Why facets matter more than traits
A deep dive into the BFI-2 Rasch study (N=5,362) and what 15-facet resolution changes about dating compatibility — with implications for our extended question bank.
Updated 2026-08-25 · 2 min read
Most dating apps treat personality as five broad scores. The best contemporary work does the opposite — it splits each trait into three facets and insists each facet earn its keep as a measurement.
- Participants
- 5,362
- Italian adults, BFI-2 Rasch validation
- Facets validated
- 15
- 3 per Big Five trait, all fit Rasch model
- Scalar invariance
- Holds
- across sex, at the Rasch level
The study that changed our routing
The 2025 BFI-2-Rasch paper did something simple and rare: it took the most-used personality inventory and asked whether its facets satisfy fundamental measurement — not just factor analysis.
A trait-level score can reliably differentiate people. A facet-level score is thinner evidence — and should look thin on the surface.
For each facet the authors checked four gates:
- Monotonicity — higher trait level should mean higher endorsement
- Local independence — items shouldn't repeat each other (residual r < 0.20)
- Unidimensionality — one underlying dimension (first contrast eigenvalue < 2.0)
- DIF — no item favors one sex over the other (P-DIF within ±0.10)
Every facet passed. Convergent validity against BFI-10 was strong (r .56–.77). Trait-level reliability was good; facet-level PSR was modest for some facets — exactly the honesty we mirror in the product.
How we encode it
Before this study, questionCategoryHint routed multiple-choice answers through keyword matching (textSignalScore). For the extended bank we route by id prefix:
| Prefix | Trait | Facets covered |
|---|---|---|
opn- | Openness | imagination, artistic interests, emotionality, adventurousness, intellect, liberalism |
con- | Conscientiousness | self-efficacy, orderliness, dutifulness, achievement striving, self-discipline, cautiousness |
ext- | Extraversion | friendliness, gregariousness, assertiveness, activity level, excitement seeking, cheerfulness |
agr- | Agreeableness | trust, morality, altruism, cooperation, modesty, sympathy |
neu- | Neuroticism | anxiety, anger, depression, self-consciousness, immoderation, vulnerability |
EQ and relationship items get their own sub-prefixes (eq-saw-, relatt-, relcom- etc.) so each taps a narrow aggregate at 1.4× weight instead of a coarse blend.
What thin evidence looks like
When a facet has only one answered item, its stability is low even if the trait-level confidence is decent. That's why the product surfaces trait-level confidence as the primary signal and treats facet stability as diagnostic.
Good reliability. Safe to rank, safe to explain. This is what discover ordering and the weighted_by_your_history rationale use.
Reading further
- Our IPIP-facet approach — how we author 1-5 items and wire them
- BFI-2 Rasch study (2025) — DOI 10.1007/s12144-025-07584-7
- HEXACO honesty-humility separability (2025) — why honesty-humility gets its own bank
- 01
Batches 01-08 (shipped): 270 extended items across Big Five facets, relationship context, EQ, values, cultural context, lifestyle and goals, safety, and trust. They are facet-namespaced, forward-scored, and
active:falsefor the core MindSync instrument.Batches 09-11 (shipped): 60 more items in three non-Likert formats: scenario vignettes with four graded responses, forced-choice pairs to dampen agree-bias, and semantic-differential scales between two plain-speech anchors. Same facet namespaces, same forward keying.
- 02
Batch 02 (next): Honesty-humility 15 — the highest incremental-validity gap; requires a new
hh-aggregate, not a fold into agreeableness. - 03
Batches 03-05: Refined values, Gottman dynamics, cultural/context — to 500, with semantic dedup (cosine >0.82 flagged) and facet-dispersed daily batches.