Why Smart People Keep Dating the Wrong People: What Is Measured, and What Is Only Repeated

Do people actually have a type? One study answers it properly

Start with the only part of this subject that rests on a clean measurement, because it is the part most worth knowing and it is almost never the part that gets cited.

Yoobin Park and Geoff MacDonald used the German Family Panel, a nationally representative longitudinal study that has followed the same people annually since 2008 and, crucially, recruits their partners too. That design allows something almost nothing else in this literature allows. Instead of asking people to describe their exes — which measures memory, mood and self-justification as much as it measures the ex — the researchers took 332 participants who had two different partners complete a personality questionnaire about themselves during the study period. Ex-partner and current partner, each describing their own personality, in their own words, years apart, with no idea they were being compared.

The question was whether an individual's current partner resembles their ex more than chance would predict. The answer was yes, and the analysis is careful about what kind of yes.

Two boring explanations had to be removed first. Normative similarity: everybody describes themselves in broadly similar, broadly flattering ways, so any two people look alike if you do not correct for it. That was large and real in the data (b = 0.53). Self-partner similarity: people tend to date people like themselves, so two partners of the same person will resemble each other simply through both resembling that person. That was there too (b = 0.20), which is assortative mating showing up as a number rather than as a saying.

With both of those subtracted, a third thing remained. Distinctive partner similarity: b = 0.22, p < 0.001. A person's current partner resembled their ex specifically, in ways that were neither generic nor a reflection of the person doing the choosing. A second and more conservative analytic approach found the same thing smaller but still significant.

So the folk claim is true, and unusually well established for this genre. People do have a type, it survives a change of partner, and it is not just an echo of themselves.

The same study contains the part that never gets quoted

Now the sentence that gets dropped every time this finding is repeated: nothing in it establishes that you are choosing.

Park and MacDonald say so themselves. They discuss an earlier study by Eastwick and colleagues that found a person's successive partners cluster on physical and some non-physical traits, and note that the clustering there appeared to be mostly attributable to partners sharing environments — a congregation, a campus, a workplace — rather than to anybody selecting anything. In their own discussion they raise the same possibility for their own result: partner consistency "may be tied to the diversity of personalities in one's environment and thus may not reflect active choice processes."

There is a hint in the data pointing the same way. People higher in extraversion or openness to experience had current partners less distinctively similar to their exes. That is consistent with a psychological story about seeking novelty, and equally consistent with a boring structural one: extraverted and open people have larger and more varied social networks, so a wider range of people passes in front of them. The authors flag this analysis as exploratory and note that it did not survive an alternative analytic approach.

This matters more than it sounds, because the entire self-help edifice built on "you have a type" assumes the type is a choice you are making, from a pattern inside you, and can therefore be corrected by insight. What has been measured is a pattern in outcomes. Whether it comes from your preferences, from who says yes to you, or from the fact that the same kinds of people keep showing up in the rooms you are in, is not settled. Probably it is some of each, and the mix likely differs by person.

A workable version: the repetition is real, the explanation is a hypothesis, and the honest place to start is not "what is wrong with my taste" but "who is actually in front of me, and why is it always these people". Some of that is geography, work and the shape of your week — nearby causes with real measured effects, laid out in what proximity and distance actually do to attraction.

Being clever does not protect you, and that has been measured too

The title of this page contains a claim, so it should be checked rather than asserted. Are intelligent people worse at this?

There is no study that ranks partner choices by IQ, and anybody who tells you otherwise is inventing. What there is, is a well-replicated finding about what cognitive ability does and does not do to your ability to see your own reasoning.

Richard West, Russell Meserve and Keith Stanovich ran two studies on the bias blind spot — the tendency to judge thinking biases as more common in other people than in yourself. Most earlier work had looked at social biases; they extended it to the classic cognitive biases from the heuristics-and-biases literature. They found the blind spot replicated across those biases. Then the result that matters here: none of the blind spots were attenuated by cognitive ability or by thinking dispositions related to bias. If anything, a larger bias blind spot went with higher cognitive ability. They also found that being free of the blind spot did not help a person avoid the actual biases.

Read carefully, that is not "smart people are worse at relationships". It is narrower and more useful: cognitive horsepower does not come with a matching improvement in seeing your own reasoning from outside, and on this particular measure it goes very slightly the other way. If your ability to construct a good argument is high, your ability to construct a good argument for the thing you were going to do anyway is also high. Nothing in the machinery flags which one is running.

That is the honest version of the intelligence trap, and it is a smaller claim than the one this page used to make. It also sets up the next problem, which is worse.

