Sequential dependency for affective appraisal of food images

The study examines whether affective responses to food images are influenced by immediately preceding images, testing for sequential dependencies in emotional appraisal. Drawing on the circumplex model of affect (valence and arousal) and prior evidence that perceptual judgments exhibit serial dependence, the authors hypothesize positive (assimilative) serial dependence if food images evoke emotions that persist long enough to influence subsequent appraisals. Given known cross-cultural differences in food perception and the growth of O2O food delivery where imagery strongly impacts choice, the study also explores whether nationality, age, gender, and BMI moderate such dependencies. The purpose is to clarify temporal carry-over in emotional appraisal of food imagery and its potential implications for consumer interfaces.

Prior work shows images strongly influence consumer attitudes and behavior, with emotions being key drivers of consumption. Serial dependencies have been widely reported for low-level visual features (orientation, motion, color) and also for complex judgments (e.g., face attractiveness, decision-making in sports). Serial effects can be positive (assimilative) or negative (repulsive), potentially reflecting higher-order decisional processes versus low-level perceptual adaptation. Limited research has assessed serial dependence for emotions; Liberman et al. (2014) found positive dependence in ratings of facial expressions, which may conflate emotional and visual processing specific to faces. Cultural factors influence food perception; e.g., Japanese participants show different arousal/valence use compared to Europeans. Demographic factors (age, gender, BMI) also relate to food-evoked emotions and cravings. The present study extends this literature by testing serial dependence for amodal emotional dimensions (valence, arousal) evoked by food images, across diverse nationalities, and by examining demographic moderators.

Design: Exploratory re-analysis of a cross-national dataset including both published and unpublished data. Within-subjects inter-trial analyses assessed whether ratings on a given trial depended on ratings on up to four preceding trials. Participants: 1322 recruited from 16 countries; after exclusions based on age (<18, >80, or missing), unrealistic/missing height or weight, N=1278 (795 females). Mean age 29.8 years (18–72). Mean BMI 23.6 (13.2–70.2). Recruitment via social media/email; Germany/UK via Prolific; Japan via Crowdworks. Ethics: TNO Ethics Committee (2017-011), in accordance with the Helsinki Declaration. Stimuli: 60 food images (850×640 px): 50 from CROCUFID and 10 from FRIDa, selected to span the valence range from aversive to highly pleasant. Stimulus set and screenshots available at https://osf.io/cyqg7/download. Measures: Valence and arousal collected with the EmojiGrid (0–100 for both axes), a pictorial 2D affective self-report tool validated across ages and cultures for food and other affective stimuli. Procedure: Online survey (Gorilla). Instructions to use a computer in fullscreen and to give first-impression ratings; two practice trials with EmojiGrid. Then 60 food images presented in random order; participants rated valence and arousal for each image. Average duration ~10 minutes. Instructions in English except translated for China (Mandarin), Indonesia (Bahasa), Iran (Farsi), Turkey (Turkish). Demographics (age, gender, height, weight) collected. Data Analysis:

• Valence serial dependence: For each participant, compute median valence across 60 images; label each trial as low (< median) or high (≥ median). For each lag (t−1 to t−4), bin current-trial valence by whether the lagged trial was low vs high, compute mean difference (high−low) as inter-trial effect. Exclude first four trials for lag analyses up to t−4. Repeated-measures ANOVA on inter-trial effects with factor trial distance (1–4). One-sample two-tailed t-tests per lag against zero. Huynh–Feldt corrections as needed.
• Arousal serial dependence: Identical analysis pipeline for arousal.
• Image-level analyses (t−1 only): For each of the 60 images, independent t-test comparing current-trial ratings when preceded by low vs high prior rating (valence and arousal separately). Multiple comparisons controlled via FDR (Benjamini–Hochberg). Implemented in Python SciPy.
• Moderators: ANCOVA on inter-trial effects with age and BMI (continuous), gender and nationality (categorical). BMI computed as weight (kg) / height (m)^2. If a covariate significantly influenced serial dependence, follow-up ANOVA with previous rating (low/high) as within-subject factor and the covariate as between-subject factor. Software: JASP for main statistics; SciPy for image-level tests.

