Participants and design

We recruited 182 adult women (Mage = 31.4, age range: 18–64 years) for an online study from university and social network forums and among colleagues, friends, and acquaintances, to be randomly assigned to the conditions of a 2 (group: intervention vs. control) between-participants design. To homogenize and align the sample with previous studies [20, 24], we excluded women who disclosed a history of an eating disorder or who were currently receiving treatment for disordered eating. We initially envisioned including only healthy women with at least mild body image concerns, but dropped this criterion due to validity concerns (i.e., we intended to use a response toward a single question, which had not been validated) and because limiting the variability in the sample reduces the power for detecting moderations.

We targeted a minimum sample size of N = 128 to achieve a power of 1−β = 0.80 to detect a medium-sized difference (f = 0.25) between intervention vs. control condition for body dissatisfaction changes at α = 0.05. All data were collected from April 2022 to January 2023. Data and materials can be obtained from the corresponding author upon request. We report all measures, manipulations, exclusions and deviations from pre-registration (see Declarations).

Measures and procedure

We implemented the study in jsPsych [33]. Upon accessing the study’s website, participants were randomly assigned to the intervention or control condition and read the study information and consent forms. Study participation required a physical keyboard. After providing informed consent, participants self-reported their age, weight, height, regular medications via open-ended questions and gender, language ability, years of education, dominant hand, pregnancy status, history of eating disorders, current eating disorder treatment, and previous study participation via multiple-choice questions. We then assessed, in random order, disordered eating symptoms via the German short version of the Eating Disorder Examination-Questionnaire (EDE-Q8, [34] as a control measure [20,21,22,23, 25, 26] and, more importantly, pre-existing trait-like body image concerns as a potential moderator of pairing procedure effects using the German Body Shape Questionnaire (Fragebogen zum Figurbewusstsein, BSQ; [35]. The BSQ´s 34 items assessed thoughts and concerns about one’s body shape during the past 28 days on a scale from 1 (never) to 6 (always), with the sum score serving as a validated measure of (trait-like) body image concerns.

Target body selection

After completing the initial assessments, participants commenced to the main procedure. First, participants rated 44 standardized, black-and-white photographs of eleven women’s bodies (without heads and in neutral underwear), taken from four perspectives (standing-front, standing-back, standing-side, and sitting-side; for examples, see [36]. The models’ Body-Mass-Index (BMI) ranged from 13.8 to 61.3 kg/m2, and participants rated How similar does this body looks to your body? on a visual analog scale (VAS) ranging from -100 (not similar at all) to 100 (highly similar). We used each participant’s ratings to select the bodies with the highest and the median average similarity to serve as similar and dissimilar target stimuli, respectively. We chose the median rather than the least similar body for comparison, as we expected severely underweight or overweight bodies always to receive the lowest similarity ratings. Selecting the least similar body may thus have unintentionally confounded similarity with a specific weight dimension.

Baseline state body dissatisfaction

Next, participants completed the German Body Image State Scales (BISS, [37, 38], which was also used in recent studies [23, 24, 26] and thus served as this study’s change-sensitive measure of state-like body dissatisfaction. The mean across its six items measured the current (“right now, in this very moment”) evaluation and affect about the physical appearance and attractiveness using different 9-point scales (from − 4, very dissatisfied/very unattractive/very much worse than usual to 4, very satisfied/very attractive/very much better than usual). We coded the scale such that higher scores indicate a more positive body evaluation.

