What Actually Predicts a Happy Relationship

Search for relationship advice and you get lists. Twenty-two tips. Nine things she wishes you knew. Ten habits of couples who last. None of them tell you where the advice came from, and almost none of it has been tested against anything.

The strange part is that this question has been studied properly, by a lot of people, for decades. In 2020 a team pooled the results and asked the obvious question nobody had been able to answer: out of the hundreds of things relationship science claims matter, which ones actually predict whether people are happy?

The answer is not what the lists say, and the most useful part of it is the bit that almost never gets quoted.

The short answer

Five things predicted relationship satisfaction more than anything else, and all five are perceptions rather than facts about your partner.

What did not predict it: either person’s personality, either person’s traits, and the partner’s own ratings of anything.

The study this rests on

This is one of the largest efforts ever made (official source)Official source: sites.lsa.umich.eduA university source, not a blog or a shop. Read the original and judge it yourself. to settle a question in relationship science, and it was built by pooling other people’s data rather than running one more study on one more sample.

Two things about that are worth slowing down for, because they cut against most of what gets written on this subject.

It is the relationship, not the person

The lists all assume the problem is selection. Find the right person, avoid the wrong ones, learn to read the signs. Whole industries run on that assumption.

The data points somewhere else. Once you account for how somebody sees their relationship, knowing anything about their partner as a person adds nothing. Not their personality. Not their traits. Not even the partner’s own account of how the relationship is going.

Katerina Holmes / Pexels

That is a genuinely strange result if you take dating advice seriously, and a hopeful one if you are in a relationship already. The thing that predicts satisfaction is the part you both have some purchase on, rather than a fixed property of the person you happened to meet.

It also reframes the word compatibility. Used properly it should describe what happens between two people, which changes. Used the way dating advice uses it, it describes a match between two lists of traits, and that version had no predictive power here.

Why the advice you find is shaped the way it is

Before the findings, it is worth understanding why almost nothing you read on this subject looks like them.

Relationship advice is one of the highest-volume categories on the internet and one of the least accountable. The claims are unfalsifiable in the short term, the reader has no way to check them, and nobody comes back a year later to report that tip fourteen did not work. That combination produces a specific kind of writing: confident, unsourced, and organised as a list because lists are easy to produce and easy to skim.

Three patterns are worth being able to spot, because once you can, most of the field becomes readable.

The unattributed study. “Researchers found that couples who do X are more likely to stay together.” No name, no year, no sample size. Sometimes there is a real paper underneath, often the same handful of papers, and the details that would let you judge it have been removed. Two of the most repeated claims in this entire category trace back to studies of fewer than a hundred couples.

The reversed arrow. Happy couples do a thing, therefore doing the thing makes couples happy. Almost every correlation in this field gets reported this way, and almost none of the underlying research can support it. Couples who go on date nights are happier; it does not follow that scheduling a date night produces happiness rather than happiness producing date nights.

The universal claim from a narrow sample. A finding from married American couples in one decade, restated as a fact about men and women. The finding may be sound and the generalisation is a separate claim that usually goes untested.

None of this means the field is worthless. It means the useful move is to look at the largest, best-powered work and to read the limits alongside the results, which is what the rest of this guide does.

The five, one at a time

Ranked lists invite skimming, so it is worth saying what each of these actually is.

Perceived partner commitment. The strongest single predictor, and note the first word. What mattered was whether you believe your partner is in this, not what your partner reports. Two people can be equally committed and only one of them is being told, which is a solvable problem and a surprisingly common one.

Appreciation. How appreciative you feel toward your partner. Again the direction is worth noticing: it is your gratitude toward them that predicted your satisfaction, not their gratitude toward you. Noticing what somebody does for you appears to do something to the person doing the noticing.

Sexual satisfaction. Third, above conflict. Worth stating plainly because advice written for men tends to either lead with this or pretend it does not signify. It is neither the whole thing nor a side issue.

Perceived partner satisfaction. Whether you think they are happy. Related to the first, and it has the same property: it is a belief about them that lives in you, and beliefs can be out of date.