Why the reason you give is not the reason

In 1977 Richard Nisbett and Timothy Wilson published a review that has been unsettling people ever since. Their argument, assembled from a long run of experiments where the real cause of a behaviour was known to the experimenter and invisible to the participant, was that people have little direct access to the processes producing their judgements. When asked why, they do not read off a cause. They generate a plausible one, from implicit theories about what ought to have caused it, and report it with the confidence of an observation.

The consequence for anyone analysing their own romantic pattern is direct. The story you tell about why you were drawn to someone is a theory you are constructing, not a memory you are retrieving — and there is no internal signal distinguishing a good theory from a fluent one.

Wilson and Schooler then showed the practical cost. Across two studies, people who were asked to analyse the reasons for their preferences ended up with preferences that corresponded less with expert judgement — about strawberry jams in one study, about university courses in the other. Analysing reasons appears to focus attention on whatever is easiest to put into words, and choices then get made on those criteria.

The limits here need saying plainly, because this is exactly the point where popular psychology stops being careful. Those experiments were about jam and course catalogues, not partners. Nothing licenses reading them as "thinking about your relationship makes it worse". What they establish is a mechanism — verbalising reasons shifts which criteria drive a choice — and that mechanism is a warning against one specific move: treating a long, articulate account of your romantic history as though it were data about your romantic history. It is a product of the same system that produced the history.

Which raises an obvious empirical question. If people cannot report the causes of a jam preference, can they report what they want in a partner?

What people say they want does not predict who they want

Paul Eastwick and Eli Finkel tested exactly that, and the design is why the result stuck.

Before a speed-dating event, participants stated how much they valued physical attractiveness and earning prospects in an ideal partner. The familiar sex differences appeared, cleanly, as they always do when the target is hypothetical. Then everyone met actual people. In the live event, and in a longitudinal follow-up on partners met outside it, there were no sex differences in how attractiveness and earning prospects related to actual romantic interest — and participants' stated ideals, measured beforehand, failed to predict what actually inspired their desire. The authors close by pointing back at Nisbett and Wilson: even here, on something this consequential, people appear to lack introspective access to what is driving them.

That is a single event, so the meta-analysis matters. Eastwick, Luchies, Finkel and Hunt pooled 97 studies across the attraction and relationships literatures. Physical attractiveness predicted romantic evaluations at r ≈ .40 for both sexes; earning prospects at r ≈ .10 for both sexes; the sex difference in those correlations was r = .03 and uniformly non-significant. The gap is not between men and women. It is between hypothetical targets and people who are actually in the room.

And when researchers tried to predict the room itself, they could not. Samantha Joel, Eastwick and Finkel ran more than a hundred self-report measures on speed daters and applied machine learning: they could predict how much a person tended to desire others in general, and how much others tended to desire them, but the pair-specific component — whether these two would click — was not predictable from anything gathered in advance.

Put the three together and you get the sentence this page exists to deliver. Your criteria are not the problem, because your criteria were never running the choice. Refining a list is not a fix; it is the same operation performed more carefully. That is the same wall the compatibility-question genre runs into, worked through in detail in what questions can and cannot reveal about compatibility, and it is why milestone schemes like the 3-6-9 rule feel so satisfying and predict so little.

The intermittent-reinforcement story, and why this page no longer tells it

Until today this article contained the following, stated as fact: that anxiously attached people are drawn to avoidant partners because intermittent reinforcement — connection, distance, connection — is more neurologically activating than consistent warmth, that the reward system responds more strongly to variable schedules, and that this is therefore not a cognitive error but a feature of how reward circuits work.

We have removed it, because we could not source it. It is on the corrections page with the rest.

The claim is everywhere. It is on therapist blogs, in coaching newsletters and across short video, usually with a mention of dopamine and often with an implied citation to "the research". Following it back, what we found were other articles making the same claim. The underlying phenomenon — that behaviour reinforced on a variable schedule is more persistent, and that dopamine neurons respond to reward prediction error — is genuine work in learning and neuroscience. What is missing is the step that matters: a study showing that a variable schedule of a partner's warmth causes people to select or prefer that partner. We did not find one. Until somebody does, the mechanism is an analogy, and this site does not publish analogies as findings.

The nearest thing to real evidence points at attention rather than reward, and it is much thinner than its reputation. Erin Whitchurch, Timothy Wilson and Daniel Gilbert had college women view the profiles of four male students who had supposedly already seen theirs. Told the men liked them a lot, they liked the men more than when told they were liked an average amount, which is the ordinary reciprocity effect. But the group told the men might like them a lot or an average amount — the uncertain condition — was the most attracted of all, more than the certainly-liked group. The mediator the authors identified was rumination: uncertain participants thought about the men more, and that predicted the attraction.