Pairing task and contingency learning assessment

After completing the BISS, we explained that the next task would show photos of bodies as well as positive and neutral images. Each task trial (see Fig. 1) started with the display of a fixation cross presented for 500, 1000, 1500, or 2000 ms (selected at random). Next, one of four similar or dissimilar body images was displayed for 1000 ms (i.e., the two selected bodies across the four perspectives), followed by the presentation of a positive or neutral target image (e.g., of cute animals, happy people, landscapes vs. daily objects). We selected a total of 48 positive and 48 neutral target images from various validated databases [39,40,41]. Upon presentation of the positive or neutral image, participants were asked to categorize it as either positive or neutral using their keyboard’s m and c keys, respectively. This ensured focusing attention on the valence of the source stimuli, which has been known to strengthen pairing procedure effects [32]. The categorization further allowed us to measure contingency learning behaviorally [30]. If categorized incorrectly, or if no response was given within 3000 ms, an error message was displayed for 500 ms before advancing to the next trial (a blank slide was displayed for correct responses). In total, the task comprised 96 trials, presented in random order, such that each positive or neutral image was displayed only once. The task took approx. 8 min to complete.

Fig. 1

Trial sequence illustration of the pairing procedure. Each trial started with the display of a fixation, followed by a similar or dissimilar body image, then followed by the presentation of a positive or neutral target image. Participants were asked to categorize the target images upon presentation using their keyboard c and m keys. Trials were presented in random order

Critically, in the intervention condition, pictures of the self-similar body were followed mainly by positive images (40 vs. 8). In contrast, pictures of the self-dissimilar body were followed less often by positive than neutral images (8 vs. 40). Thus, the similar body → positive contingency was ϕ = 0.67 in the intervention condition. In the control condition, there was no contingency between self-similarity and the valence of the source pictures (24 trials per each combination; similar body → positive contingency ϕ = 0). However, the overall number of similar and dissimilar body images did not differ between conditions, such that the total number of stimulus exposures and response base rates remained identical.

State body dissatisfaction post-intervention and conclusion

After completing the pairing procedure, participants next completed the BISS for a second time to detect induced changes in body dissatisfaction. Finally, we asked participants to rate How attractive do you find each body? on a VAS ranging from − 100 (not at all attractive) to 100 (very attractive) for each picture presented for target body selection. We collected these ratings to explore any potential effects of the procedure on a perceived shift in body attractiveness [36]. After these ratings, participants were thanked and dismissed.

Data aggregation and analysis

We extracted relative frequencies and median latencies of correct responses (i.e., response accuracy and response time in ms) for each participant from the data collected during contingency learning. These were submitted to two separate 2 (group: intervention vs. control) × 2 (target body: similar vs. dissimilar) × 2 (source valence: positive vs. neutral) analyses of variance (ANOVAs), with target body and source valence as repeated-measures factors, to assess the predicted effect of contingency learning.

The questionnaires showed high levels of internal consistency (Cronbach’s α = 0.96, 0.91, and 0.91/0.92 for the BSQ, EDE-Q8, and BISS baseline/post-intervention, respectively) and were thus aggregated according to their convention. We compared baseline scores (as well as participant characteristics) between conditions, and before and after the intervention with independent samples t-tests. Following recommendations [42], the effect on BISS scores post-intervention was examined in a 2 (group: intervention vs. control) analysis of covariance (ANCOVA), with baseline scores as covariate, as well as in a mediation analysis with contingency learning as a mediator and trait-like body image concerns as a moderator of the effect of contingency learning on current body dissatisfaction improvement (again, baseline BISS scores were included as covariate). We initially intended to include BSQ scores as a covariate in the ANCOVA, but omitted reporting this analysis as it yielded comparable effects, further moderations were described as exploratory analyses in the study registration.

The significance level for all analyses was set at p ≤ 0.05. Post hoc pairwise comparisons report Bonferroni-adjusted p-values for multiple comparisons. Effect sizes are reported as ηp2. Variable values are reported as means and standard deviations (SDs). The data were aggregated and analyzed with IBM SPSS 28 [43]. We used PROCESS v4.2 model 14 [44] with boot-strapped (10,000 samples) bias-corrected 95% confidence interval (CI) for moderated mediation analysis. The effects of the moderator were evaluated at its mean (M) and at M ± 1 SD.