what the measured variables predict, and most of them have nothing to do with how good the relationship is. Knowing which of them is holding a particular relationship up is a question with an answer, and it is the one our relationship satisfaction test is built to print rather than to judge. More of this under Mindset. Claims this site has had to retract are listed on the corrections page.