Paper scaffold — RQ2: Compositional features predicting desire response

Status: Scaffold — not a submission draft
Source RQ: docs/RESEARCH-PROGRAM.md §II — What compositional features of figurative imagery predict desire response?
Target venues (choose one primary): British Journal of Psychology or Frontiers in Psychology (Perception / Perception Science track)

---

Title (draft)

The grammar of desire: Which compositional features of figurative images predict approach-oriented aesthetic response?
(Alternative: Embodied composition over art-historical label: multilevel models of figurative viewing behavior)

---

Abstract

A growing literature links depicted gaze, pose, and lighting to attention and empathy in museum contexts, but rarely at feature-level granularity across hundreds of heterogeneous figurative stimuli paired with multi-signal behavioral outcomes. We merge consent-filtered responses with AI-assisted visual_decompositions (45+ structured fields per experience_unit) and fit multilevel regressions: trial-level outcomes (rating, save, log dwell) nested within participants and crossed with stimulus units. Random intercepts for users and crossed random effects for units (where identified) quantify stable individual and image contributions. We test whether embodied/relational features (gaze, intimacy, negative space, light quality) explain variance beyond style tier, medium, and period proxies. Complementary SHAP analysis on a gradient-boosted tree provides interpretable global feature rankings without replacing the inferential model. Evidence is thin until the active corpus accumulates sufficient per-unit response diversity (docs/engine-room/02-instrument-validation.md); we pre-register minimum per-unit n for entering decomposition interactions. If hypotheses hold, results show which compositional variables deserve causal follow-up (e.g., sequence experiments).

Word count: ~210

---

Introduction — prose prompts (~400 words)

1. Hook with embodied simulation (1 paragraph). Freedberg & Gallese (2007) and follow-on neuroaesthetics (Chatterjee & Vartanian, 2014): why compositional grammar should matter for approach responses, not only recognition.

2. Limit of prior work (1 paragraph). Lab studies use small stimulus sets; art-historical labels confound medium with composition. Vela’s decomposition schema is instrumental—document its provenance (Claude vision pipeline) and the need for expert validation (tie to RQ12 / future work, not this paper’s core).

3. Present RQ2 (1 paragraph). State the hypothesis from the research program: embodied/relational features dominate over coarse style tags. Preview multilevel structure and SHAP as complementary lenses.

4. Contribution (2–3 sentences). First large-N mapping (once recruited) from fine-grained decomposition to multi-signal outcomes in figurative art + editorial photography jointly.

5. Ethics (brief). Figurative content; consent; avoid objectification language in Discussion.

---

Hypothesis (formal)

Let $y_{ijk}$ be outcome $k$ (e.g., save, log dwell, ordinal rating) for user $i$ on unit $j$ at trial position $t$. Let $\mathbf{x}_j \in \mathbb{R}^p$ be decomposition predictors for unit $j$ (standardized), and $\mathbf{z}_j$ coarse tags (style_tier, medium).

• H1 (dominance): Coefficients on gaze-related, intimacy, negative space, and light-quality features jointly explain strictly greater reduction in deviance than $\mathbf{z}_j$ alone (nested model test, $\alpha = .05$, pre-register df).

• H2 (specific direction): Direct gaze and tight framing increase dwell conditional on intimacy_level controls; boundary_flag probability increases in pre-registered high-intimacy × tight-framing cells (exploratory if sparse).

• H3 (robustness): SHAP top-$m$ features overlap substantially with multilevel significant predictors after FDR correction within the decomposition family.

---

Methods

Participants and design

Same consent filter as RQ1. Trials nested in users; units crossed (multiple ratings per unit across users encouraged for generalization).

Outcomes (level 1)

Outcome	Source	Model family
rating	responses.rating	Ordinal mixed logit or continuous robust LM
saved	responses.saved	Mixed-effects logistic
dwell_ms	responses.dwell_ms	LMM on log(dwell + 1); sensitivity on Tobit/censored

Predictors (level 2 / unit)

From visual_decompositions (exact column names in DB migration + lib/types.ts — refresh list at analysis time):

• Primary theory cluster: gaze (or equivalent), pose_type, light_quality, intimacy_level, fields encoding negative space / framing if present.
• Controls: rendering, medium, period_reference (if sparse, collapse categories).
• From experience_units: style_tier, primary_dimension, optional tags (dimensionality reduction or hand-picked tags pre-registered).