Conflict. Present, and fifth. Not first. How much you fight matters less than four things that are mostly about perception, which is not what most people would guess.

Perception is the mechanism, and perceptions go stale

Four of the five predictors are beliefs about your partner rather than facts about them. That is the most actionable thing in the whole study and it is easy to read past, so it is worth making concrete.

A belief about somebody you live with is not continuously updated. It gets formed at some point and then coasts, and it coasts on evidence that may be years old. You can be entirely committed and your partner’s estimate of your commitment can be running on a version of you from three years ago, because nothing has happened since to revise it.

This is why the strongest predictor in the study is also, in a sense, a communication problem. The variable is not commitment. It is perceived commitment. The gap between those two words is where a lot of otherwise sound relationships quietly sit.

Two consequences follow, and both cut against how men are usually advised to behave here.

Demonstrating is not the same as saying. The standard advice is that actions speak louder than words. As a description of what other people can observe, that is true. As a strategy for updating somebody’s beliefs, it is unreliable, because actions are ambiguous and get read through whatever the person already thinks. Paying the bills demonstrates commitment to somebody who already believes you are committed, and demonstrates obligation to somebody who does not.

Assuming it is obvious is the failure mode. The most common version of this is a man who is genuinely, unambiguously committed, has never said so in plain language since the early part of the relationship, and is confident it goes without saying. The study’s top predictor is precisely the thing he is leaving to inference.

The fix is not romantic and it is not a technique. It is saying the thing out loud, in words, at intervals, in a way that can be understood without interpretation.

The finding almost nobody quotes

The headline results above got written up widely. This one did not, and it is the most important sentence in the paper.

None of it predicted the future.

The team tested those hundreds of variables against how relationship quality changed over the course of each study, and nothing reliably predicted whose relationship would improve and whose would decline. Not the perceptions, not the personalities, not the partner’s reports.

Leticia Vieira / Pexels

Hold that next to the amount of content promising to tell you whether your relationship will last, whether they are the one, and which green flags to watch for. The largest study of its kind ran the numbers and could not do it.

That is not a counsel of despair. It is a reason to stop treating the question as a prediction problem to be solved with better screening, and to treat it as a maintenance problem instead. The measurable stuff is about now. Now is also the only part you can act on.

What predicts it ending

There is a second body of work (official source)Official source: bpl.studentorg.berkeley.eduA university source, not a blog or a shop. Read the original and judge it yourself. approaching from the other side, and it is the one that filters into popular advice, usually without attribution.

Read the sample size before you read the conclusion. Seventy-three couples, married, in the United States, in the 1980s. That is a small and specific group, and this study is often cited as though it were a law of nature. It is not. It is a careful description of a pattern in one cohort.

What makes it worth including anyway is that the pattern is behavioural rather than dispositional. Withdrawal, defensiveness and the ratio of negative to positive expression are things people do, which means they are things that can change. That is the same shape as the finding above.

What the conflict findings actually describe

Conflict came fifth in the 2020 study, and the 1992 work is entirely about how conflict goes. Between them they point at something more specific than “communicate better”, so it is worth unpacking what the coded behaviours look like from the inside.

Withdrawal from interaction. Coded from the outside it looks like one person going quiet and disengaging while the other keeps talking. From the inside it rarely feels like a tactic. It usually feels like being overwhelmed and running out of anything useful to say. That distinction matters, because the partner still talking almost always reads it as punishment or indifference, and the person withdrawing almost always experiences it as self-protection.

Defensiveness. In practice this is any response whose function is to establish that the complaint is unfair rather than to address it. Explaining. Correcting a detail. Pointing out that they do it too. Each individual instance is reasonable and the pattern is corrosive, because a complaint that never lands has to be repeated, and repeated complaints escalate.

The ratio, not the incident. The 1992 finding is about the balance of negative to positive expression across an interaction rather than about any single exchange. That is a more forgiving standard than it sounds. It does not require that arguments be pleasant. It requires that the relationship not be overwhelmingly negative in aggregate.

