peopleanalyst

research / vela / reports

RQ4 — Individual differences

Which dimensions of variance separate participants and how stable they are.

Vela·Reports·source: people-analyst/vela/docs/research/papers/rq4-individual-differences.md

Paper scaffold — RQ4: Individual differences in desire profile structure

Status: Scaffold — not a submission draft
Source RQ: docs/RESEARCH-PROGRAM.md §II — What individual differences predict desire profile structure?
Target venues (choose one primary): Personality and Individual Differences or Empirical Studies of the Arts

Title (draft)

Viewer types in figurative aesthetics: Factor structure and behavioral correlates of multidimensional desire profiles
(Alternative: Eight dimensions, how many people? Clustering aesthetic response in a behavioral corpus)

Abstract

Psychometric approaches to aesthetic taste often rely on self-report inventories; Vela instead induces implicit multidimensional profiles from Pass/Resonates, ratings, saves, and related signals mapped into eight desire dimensions (softness, intensity, narrative, structure, texture, abstraction, classical, contemporary). We ask how many latent factors explain covariance among these dimensions across users, whether finite mixture (cluster) models recover interpretable viewer types, and whether behavioral session features—exploration breadth, rating variance, boundary rate, session frequency, mode progression—predict cluster membership better than demographic placeholders (which Vela largely does not collect). Analyses use consent-filtered exports of user_desire_profiles and, for temporal sensitivity, user_desire_profile_versions. Construct validity and factor retention cannot be responsibly claimed until the eligible-user count exceeds thresholds in docs/engine-room/02-instrument-validation.md (spec §4.3 cites denominator 20 per dimension). We pre-register parallel analysis for snapshot-only vs most-recent-version-only profiles. Successful replication would yield a data-driven taxonomy of figurative “desire types” for theory and for cold-start personalization.

Word count: ~215

Introduction — prose prompts (~400 words)

  1. Individual differences without Likert blocks (1 paragraph). Contrast traditional aesthetic sensitivity inventories with behaviorally inferred profiles; cite Rentfrow & Gosling-style work only as analogy—be clear Vela is not using BFI.

  2. Eight dimensions (1 paragraph). Summarize their origin (docs/engine-room/01-math-spec.md, REINCARNATION_CONFIG): these are system constructs, not user-facing labels—Discuss limitations honestly.

  3. Research questions (1 paragraph). EFA → factor retention; cluster on factor scores or raw dimensions (pre-register); multinomial or latent class regression predicting cluster from aggregates computed in scripts/research/compute-features.ts style features (save rate, boundary rate, exploration index).

  4. Ethics. Profiles are sensitive; aggregation only; no re-identification narrative.

Hypothesis (formal)

Let $\mathbf{d}_i \in \mathbb{R}^8$ be the dimension score vector for user $i$ (latest snapshot unless modeling versions).

  • H1 (factor structure): A scree plot and parallel analysis favor m < 8 factors with interpretable loadings (e.g., a “classical–contemporary” bipolar plus a “soft–intense” axis)—pre-register rotation (oblimin).

  • H2 (clusters): Mixture or k-means on factor scores yields K ∈ {3,4,5} classes with silhouette > .50 or entropy criterion met—final K chosen by pre-registered information criteria, not post-hoc cherry-picking.

  • H3 (correlates): Behavioral predictors (pre-registered list from aggregated responses + player_sessions) explain incremental variance in cluster membership beyond chance (likelihood ratio test vs intercept-only).

  • H4 (demographics null): If any sparse demographic proxy exists in export, its incremental $R^2$ is negligible vs behavioral block (exploratory if absent).

Methods

Sample

Users with research_consent and total_responses ≥ threshold (align with profile manager eligibility, e.g., ≥5 eligible responses—cite lib/reincarnation/profile-manager.ts rules at analysis freeze).

Dimension scores

Source: user_desire_profiles.dimension_scores (JSONB) — parse to matrix $\mathbf{D}$ ($N \times 8$). Optional: use first vs last row per user from user_desire_profile_versions for sensitivity.

Exploratory factor analysis (EFA)

  • Determine factorability (KMO, Bartlett).
  • Maximum likelihood EFA with promax or oblimin rotation.
  • Retain factors with parallel analysis + scree + theoretical interpretability.

Cluster analysis

  • Gaussian finite mixture (mclust) preferred over naive k-means; report BIC.
  • Validate clusters with multivariate ANOVA on raw behavioral features not used to build profiles (hold-out behavioral set if engineered features overlap—split features into profile-training vs validation sets pre-registered).

