Facebook Personality Data: Age, Gender and Relationship Patterns

The Most Interesting Patterns

• Dating versus single: higher sexual interest (+6.9 points), less emotional distance (−4.9) and higher self-rated attractiveness (+4.0).
• Married versus single: stronger desire for children/family (+8.5), less emotional distance (−5.1) and less desire to be different (−3.8).
• Complicated versus dating: more emotional distance (+5.1), more distress at being ignored (+3.1) and lower sexual interest (−3.7).
• Older versus younger adults: the age-50 model predicts more willingness to admit mistakes (+11.0), more orderliness (+6.5), and less thrill seeking (−10.2) than the age-20 model.
• Trait clusters: emotional distance goes with suspicion of others (r = +.403); money importance goes with beauty importance (r = +.408).

Differences are percentage points on 0–100 questionnaire scores, not percent increases. These are group averages with substantial individual variation, not diagnoses or predictions about particular people.

Relationship Status Is Worth Separating

Each category is compared with single users under the same modeled age and gender distribution. “In a relationship” is kept separate from engaged and married; open and complicated statuses are not lumped into the partnered group.

An interesting contrast: religious faith is lower among dating than single users (−5.8 points), but higher among married than single users (+4.5). Combining all partnered statuses would obscure this pattern.

Open-relationship users show higher alcohol/tobacco scores (+8.2 points versus single), but that group has only 677 retained profiles. Treat its estimates more cautiously than the much larger dating, single and married groups.

Age and gender adjustment reduces two obvious sources of confounding, but does not control education, country, income, selection into this app, social networks or prior relationship history.

'It's Complicated' Versus 'In a Relationship'

This is a direct comparison, not two separate comparisons with single users. Positive values mean a higher score in the complicated-status group.

The two groups differ most clearly in emotional distance (+5.14 points; d = +0.249). The estimate is +4.38 in one random half and +5.91 in the other; age-band adjustment gives +5.16.

A complicated status could reflect relationship circumstances, personality, reporting style or other factors. The data cannot establish which comes first, and the label is not a measure of relationship quality.

Age Patterns

These are modeled cross-sectional age differences, holding recorded gender and relationship-status mix constant. They are not evidence that the same people changed by this amount as they aged.

Two findings need particular care: older profiles score higher on self-rated excess weight (+17.7), and lower on “financial wealth” (−8.8). Weight is a short self-rating, not measured BMI. The wealth items mix personal wealth, family wealth and wealthy friends; they are not a measure of income.

Observed Age-Band Averages

These are raw means, not adjusted estimates. They help show the actual age coverage and are not interchangeable with the age-50 minus age-20 model.

60.5% of retained profiles are ages 18–24. Ages are approximate: the old application sometimes inferred birth year from education/graduation dates, and the database does not identify which ages were inferred. Age, birth cohort and historical period cannot be disentangled here.

Recorded Gender Patterns

81,494 profiles were recorded female and 54,488 male. Results below compare those two historical metadata categories under the same modeled age and relationship-status distribution. The archive does not support a meaningful analysis of other gender categories.

Female-group averages are higher on self-rated attractiveness (+4.7), family orientation (+4.4), humanitarian concern (+3.8) and putting others first (+3.4). Male-group averages are higher on emotional calm (+10.4), thrill seeking (+6.7), exercise (+6.4) and enjoying conflict (+5.8).

Labels matter: the “Intellectual interests” difference (female minus male −7.8 points) comes from physics, theory and difficult-vocabulary interests. It is not an intelligence or ability test. “Self-rated competence/superiority” is not a clinical narcissism diagnosis.

Age-by-Gender Patterns

The female-minus-male emotional-calm gap is 6.2 points less negative at modeled age 50 than 20. Other gaps vary with age too. These are interactions in a cross-sectional sample, not trajectories within individuals.

How the Traits Relate to Each Other

These are ordinary Pearson correlations across all 253,717 valid profiles, without age or gender adjustment. Unlike the demographic models, they can include users without usable demographic metadata. Related item wording, common response style and demographic composition can contribute to these associations.

Emotional distance forms a recognizable cluster: less outgoing disclosure (r = −.454), more suspicion (+.403), and less acceptance of help (−.385). Another cluster joins money importance, beauty importance and expensive consumption. These clusters suggest overlapping dimensions rather than 39 completely independent traits.

Correlations do not show that one trait causes another, nor do they establish a latent personality structure. Formal factor analysis would be a separate analysis.

All Results

All 39 dimensions are shown for every contrast, sorted by absolute standardized difference. “Effect / SD” is the point difference divided by the trait's SD in the pooled retained cohort; it is not a probability. Half A and Half B are independently fitted person-level random splits. The age-band check replaces the quadratic age curve with age-band indicators and gender-by-band interactions.

The final dimension, Likes indie rock, uses liking the band Pixies as a narrow historical proxy, not a multi-artist measure of indie-rock preference. It is retained for completeness, but excluded from personality highlights. Same-gender attraction and relative height are not part of these 39 dimensions.

Download all demographic estimates (CSV) / Download all trait correlations (CSV)

Methods and Limits

1. One stored profile per Facebook user ID. Eligible scores are 0–100, tests are dated April 26, 2008 onward, and entirely-50 profiles are excluded as possible defaults/straightlining. This is a heuristic, not a complete response-quality assessment.
2. Of 253,717 valid profiles, 136,241 have recorded m/f and plausible approximate adult age 18–80. To avoid exporting small cells, each random-half × gender × relationship status × age-band cell must contain at least 30 profiles. This drops 259 more profiles and leaves 135,982.
3. Each trait is modeled by OLS with quadratic age, recorded gender, gender-by-age interactions, seven status categories, and gender-by-status interactions. Age is centered at 30 and scaled in decades. Exact sufficient statistics reconstruct the same coefficient estimates as individual-level OLS for this design; no individual rows leave the database.
4. Contrasts use a common reference distribution from the pooled retained cohort. Relationship-status differences assume the modeled status effect does not vary with age; sparse age/status combinations and model extrapolation remain limitations. Gender-specific status effects are allowed.
5. Fits are repeated in two deterministic random user-ID halves (68,263 / 67,719), using a fixed analysis salt. This is a stability check, not a preregistered external validation sample; both halves belong to the same social network and historical sample.
6. An alternative model uses five age-band indicators and their gender interactions. The maximum change across all 39 complicated-versus-relationship estimates is 0.15 points. This does not rule out other confounding or measurement error.
7. All estimates are exploratory. No statistical-significance claim is made. Highlighting |effect/SD| ≥ .15 and agreement in split direction is a descriptive screen, not multiple-testing control. Strongest-absolute rankings can exaggerate selected estimates.

Important Limits

• Self-selected historical app users are not representative of the public or today's Facebook users.
• Age can be inferred from graduation year; metadata may be updated separately from the test date. The exact contemporaneous age and relationship status are therefore uncertain.
• Scores were stored by the old application. Some historical profiles could originate from pasted results; the source does not identify every collection route or item version.
• Self-report, gendered item content, item overlap, short scales and unmeasured cultural factors can influence differences. These data cannot demonstrate biological explanations, clinical diagnoses, causal effects or within-person development.
• Split agreement does not account for dependence among friends/family or make small differences practically important. This analysis is descriptive and does not change the live compatibility formula.

Related SimilarMinds Research and Tests

These findings describe differences between people and groups. For a separate analysis of similarity between pairs, including partners, best friends and siblings, see Personality Similarity and Romantic Compatibility.