Its limits, stated at the size they actually are: one small laboratory study, one sex, undergraduates, profiles rather than people, and no partner selection over time. It supports one narrow thing — uncertainty holds attention, and attention feels like attraction from the inside — and nothing about neurology, and nothing about who you end up with. The same wedge between intensity and information is what separates love from anxious attachment, and the same reason-to-check loop drives the psychology of jealousy, where the reassurance always expires. It is also why engineered unavailability — the delayed reply, the strategic silence — keeps failing when it is tested, which is the subject of the 72-hour rule.

Self-esteem does something real, and the popular version has the stage wrong

"You accept the love you think you deserve" is the most quoted line in this genre. There is a measured finding underneath it, and it is more specific and more interesting than the slogan.

William Swann, Chris De La Ronde and J. Gregory Hixon surveyed 176 married and dating couples, testing a prediction from self-verification theory: that people want intimates to see them as they see themselves, even when that self-view is negative. The result split by relationship stage, and the split is the whole point.

Dating participants were most intimate with partners who evaluated them favourably. Not accurately — favourably. At that stage, being seen well was what tracked closeness, regardless of self-view.

Married participants were most intimate with spouses whose evaluations verified their self-views. And for married people with negative self-views, that ran where you would expect and would rather it did not: intimacy increased as their spouses evaluated them more negatively.

So the honest statement is not "you pick partners who confirm what you think of yourself". At the choosing stage, the data say close to the opposite: people lean toward the person who thinks well of them. What appears to happen is a shift — from wanting to be seen positively to wanting to be seen correctly — as a relationship becomes the place where your identity is confirmed rather than displayed. If your self-view is poor, that shift converts a warm partner into a slightly wrong one over time, without a single decision being made.

The usual caveats: one cross-sectional survey from 1994, self-report, one cultural context, and marital status standing in for a process that surely varies. It is a well-known result rather than a settled law.

There is also a measured version of the "what I deserve" intuition that does not depend on any of this. In Rusbult's investment model, one of the three things holding people in relationships is the perceived quality of their alternatives, and a meta-analysis of 52 studies and 11,582 participants put its association with commitment at r = −.48. A belief about what you deserve is, operationally, a belief about your alternatives — which is why our relationship satisfaction test scores alternatives as a number of its own rather than folding it into how happy you say you are.

The repetition is not only about choosing. It is also about not leaving

Framing this as a selection problem hides half of it. Two people with identical taste can end up with very different histories depending on how long each stays after it stops working, and staying has been studied more directly than choosing has.

The investment model says commitment rests on three things: satisfaction, the quality of your alternatives, and what you have invested. Le and Agnew's meta-analysis — 52 studies, 60 independent samples, 11,582 participants — found all three associated with commitment (satisfaction r = .68, alternatives r = −.48, investment r = .46), together accounting for close to two-thirds of the variance in commitment, and commitment in turn predicting breakup. Note what that means for the pattern: a person can be unhappy and still highly committed if their alternatives look poor and their investment is large. No bad taste is required to produce a run of long, unsatisfying relationships. Time is enough.

Then a finding that is easy to miss and hard to unsee. Joel, Impett, Spielmann and MacDonald tracked 1,348 people in relationships over ten weeks, then ran a preregistered replication with 500 people actively contemplating a breakup over two months. In both, the more dependent people believed their partner was on the relationship, the less likely they were to initiate a breakup — and this held above and beyond the self-focused investment-model variables. People stay, in part, for the other person.

That reframes the question this page is named after. "Why do I keep choosing the wrong people" assumes the mechanism is at the front door. Some of it is at the exit, and the exit is where the more tractable questions live: how long you take to notice, what you count as sufficient reason, and whose feelings the decision is being made around. If the arrangement never quite becomes a relationship, that has its own shape and its own literature — why waiting tends to add constraint rather than commitment — and if the difficulty is that it ended and did not finish ending, the recovery markers are narrower and more useful than the folk ones, in the signs you have moved on and the signs you have not.

When the wrong person is not a type

This subject has an edge that the "type" framing hides, and it is worth naming before the section on change.

Some patterns of repeated bad relationships are patterns of taste, environment and timing. Some are patterns of harm. Those are not the same problem, they do not respond to the same reading, and an article about partner preference is the wrong instrument for the second one.

The behaviours that mark the second category are describable rather than intuitive. The United Kingdom Home Office's statutory guidance on controlling or coercive behaviour lists, among others: making and enforcing rules with punishments for non-compliance; monitoring time and requiring an account of it; isolating a person from friends, family and support; controlling access to money, transport or work; repeated belittling that makes someone feel worthless; and threats. None of that is a compatibility issue and none of it is fixed by better partner selection. If reading a page about why you keep choosing badly feels like it is describing something that has been done to you, that is the more urgent subject, and the resources at the foot of this page are the better starting point.