• Valence serial dependence:
  • Significant effect of trial distance: F(3,3831)=34.855, p<0.001.
  • t−1: Positive dependence, inter-trial effect=+2.4 points, t(1277)=9.277, p<0.001.
  • t−2: Not significant, −0.1, t(1277)=0.457, p=0.647.
  • t−3: Not significant, −0.4, t(1277)=1.749, p=0.081.
  • t−4: Significant repulsive effect, −0.7, t(1277)=2.824, p=0.005.
• Arousal serial dependence:
  • Significant effect of trial distance: F(3,3831)=80.590, p<0.001.
  • t−1: Positive, +4.3, t(1277)=16.907, p<0.001.
  • t−2: Positive, +1.6, t(1277)=6.267, p<0.001.
  • t−3: Positive, +0.5, t(1277)=2.085, p<0.05.
  • t−4: Not significant, +0.2, t(1277)=0.605, p=0.545.
• Image-level analyses (t−1):
  • Valence: Positive dependence for 55/60 images (91.7%); significant in 50.0% pre-FDR and 35.0% after FDR. Strongest effects for images with mid-range mean valence (reduced near ceiling/floor).
  • Arousal: Positive dependence for 59/60 images (98.3%); significant in 66.7% pre-FDR and 63.3% after FDR; effects relatively consistent across arousal levels.
• Moderators (ANCOVA):
  • Gender significantly moderated arousal serial dependence: F(1,1244)=5.266, p=0.022. Trend for valence: F(1,1244)=3.450, p=0.063 (significant without nationality covariate: F(1,1274)=4.712, p=0.030).
  • Follow-ups: Serial dependence larger in males than females:
    • Valence: 3.4 vs 2.2; interaction F(1,1276)=5.217, p=0.023; both sexes significant (males t(482)=8.380, p<0.001; females t(794)=6.671, p<0.001). Mean valence higher in males (43.3) than females (41.8), F(1,1276)=9.217, p=0.002.
    • Arousal: 5.1 vs 4.1; interaction F(1,1276)=3.945, p=0.047; both sexes significant (males t(482)=12.873, p<0.001; females t(794)=12.947, p<0.001). Mean arousal higher in females (59.1) than males (55.5), F(1,1276)=23.135, p<0.001.
  • Nationality, age, and BMI did not significantly moderate serial dependence (all p≥0.108–0.133).

Findings show that affective appraisals of food images are not independent across trials: current ratings assimilate toward recent ratings, evidencing positive serial dependence. Arousal exhibits a broader temporal window (up to three trials back) than valence, which shows immediate assimilation (t−1) and a delayed repulsive effect at t−4, suggesting different temporal dynamics for the two affective dimensions. The positive dependence supports an interpretation of emotional carry-over or integration over short timescales, potentially serving to stabilize affective perception in noisy environments. Alternative explanations include trial-to-trial carry-over chains (indirect effects). The lack of nationality moderation implies generalizability across diverse cultural backgrounds despite known cross-cultural differences in absolute valence/arousal ratings using the same tool. Gender differences indicate larger sequential effects in males, with females showing higher overall arousal and males higher overall valence; this may relate to sex differences in neural processing of food cues. The results have practical relevance: ordering of images in digital menus and O2O platforms could subtly bias users’ affective responses and potentially choice by leveraging serial dependence.

This study demonstrates that affective appraisals of food images exhibit sequential dependencies: both valence and arousal ratings are assimilated toward prior ratings, with arousal effects persisting up to three trials and valence showing immediate assimilation and a later repulsion at four trials. These effects are robust across most images and across 16 countries, and are moderated by gender (stronger in males) but not by age or BMI. The findings contribute to understanding temporal integration in emotional appraisal and suggest applications for structuring food imagery in digital interfaces. Future research should test whether passive viewing (without explicit rating) yields similar dependencies, clarify mechanisms behind different temporal profiles for valence versus arousal, examine how serial dependence impacts actual choice behavior, and assess effect sizes in ecologically valid settings with controlled image order manipulations.

• Exploratory design with single presentation per image per participant increases noise in image-level inter-trial estimates.
• The task required explicit ratings on every trial; it remains unknown whether mere exposure without task elicits similar serial dependence.
• Effect sizes, while statistically significant, are modest (typically a few points on a 0–100 scale), raising questions about practical magnitude in real-world contexts.
• Online convenience sampling and varied instruction languages may introduce uncontrolled variability (e.g., device, environment), although procedures attempted to standardize display conditions.
• The analyses primarily address short-term sequential effects; causal mechanisms (e.g., emotional integration versus response bias) cannot be conclusively separated.