What none of the research supports is the popular idea that healthy couples do not fight. The couples whose relationships stayed regulated in the 1992 study were not conflict-free. They were differently shaped in conflict.

The other half: what you bring

The individual-difference predictors explained less than the relationship perceptions, but they were not nothing: life satisfaction, negative affect, depression, attachment avoidance and attachment anxiety.

The honest reading of that is uncomfortable and useful. How your life is going in general, and how you are doing, shows up inside the relationship. It is not a separate department.

Alex Green / Pexels

Two practical consequences:

• Work on your own life is relationship work. Not a substitute for it, and not separate from it. If your life outside is bad, the data says that turns up inside.
• Attachment style predicted something, and it is not a personality test result. Avoidance and anxiety describe how you behave when closeness or distance increases, and both are patterns rather than fixed types.

Attachment, without the personality-test version

Two of the five strongest individual predictors were attachment avoidance and attachment anxiety. The word attachment has been through the internet and come out the other side as a four-box personality quiz, so it is worth recovering what it means in the research.

It describes what you do when closeness or distance increases. Not who you are.

Avoidance is the tendency to create distance when things get close. It shows up as needing space during conflict rather than after it, discomfort with dependence in either direction, and a preference for handling difficulty alone. From the outside it reads as not caring. From the inside it is usually the opposite: it is caring plus a strong instinct that leaning on somebody is unsafe.

Anxiety is the tendency to seek reassurance when distance increases. It shows up as reading tone for signals, needing confirmation that things are fine, and difficulty settling until contact is restored. From the outside it reads as neediness. From the inside it is an alarm that will not switch off until the connection is verified.

Three things the research supports that the quiz version obscures:

• These are dimensions, not types. You have some amount of each, not one of four labels, and the amount is not identical in every relationship.
• They move. Attachment patterns shift over time and with experience. Treating yours as a permanent diagnosis is not what the literature says.
• The combination matters more than either. The classic difficult pattern is one partner pursuing closeness while the other creates distance, with each response confirming the other’s expectation.

What this is useful for is not labelling yourself. It is recognising a pattern in the middle of it, which is the only point at which it can go differently.

The two predictors that are really one problem

Perceived-partner commitment came first and perceived-partner satisfaction came fourth. They are separate variables and in practice they fail together, because both are answers to a question nobody asks out loud: how is this actually going for you?

The reason that question goes unasked is worth being honest about. Asking it invites an answer you might not want, and the longer the relationship has run without asking, the higher the stakes of the answer feel. So the question gets deferred, and the estimate keeps running on old data, and the gap between what you believe and what is true grows in the dark.

What makes this specifically a problem for men is not emotional incapacity. It is that the available scripts are bad. “Are you happy?” is too large to answer honestly and usually gets “yes, why?”. “Is everything okay?” is a question about crisis, so the true answer to it is almost always yes even when things are quietly not fine.

Better shapes, because the shape is the whole difficulty:

• Ask about a specific stretch of time. “How has the last couple of months felt to you?” is answerable in a way that “are you happy” is not.
• Ask what is missing rather than what is wrong. Wrong invites defence of the status quo. Missing invites description.
• Ask when nothing is happening. The same question during an argument is a move in the argument.
• Say your own first. It sets the register, and it makes clear you are asking for information rather than reassurance.

None of this is a technique for extracting a favourable answer. The point is that two of the five strongest predictors of your own satisfaction are estimates you are carrying about somebody else, and estimates that are never checked are estimates that drift.

The conversations most couples never have

The predictors point at three subjects, and all three tend to go undiscussed for the same reason: the conversation feels like it implies a problem.

Money. Not budgeting. Expectations: what money is for, what counts as security, what a reasonable risk is. Two people can share an account for years without either knowing what the other believes about any of that, and the discovery usually happens during a decision that is already contentious.

Sex. Third-strongest predictor, and the subject with the least available language. The obstacle is rarely disagreement; it is that the first conversation has to happen with no established way of talking about it, and both people are braced for the other to hear criticism.

