Personality Similarity and Romantic Compatibility

Romantic partners and best friends were more similar than randomly paired participants in historical SimilarMinds data. Similarity was a recurring feature of both romantic pairings and friendships, while its association with relationship satisfaction was positive but modest.

SimilarMinds research · Analysis updated October 19, 2026

Do people in close relationships tend to be alike? In the historical Facebook sample, mutual romantic partners, mutual best friends and the broader friendship group all had higher average personality similarity than the random-participant baseline. Resemblance was not confined to romantic relationships.

A separate, older SimilarMinds compatibility dataset addresses whether more similar partners also reported happier relationships. Smaller differences between their profiles were associated with higher average relationship satisfaction, although that association was modest.

These are two different questions: who resembles whom, and whether resemblance predicts relationship quality. Greater average similarity among romantic partners and friends supports a shared pairing pattern; it does not establish that similarity caused those relationships or made them successful.

What the Data Contains

The workbook contains 77,259 relationship records, two sets of 21 profile scores, and ratings in both directions. Fields qq1 through qq7 contain partner A’s relationship ratings; the corresponding b fields contain partner B’s ratings.

For the main analysis, we used 23,458 records with usable profiles, positive relationship-duration responses from both partners, and valid satisfaction ratings from both partners. Each relationship’s mutual satisfaction score is the average of those two ratings, from 1 to 10. A secondary analysis uses the lower rating, so that one partner’s unhappiness is not hidden by the other’s high rating.

These are records, not a claim of 23,458 independent, verified couples. Repeated identifiers were kept together when estimating uncertainty and testing predictions.

Similarity Was Associated With Satisfaction

We recomputed a transparent similarity index from 20 dimensions in Version 1.02: smaller average profile differences produced a higher score. The orientation-related dimension was excluded from this general resemblance index. With the wording now available, orderliness, work orientation, and self-rated competence were included rather than discarded because of conflicting spreadsheet labels.

The relationship between similarity and mutual satisfaction was r = 0.173, with an approximate cluster-adjusted 95% confidence interval of 0.158 to 0.188. This is a positive but modest association, not a near-deterministic relationship.

A simple linear model using only similarity explained approximately 3.0% of the variation in mutual satisfaction in held-out groups from this dataset. Most of the variation remained unexplained.

The average rose through the 80-to-below-90 band, but the 90-to-100 band had a lower average. That highest band was much smaller. The data does not justify a rule that every increase in resemblance improves a relationship, or that a particular similarity threshold guarantees success.

Common Similarities Were Not Always the Most Useful Ones

There are two different questions: Do partners tend to resemble each other on this dimension? and Are smaller differences on this dimension associated with higher satisfaction? They lead to different rankings.

Labels describe the supplied item wording; they are not claims that these short measures are validated scales. “Agreement” means a smaller absolute difference between the two profile scores. The second column describes resemblance between partners across records; the third relates each pair’s agreement to its satisfaction rating. These are unadjusted, exploratory associations, not ready-made compatibility weights or a test that every ranking difference is meaningful.

For example, religious faith showed substantial resemblance between partners, but agreement in sexual interest had a larger association with satisfaction in this analysis. Similarity in indie-rock preference and self-rated financial wealth had almost no unadjusted linear association with satisfaction, despite partners resembling one another on those dimensions.

This does not mean those subjects never matter to a particular couple. It means that a feature can be common among existing partners without being a strong general predictor of their relationship satisfaction.

Similarity Related to Different Parts of the Relationship

The strongest association among the six averaged relationship ratings was with perceived life-path compatibility. The correlations with the other ratings were smaller.

Sex-life satisfaction has fewer usable responses because “Not Applicable” was not treated as a zero satisfaction rating.

Partner Resemblance Does Not Establish a Success Threshold

In an illustrative check, actual pairs averaged 75.9 on the recomputed similarity index. Re-pairing the same profiles within test-version and reported-gender groups produced an average of 71.0. This suggests that the observed pairings were more alike than such re-pairings of this sample.

It does not tell us how those imaginary relationships would turn out. The re-paired profiles have no observed relationship outcome, and this volunteer sample is not a norm for everyone who might use a future app.