There is a second, quieter version, and it is the one this genre gets wrong most often. Some people arrive here having concluded that repeatedly being treated badly is evidence of a defect in themselves. Nothing in the research supports that inference. The measured findings on partner consistency describe a statistical regularity across a population, with a mechanism nobody has established, and possibly not a choice at all. A pattern is not a verdict, and being harmed is not a preference.

Does the pattern change? Three answers, at the sizes they come in

The popular literature is confident about this in both directions — attachment style as destiny, and attachment style as something you fix in six weeks of journalling. The evidence is less dramatic than either.

On its own, slowly, in the direction you would hope. William Chopik, Robin Edelstein and Kevin Grimm pooled three long-running longitudinal studies to track attachment from age 13 to 72 (N = 628). Attachment anxiety declined on average with age, particularly through middle and older adulthood; avoidance declined in a linear fashion across the lifespan. Being in a relationship predicted lower anxiety and lower avoidance across adulthood. Nobody in that dataset was doing anything about it.

Deliberately, yes, and the effect is small. Nathan Hudson, Chopik and Daniel Briley built a measure of people's goals to change their own attachment anxiety and avoidance, then ran a sixteen-wave weekly study across a university semester. Change goals predicted corresponding change: wanting to be less anxious predicted declining anxiety, wanting to be less avoidant predicted declining avoidance, with no cross-contamination between the two, and the effects survived controls for neuroticism and for goals to change neuroticism. The coefficients were small — comparable, as the authors say, to what is found for the Big Five. This is real, and it is a slow drift over months, not a switch.

In therapy, probably, on weak-to-moderate evidence. Paul Taylor, Jacqueline Rietzschel, Adam Danquah and Katherine Berry systematically reviewed studies of attachment representations before and after psychological therapy, across interview and self-report methods. Security increased and anxiety decreased following therapy; findings for avoidance were unclear. The improvements appeared across different patient groups, therapeutic approaches and settings, and across levels of study quality. Their own conclusion is the appropriate one to quote: further controlled trials are required.

And one uncomfortable correction to the story people tell about having overcome their childhood. "Earned security" — coherent, secure adult functioning despite a difficult early history — is normally identified retrospectively, from how an adult narrates their childhood. Glenn Roisman, Elena Padrón, Alan Sroufe and Byron Egeland checked that against 23 years of prospective data on the same people. Retrospectively defined earned-secures had not been more likely to be anxiously attached as infants, and had been observed in childhood and adolescence to receive among the most supportive and structured parenting in a high-risk sample. The narrative of having come from a bad place and made it to a good one, told by adults with a documented past, did not match the documented past. What does hold up: people who were actually insecure as infants and secure as adults did go on to do well in close relationships.

Which is a good result and a warning at once. Change happens. The autobiography you have built about how it happened is a separate object, and the two do not have to agree — the same gap between an account and the thing accounted for that runs through this whole page. The full version of this question, with the stability coefficients, the prototype model and what couple therapy has actually measured, is in can attachment styles change.

What is worth watching instead

If ideals do not predict, if pair-specific desire is unpredictable, and if the reasons you can articulate are constructions, what is left to pay attention to?

The largest attempt to answer this pooled 43 longitudinal couples studies and 11,196 couples through machine learning. The strongest predictors of relationship quality were a person's own experiences inside the relationship — perceived partner commitment, appreciation, sexual satisfaction, perceived partner satisfaction, conflict — accounting for up to 45% of the variance at baseline. Individual differences measured outside the relationship added far less. A partner's own ratings added almost nothing on top of what the person themselves reported. And change over time was barely predictable at all, at around 5% of variance.

There is no partner-selection insight in that. Which is the finding. Nothing measured before a relationship exists tells you much; almost everything that predicts is generated inside one, and only becomes available once you are in it.

That points at a practical shift, and it is smaller and duller than the shift this page used to recommend. Stop auditing the type. Start noticing, early and specifically, what happens in you around this particular person: whether the reassurance holds past the afternoon, whether you are collecting evidence instead of asking, whether affection is increasing or draining. Those are observations rather than theories, they are available within weeks, and they are the same variables the attachment style test and the relationship satisfaction test score — both of which say plainly on the page what a result cannot tell you.

The behaviour-level version of the same shift is what has been measured about emotional maturity, and what only gets repeated; the longitudinal version, in which personality measured outside a relationship predicted breakups at close to zero, is what the 137-study count of dissolutions found; and the version measured in the hour after a fight, where analysing your partner is precisely the thing that makes it worse, is what the repair research supports. More of this under Mindset.