Predicting cluster membership

Multinomial logistic regression or latent class regression if integrating mixture steps:
$\Pr(c_i = k) = \text{softmax}(\boldsymbol{\eta}_k^\top \mathbf{b}_i)$ where $\mathbf{b}_i$ is vector of session aggregates.

Vela data export path (by table)

TableColumns / role
user_desire_profilesdimension_scores, axes, boundary_tags, total_responses, session_count, profile_version
user_desire_profile_versionsLongitudinal replicate of dimension snapshots + created_at
responsesAggregates: mean rating, SD, save rate, boundary rate, counts per user
player_sessionsSession counts, mode distribution, mean duration

Feature engineering may mirror scripts/research/compute-features.ts outputs—version that script in the OSF component.

Pre-registered analysis plan (OSF-ready)

  1. EFA: 8 indicators, max 3 factors initially, expand only if parallel analysis supports.
  2. Clustering: Compare k=3..6 with BIC; pick winner by pre-registered rule.
  3. Stability: Bootstrap adjusted Rand index vs random halves.
  4. Regression: Block-wise entry of predictors.

Power analysis

  • EFA / CFA planning: MacCallum et al. (1999) sample size guidelines; conservative N ≥ 200 subjects for stable eight-indicator EFA; N ≥ 100 minimum for exploratory only with prominent caveats.
  • Cluster stability: Dolnicar (2002) bootstrapping mixtures; aim ≥150 users.
  • Multinomial regression: Peduzzi et al. (1996) events-per-variable rule—limit predictors to ≤ events/10.

Citations:
MacCallum, R. C., Widaman, K. F., Zhang, S., & Hong, S. (1999). Sample size in factor analysis. Psychological Methods, 4(1), 84–99.
Dolnicar, S. (2002). A review of data-driven market segmentation in tourism. Journal of Travel & Tourism Marketing, 12(1), 1–22.

Data sources

Internal: Profile tables + aggregated responses/sessions.

External validation (optional extension): If a short self-report aesthetic sensitivity scale is administered in a future wave, MTMM matrix—outside core paper. TODO(ASN-575) for relevant scales (e.g., Aesthetic Experience Scale).

Expected figures

#Sketch
F1Scree + parallel analysis for eight dimensions.
F2Factor loading heatmap after rotation.
F3Radar charts of cluster centroids on dimensions (normalize 0–1).
F4Multinomial coefficient plot for behavioral predictors.

Limitations

  • Construct confound: Dimensions are mathematically coupled through shared scoring code—factor structure may reflect engine design more than psyche; discuss need for independent validation.
  • Selection: Highly active users dominate early exports.
  • No demographics: Limits generalization claims—frame as behavioral typology.

Contribution statement

We offer a behavior-first typology of aesthetic engagement with figurative imagery, grounded in implicit signals rather than self-report, with explicit links to the adaptive engine’s dimension basis. If clusters replicate, they become hypotheses for personalization ethics (stereotyping risk) and for cross-domain prediction work (RQ7).

References (seed list — TODO ASN-575)

  1. Fabrigar, L. R., Wegener, D. T., MacCallum, R. C., & Strahan, E. J. (1999). Evaluating the use of exploratory factor analysis in psychological research. Psychological Methods, 4(3), 272–299.
  2. MacCallum, R. C., Widaman, K. F., Zhang, S., & Hong, S. (1999). Sample size in factor analysis. Psychological Methods, 4(1), 84–99.
  3. McCrae, R. R., & Costa, P. T., Jr. (1997). Personality trait structure as a human universal. American Psychologist, 52(5), 509–516. (analogy only)
  4. Rentfrow, P. J., & Gosling, S. D. (2003). The do re mi’s of everyday life: The structure and personality correlates of music preferences. Journal of Personality and Social Psychology, 84(6), 1236–1256. (method analogy)
  5. Peduzzi, P., Concato, J., Kemper, E., Holford, T. R., & Feinstein, A. R. (1996). A simulation study of the number of events per variable in logistic regression analysis. Journal of Clinical Epidemiology, 49(12), 1373–1379.
  6. Dolnicar, S. (2002). A review of data-driven market segmentation in tourism. Journal of Travel & Tourism Marketing, 12(1), 1–22.
  7. Chamorro-Premuzic, T., & Furnham, A. (2004). Art judgement: A measure related to both personality and intelligence? Imagination, Cognition and Personality, 24(1), 3–24.
  8. McManus, I. C., & Furnham, A. (2006). Aesthetic activities and aesthetic attitudes: Influences of education, background and personality on interest and involvement in the arts. British Journal of Psychology, 97(4), 555–587.

TODO(ASN-575): Add empirical aesthetics papers on individual differences, chills, Openness correlates, and digital behavior typologies.