What each of you is actually for. Whether you are building something together or living adjacent lives that overlap. Long relationships can proceed a long way without either person saying which of those they think is happening.

The practical version is unglamorous. These are conversations to have when nothing is wrong, because the alternative is having them during the argument that made them unavoidable, and that version goes badly for reasons that have nothing to do with the subject.

What this means if you are not in one yet

Everything measured here was measured inside existing relationships, which raises an obvious question for anybody currently looking: what does it tell you about who to look for?

Less than you would hope, and that is the useful answer.

The variables that predicted satisfaction are properties of a relationship that does not exist yet. You cannot assess perceived-partner commitment on a second date because there is nothing to perceive. And the variables that do exist before a relationship starts, personality and traits, were the ones that added nothing.

Three things follow that are worth more than a checklist.

Screening has a low ceiling. Effort spent evaluating whether somebody meets a specification is effort spent on the category of information the data was least impressed by. That does not mean anything goes; it means the returns on getting more selective are smaller than dating advice implies.

How it goes between you is observable early. Not their traits, the dynamic. Whether conversation is work. Whether disagreement is survivable. Whether you like who you are in their company. Those are early-visible versions of what actually predicted satisfaction later.

Your own state comes with you. Life satisfaction, mood and depression were among the strongest individual predictors, and they do not wait outside until you find somebody. This is the one piece of the standard advice the data supports: sorting out your own life is not a delaying tactic.

The most honest thing that can be said to somebody looking is that the search matters less than the maintenance, and almost all the available advice is about the search.

A bad patch or a bad fit

Since nothing in the data predicts a trajectory, the question of whether to stay cannot be answered by a checklist. It can be made less confusing.

Two distinctions are worth holding.

Is this a state or a pattern? A state is bad and identifiable: a bereavement, a job loss, a new baby, an illness. Relationships get worse during those and it does not tell you much about the relationship. A pattern is the same shape recurring across states, and that is more informative than any individual bad month.

Is the thing you want available? Not whether they are a good person, and not whether you love them. Whether the specific thing you need is something they can give, and whether the version of the relationship you want is one they want. Those are answerable questions and they are different from the question people usually torture themselves with, which is whether they are the right person.

What the evidence does say is that the direction of travel is not visible in any measured variable, including to the people inside it. So the honest framing is not “will this work”. It is “is this good enough now, and is the thing making it not good enough something either of us can change”.

So what do you actually do

Nothing here is a technique. The predictors are broad, which means the actions are simple and boring, which is why they do not make good lists.

Make commitment visible rather than assumed. The single strongest predictor was a belief that lives in the other person. Being committed and never saying so leaves the strongest lever in the relationship untouched.

Say the appreciation out loud. Feeling appreciative predicted your own satisfaction. Saying it is what gives your partner the same thing on their side of the ledger.

Treat sex as a subject you can discuss. It was the third strongest predictor and it is the one couples are least likely to have language for.

Change how conflict goes rather than how often. Conflict was on the list but it came fifth, and the 1992 work points at the shape of the fight rather than the count. Withdrawal and defensiveness were the markers, not disagreement.

Look after your own life. Because it shows up in the measurement whether you intend it to or not.

Repair, which is the part nobody teaches

The 1992 findings describe what goes wrong during conflict. What they leave implicit is that the couples who stayed regulated were not avoiding those behaviours perfectly. Everybody gets defensive. Everybody withdraws sometimes. The difference is what happens next.

Repair is the unglamorous mechanism underneath, and it has three parts.

Naming it while it is happening. “I am getting defensive” said out loud, mid-argument, does more than any amount of not being defensive. It converts a behaviour the other person is experiencing as an attack into information you are both looking at.

Coming back. For anybody whose pattern is to withdraw, the single highest-value habit is returning to the conversation without being fetched. Not immediately, and not never. The withdrawal itself is survivable; the open question of whether the person is coming back is what the other person is actually reacting to.

Being specific about what you got wrong. A general apology ends an argument. A specific one updates the other person’s belief about whether you understood. Given that four of the five predictors are beliefs, that difference is not cosmetic.