Multilevel model

Recommended specification (R lme4 / glmmTMB; or Stata mixed):

$$ y_{ij} = \beta_0 + \mathbf{x}j^\top \boldsymbol{\beta} + \mathbf{z}j^\top \boldsymbol{\gamma} + u{0i} + v{0j} + \epsilon_{ij} $$

• $u_{0i}$: user random intercept.
• $v_{0j}$: unit random intercept only if units have sufficient repeated ratings; else unit as fixed effect cluster-robust SE (Snijders & Bosker, 2012).

Interactions: Pre-register at most two decomposition × position interactions (e.g., fatigue × intimacy) to limit multiplicity.

SHAP complement

Train XGBoost or LightGBM on pooled trial matrix (user-level aggregates as additional columns). Use shap library; report mean |SHAP| ranking with bootstrap CI. SHAP is descriptive, not inferential.

Software

Python (pymer4 calling R), or pure R; marginaleffects for contrasts. Version-pin decomposition export date because visual_decompositions can be backfilled.

Vela data export path (by table)

Table	Role
responses	Outcomes + unit_id, user_hash, session_id, timestamps
visual_decompositions	Unit-level predictors (unit_id FK)
experience_units	style_tier, primary_dimension, status
profiles	Consent gate

Join path: responses.unit_id → visual_decompositions (1:1 per unit) → experience_units.id.

---

Pre-registered analysis plan (OSF-ready)

1. Primary model: Mixed model for saved with decomposition block + controls.
2. Secondary: Same for log(dwell).
3. Multiplicity: FDR within the decomposition coefficient family per outcome.
4. Missing data: Listwise deletion vs multiple imputation for sparse decomposition fields—choose one pre-registered.
5. Subgroup: Photographic (editorial_soft, cinematic) vs classical painting strata—report if cell sizes allow.

---

Power analysis

• Clustered data: Maas & Hox (2005): aim for ≥50 level-2 units (users) with ≥20 observations each for stable random slopes; Vela Phase 1 table targets ~50 users and 1,000 responses total (docs/RESEARCH-PROGRAM.md §IX).
• Crossed random effects for stimuli: Schielzeth & Forstmeier (2009) caution on random slopes for stimuli with few repeats—pre-register minimum responses per unit (e.g., ≥5) for unit random intercepts.
• Effect sizes: For logistic mixed models, simulate power with simr package given pilot variance components.

Citation: Maas, C. J. M., & Hox, J. J. (2005). Sufficient sample sizes for multilevel modeling. Methodology, 1(3), 86–92.

---

Data sources

Internal: As above; optionally unit_pool_history as time-varying confound if pool changed during data collection.

External: TODO(ASN-575) — cite normative pose/gaze perception studies from literature map. Optional: WikiArt metadata for period labels not used as primary predictors (avoid double-counting with decomposition).

---

Expected figures

#	Sketch
F1	Coefficient forest plot for decomposition block (95% CI).
F2	Predicted probability of save across intimacy × gaze (marginal means).
F3	SHAP beeswarm for top 15 features.
F4	Variance partition diagram: ICC_user, ICC_unit, residual.

---

Limitations

• Decomposition is AI-generated — measurement error and drift across model versions; freeze ANALYSIS_PROMPT_VERSION / pipeline commit in supplement.
• Confounding: Popular units get more responses; joint modeling of exposure needed for some robustness checks.
• Causal language: Multilevel coefficients are associational unless paired with experimental sequence manipulations (future RQ5–7).

---

Contribution statement

We provide a replicable multilevel framework linking fine-grained figurative composition features to approach-oriented behavioral outcomes in the wild, with explicit handling of nested observations and AI-labeled stimuli. The decomposition schema itself becomes a hypothesis list for curators and for adaptive systems that currently treat style tier as a proxy for content.

---

References (seed list — TODO ASN-575 merge)

1. Chatterjee, A., & Vartanian, O. (2014). Neuroaesthetics. Trends in Cognitive Sciences, 18(7), 369–375.
2. Freedberg, D., & Gallese, V. (2007). Motion, emotion and empathy in esthetic experience. Trends in Cognitive Sciences, 11(5), 197–203.
3. Lundberg, S. M., & Lee, S.-I. (2017). A unified approach to interpreting model predictions. NeurIPS.
4. Maas, C. J. M., & Hox, J. J. (2005). Sufficient sample sizes for multilevel modeling. Methodology, 1(3), 86–92.
5. Schielzeth, H., & Forstmeier, W. (2009). Conclusions beyond support: Overconfident estimates in mixed models. Behavioral Ecology, 20(4), 416–420.
6. Snijders, T. A. B., & Bosker, R. J. (2012). Multilevel analysis: An introduction to basic and advanced multilevel modeling (2nd ed.). Sage.

TODO(ASN-575): Add perception literature on gaze cueing, personal space in images, and museum eye-tracking.