Romantic Partners and Friends Were More Similar Than Random Pairs

The historical Facebook dataset contains 253,717 eligible stored test profiles. We reconstructed relationship groups from the original directed labels, deduplicating unordered pairs within each group. For the trait-by-trait comparison, both people had to label one another as romantic partners, siblings or best friends: 478 partner pairs, 233 sibling pairs and 885 best-friend pairs.

These reciprocal labels are important. The old app also stored a Facebook partner field, but its code automatically mirrored that field to the other person. A mirrored field does not establish two independently entered relationship labels; it is not the definition used for the graph.

The shared pattern is similarity in both romantic pairings and friendships. Mutual romantic partners averaged 79.21 in closeness and mutual best friends averaged 79.10, compared with 77.07 for randomly paired participants. The broader friendship group, with known romance and family labels excluded, also exceeded that baseline, averaging 78.52. These are group averages: not every romantic pair or friendship was more similar than every random pair.

Closeness is 100 minus the mean absolute percentage-score difference. Height and the orientation-related dimension are excluded from this 39-dimension summary. This is not a percentile or the current weighted compatibility score. Groups can overlap, and one person can occur in multiple pairs.

Mutual romantic partners, siblings and best friends were almost equally similar overall. Their means were separated by less than 0.11 points on this scale, so this analysis does not establish a reliable ranking among those three groups. Disliked connections were the least similar of the relationship groups examined, but were approximately as similar as random participants—not clearly less similar than random people.

Which Similarities Stand Out in Romantic Partners?

Overall resemblance can hide differences between traits. The graph compares the same trait’s partner-to-partner correlation across all three reciprocal relationship groups. Traits appear first when the romantic-partner correlation exceeds both the sibling and best-friend correlations by the largest amount.

• Desire for children/family had the clearest partner lead: r = 0.377 in romantic partners, compared with 0.106 in siblings and 0.221 in best friends. This makes explicit family preferences a useful candidate for further compatibility research.
• Exercise and resisting subordination also showed partner leads: exercise was 0.253 versus 0.107 and 0.158; resisting subordination was 0.173 versus 0.044 and 0.083. These are exploratory contrasts, not evidence that every observed gap will replicate.
• Religion showed the strongest absolute partner resemblance, but was not unique to romance: correlations were 0.515 in partners, 0.433 in siblings and 0.376 in best friends.
• Alcohol/tobacco attitudes were almost as similar among best friends as partners: 0.401 versus 0.385. It would be misleading to treat this strong partner resemblance as specifically romantic.
• Financial-wealth resemblance was stronger in siblings and best friends: 0.432 and 0.356, compared with 0.314 in partners. Shared background is one plausible explanation; these data do not identify the cause.
• Some partner correlations were near zero or negative. Disclosure/outgoingness, emotional calm and cautiousness differed from the sibling pattern. This does not establish that opposites have happier relationships, and the formula does not convert negative correlations into rewards for opposing trait scores.

These comparisons identify candidate romantic-specific patterns, not proven unique traits. The samples are unequal, people repeat, groups overlap, and demographics and network composition have not been adjusted. A stronger follow-up would use matched random pairs and uncertainty estimates that account for repeated people and overlapping networks.

What This Means for Compatibility Scores

These results support keeping three outputs separate:

An 80th-percentile compatibility rank would mean that the estimated score is higher than approximately 80% of scores in the specified comparison group. It would not mean an 80% chance of a successful relationship.

The older questionnaire is not the current Distinct Traits test. Its findings can guide hypotheses, but its weights and percentiles cannot simply be transferred to the newer instrument. Some potentially important topics, including plans for children and relationship preferences, may require separate questions rather than being inferred from a loosely related personality trait.

The Version 1.1 formula combines mapped Version 1.02 evidence with the audited reciprocal Facebook correlations. Where a defensible matching construct exists, the two source-based weights receive equal influence; a small, bounded adjustment gives additional emphasis to similarities that were stronger in partners than in both comparison groups. Traits without a matching construct retain equal baseline weight. Cross-test mapping strengths and the source blend remain design judgments pending direct bridge data and outcome validation.

The earlier weight update increased the emphasis on activity and anti-authority preferences and added the disclosure and self-reliance evidence. It does not infer religious faith, plans for children, wealth or physical attractiveness from unrelated personality scores. Version 1.2 retains those weights and the absolute-distance definition, while improving score precision, native percentile calibration and the norm context in trait descriptions.