None of this requires being good at conflict. It requires being willing to re-enter it after handling it badly, which is a lower bar and a more useful one.

Your reading of it moves more than you think

Everything above is measured through what people report, and there is a German study from last year on how stable those reports actually are. It matters for anyone trying to answer “is this working” by consulting their own impression.

Two things follow. A bad week is not a measurement, and a good week is not one either. And the number you would give if asked today is a sample from a distribution rather than a fact about the relationship, which is why the useful question is about the trend across months.

The pattern that does the damage

The Gottman behaviours above describe couples heading for dissolution. A separate and much larger body of work isolates one specific pattern, and it is the one worth learning to recognise because both people usually think the other is causing it.

The reason it is worth naming rather than just avoiding is that it is self-feeding. Pressing produces retreat, retreat produces more pressing, and each person can point at the other’s behaviour as the cause with complete sincerity. How to stop fighting covers what to do with it; here the point is narrower. If you are trying to work out whether a relationship is in trouble, this pattern is more informative than how often you argue.

You are not a neutral judge of this

The last thing worth adding is a correction factor, because everything in this guide is something you will apply to your own relationship, and there is research on which direction people lean when they do that.

There is a related measurement problem, set out in a theoretical review published this year (official source)Official source: pmc.ncbi.nlm.nih.govThe study itself, on the US National Library of Medicine database. Read it and check the claim yourself.: positive and negative feelings are independent dimensions, not two ends of one scale. Someone who feels strongly both ways lands in the middle of a satisfaction scale, and so does someone who feels very little either way. Those two states are completely different and a single score cannot separate them. If you want a usable reading, score the good and the bad separately rather than netting them.

And on the cost of getting the answer wrong in the direction the bias favours: a twelve-year study of 1,150 people (official source)Official source: psu.eduA university source, not a blog or a shop. Read the original and judge it yourself. found that those who stayed in long-term low-quality marriages had lower self-esteem, health, happiness and life satisfaction, with no evidence of being better off in any respect than those who divorced. That is a study of marriages rather than of relationships in your twenties, and the comparison is not a controlled one. What survives the caveats is that staying is not the safe default it feels like. How to know when to end a relationship works through that properly.

What changes over ten years

Most of what gets written about relationships describes the first eighteen months, because that is the part with the most visible events. The measured predictors do not obviously change with time, but what makes them hard does.

Commitment stops being demonstrated by default. Early on, the relationship generates constant evidence: choosing to see each other, rearranging things, the visible fact of selection. Later that evidence stops being produced as a byproduct, and the strongest predictor in the study loses its automatic supply.

Appreciation gets harder to feel, not less deserved. What somebody does regularly becomes invisible. That is a property of attention rather than a failure of character, and it means the second-strongest predictor decays unless something interrupts it.

Conflict acquires a history. An argument in year one is about the thing. An argument in year eight is about the thing plus every previous instance of it, which is why long-term couples can escalate faster over smaller matters.

The perception gap widens. Beliefs formed early get less and less updating, for the reasons above.

None of that is decline. It is the same five variables under different conditions, and it explains why relationships need active maintenance in a way that early relationships do not. The early version supplies the inputs for free. Later, somebody has to put them in deliberately.

What this cannot tell you

Three limits, stated because a guide that reports only the confident parts of a literature is not reporting the literature.

These are correlations. Feeling appreciative and being satisfied go together. The study cannot tell you that manufacturing gratitude causes satisfaction, and anybody who tells you it can has gone past the data.

The samples skew. Pooled longitudinal couples research is disproportionately Western, educated and heterosexual. The findings may travel; the study cannot show that they do.

Self-report has a known problem. Everything here except the 1992 behavioural coding rests on people describing their own relationships, and people are not neutral observers of their own relationships.

The advice this evidence does not support

Being clear about what the data does not say is as useful as reporting what it does, particularly in a field where confident claims are the norm.

“Never go to bed angry.” No support. The 1992 behavioural work points at withdrawal and defensiveness within an interaction, not at bedtimes. For somebody whose pattern is to flood and shut down, forcing resolution before sleep is likely to produce the exact behaviour the research flags.

“Relationships are 50/50.” Nothing in the predictors resembles an accounting model, and the strongest ones are perceptions rather than contributions. A relationship where both people are tracking their share is one where perceived appreciation is likely to be doing badly.

“If they wanted to, they would.” Circular, unfalsifiable, and it converts an observable behaviour into a verdict about someone’s inner state. The measured variable is your perception of their commitment, and the honest way to update a perception is to ask.

“Opposites attract” and its mirror, “find someone just like you.” Partner traits added nothing to the prediction of satisfaction. Both slogans are claims about matching, and matching is the thing the data was least impressed by.

“You complete each other.” The individual predictors say the opposite in a mild way: your own life satisfaction and mental health predicted relationship quality, which means arriving depleted is a burden the relationship carries rather than a hole it fills.

“Happy couples do not fight.” Conflict came fifth, and the regulated couples in the 1992 study were not conflict-free.

Reading the next thing you find

You will read more about this subject than this guide, most of it worse-sourced. Four questions strip out the majority of it quickly.

Who was studied, and how many? A finding from 73 couples and a finding from 11,196 are not the same kind of object, and both get reported as “research shows”.

Is it a correlation being sold as a cause? Look for the word “leads to” or “makes”. If the study measured two things at the same time, it cannot support either verb.

Is the claim about prediction? Anything promising to tell you whether a relationship will last is claiming something the largest study in the field could not do.

Does the person telling you have something to sell? Course, book, app, coaching. Not disqualifying, and worth knowing before you weigh the confidence of the claim.

Applied honestly, those four questions will make most of this category unreadable. That is the correct result.

The mistakes worth avoiding

• Treating relationships as a selection problem. The evidence says the thing that predicts satisfaction is the state of the relationship, and the thing that predicts its direction is nothing anybody has measured.
• Assuming your partner knows. Perceived commitment was the top predictor and it lives in their head, not in your intentions.
• Counting fights. Conflict came fifth, and the behavioural work points at how conflict goes rather than how often.
• Reading a checklist as a forecast. Green flags, red flags, compatibility scores. The largest study on this could not forecast the trajectory, and it had 11,196 couples to work with.
• Splitting your life from the relationship. Life satisfaction, mood and depression were among the strongest individual predictors.
• Trusting advice with no source. Almost every claim in this field circulates without one, including several that trace back to the 73 couples above.

Where to go from here

This guide is the map. Each of the five predictors has enough underneath it to be its own subject, and the conflict findings in particular deserve more room than a section.

If you take one thing away and act on it this week, make it the first predictor: say the commitment out loud, in words, to somebody who may be running on an old estimate. It is the strongest variable in the largest study on the question, it costs nothing, and it is the one most likely to be sitting untouched.

If you are earlier than that, and the harder problem is that your life does not currently contain many people, the individual predictors point at the same place: our guide on making friends as an adult covers the mechanism, which turns out to be measurable in hours.

What we cannot tell you from experience

One honesty note about this guide specifically, because the site’s rules require it.

This is a research-backed guide, not a field-tested one. Nobody at Gritwright ran a study on couples, and the badge on this article says exactly that. What is here is the best-powered work that exists, read in full, reported with its sample sizes and its limits attached.

Where a claim in this guide could have been stated more confidently, it has not been. The 1992 study has 73 couples in it and this guide says so every time it is used. The 2020 study cannot establish causation and this guide says so. That is deliberate, and it is the difference between reporting a literature and mining it for lines that sound authoritative.

The bottom line

The largest study ever run on this question found that what predicts your happiness in a relationship is how you see the relationship: whether you think they are in it, whether you feel grateful, whether sex works, whether you think they are happy, and how conflict goes.

It also found that nothing at all predicted which relationships would get better and which would get worse.

Which is a more useful place to start than any list of twenty-two tips, because it points every bit of available effort at the present rather than at a forecast nobody can make.