Testing Alternative Similarity Formulas

We compared eight fixed methods: absolute and RMS distance, their standardized forms, regularized Mahalanobis distance, distance relative to each trait’s typical random-pair gap, and two empirical gap-percentile methods. Absolute distance ranked first on the reported main measures. More elaborate normalization did not improve these comparisons.

Version 1.02: 23,458 relationship records, 20 dimensions, five folds split by 21,133 connected identifier groups. Facebook: 478 reciprocal partner pairs, 233 sibling pairs and 885 best-friend pairs across 39 dimensions. Norms were fitted on 9,607 sampled Facebook profiles; the 100,000 random comparisons used a separate pool of 9,607 profiles. Known group endpoints were excluded from both pools.

AUC 0.642 means that a randomly selected observed romantic pair ranked as more similar than a randomly selected reference pair about 64.2% of the time, counting ties halfway. It is not 64.2% classification accuracy or a probability that a relationship will succeed. Differences between the leading methods were small: for example, the partner AUC difference between absolute and RMS distance had a component-bootstrap interval that included zero.

The original method also ranked first when Facebook endpoints were re-paired within each relationship cohort rather than sampled from the separate population pool. These are exploratory comparisons in the available historical samples, not proof that absolute distance is optimal for every population or for the current 1.14 instrument. The benchmark does not independently validate the separate transferred compatibility weights.

How Much Closer Than Random Pairs?

A more intuitive expression of the historical difference is the reduction in the average gap, relative to the exact random-participant baseline. This changes the interpretation, not the evidence or ranking.

Gaps are percentage points averaged across the 39 non-orientation, non-height dimensions. The exact full-population random gap above differs slightly from the sampled reference used for AUC. Larger-looking relative percentages do not create a stronger statistical effect.

Using Native Distinct Traits Norms

The native reference starts with all 3,610 English Version 1.13 completed records available at analysis time. Thirteen constant keyed-response vectors were excluded as a conservative quality check, leaving 3,597 records. Twenty-eight retained constructs use their observed 1.13 joint distribution. Self-reliant and Present-focused retain their constructs but have wording revisions, so their norms remain provisional transfers. Orderly, Temperate, Focused and Overconfident replaced earlier constructs in 1.14 and use temporary pooled-average marginal distributions from the retained traits, independently of the other traits. The old four constructs are not relabeled as the new ones.

Norm context can be useful even though norm-standardized overall distances did not win. In the native data, a roughly 38.9-point gap is as small as or smaller than about 25.5% of random pairs on Civic-minded, versus 39.7% on Pet-centric. Typical gaps were 24.9 and 33.1 points respectively. Neither result means a literal percentage of shared personality.

The Version 1.2 update retains absolute distance and the Version 1.1 compatibility weights. It uses exact three-item sums for comparisons and rounds only displayed results. Its percentile reference samples 500,000 random distinct-record pairs, retaining the observed 28-trait joint profiles and sampling the four placeholders. The reference’s average weighted score is 70.00. This is a provisional 1.13-based reference for 1.14, not a fully measured 1.14 population norm.

The Limits of the Study

• Both sources were self-selected app or website samples, not representative surveys of couples.
• Relationships and satisfaction were observed at the time of testing. This is not a study of future breakups, marriage duration, or the effects of deliberately choosing a similar partner.
• Reports about a partner can be influenced by how someone currently feels about the relationship. The workbook alone does not establish how independently every profile description was collected.
• Most mutual satisfaction ratings were high. The average was 8.21/10, and 68.7% of records averaged at least 8. Unhappy and ended relationships may be underrepresented.
• The Facebook data measures relationship labels and personality resemblance, not relationship satisfaction. Labels are participant reports, not independently verified kinship or current relationship status.
• Random participants are not verified strangers. The baseline includes the possibility of known connections, and it does not match partner pairs on age, gender, location or other background factors.
• Possible overlap between the two historical sources has not been established. They are analyzed separately; their sample sizes are not added together as an independent replication count.

Bottom line: personality similarity was a recurring feature of romantic pairings and friendships in these historical data. Romantic partners and best friends were more alike on average than randomly paired participants; the broader friendship group also exceeded that baseline. This shared resemblance pattern does not establish that similarity causes relationship success. A compatibility score still needs to be tested against relationship outcomes.