Prior to the main analyses, we investigated whether there was any interaction between the two independent factors: the virtual reality scenario (experience versus information) and the labelling (explicit versus indulgent). As we did not find any interactions, we conducted randomisation checks for the two factors separately. We did not find any difference between the groups; therefore, we consider the randomisation procedure successful (for a complete analysis of the interactions and group differences, see Supplementary Results).
First, we investigated whether the VR experience scenario was more effective in inducing intentions to reduce meat consumption. As hypothesised, the VR experience of the future scenario led to higher intentions (M = 4.28, SD = 1.26) as compared to the VR information scenario (M = 3.79, SD = 1.24), W = 2621, p = 0.011, r = 0.2, 95% CI [0.04, 0.34]; therefore, Hypothesis 1 was supported. See Fig. 3A for the comparison.
Furthermore, as hypothesised, in VR, the participants in the VR experience condition selected food with less meat than participants in the VR information condition, W = 4090.5, p = 0.028, r = 0.17, 95% CI [0.02, 0.32]. Therefore, Hypothesis 2 was also supported (see Fig. 3B).
When analysing the impact of the conditions on the real-life buffet choice, the difference between the VR conditions was not significant, W = 3063, p = 0.218, r = 0.09, 95% CI [0.01, 0.24],[1] and therefore, Hypothesis 3 was not supported. Nevertheless, the intelligent buffet allowed us to measure the exact consumption levels for all food and, thus, the consumption of beef in grams. An independent t-test showed that participants in the VR experience condition took significantly less beef from the buffet (M = 60.5, SD = 90.3) as compared to those in the VR information condition (M = 99.4, SD = 145), ΔM = 38.9, 95% CI [1.87, 75.85], t(137.12) = 2.08, p = .040, d = 0.32, 95% CI [0.03, 0.6]. As beef consumption consisted of a two-step decision, that is, selecting the vegetarian or non-vegetarian option (1) and the amount of sauce taken (2), which resulted in a zero-inflated distribution, we opted to use the Hurdle model to investigate the impact of the VR condition on beef consumption. We found that the VR experience condition influenced the amount of consumed beef, b = -0.34, z = -18.49, p < .001, but did not significantly impact the participant’s choice of a vegetarian versus non-vegetarian dish, b = -0.31, z = -1.00, p = .316, which is in line with our previous analysis. Therefore, even when the participants in the VR experience condition chose the meat option, they ate less meat as compared to the VR information condition. See Fig. 3C for the comparison.
To investigate the effect of nudging, we compared the impacts of indulgent and explicit labelling on the real-life buffet choice. According to our results, the participants in the explicit-label group did not differ from participants in the indulgent-label condition in their real-life buffet choice, W = 3327, p = 0.811, r = 0.02, 95% CI [0.003, 0.19]; therefore, Hypothesis 4 was not supported.
Furthermore, we investigated the impact on real-world behaviour in terms of reported dietary footprint, which was measured one week after the intervention. As hypothesised, the results of the food frequency questionnaire administered one week after the intervention confirmed that participants significantly decreased their dietary footprint from pre- (M = 120.78 kg/CO2) to follow-up (M = 100.42 kg/CO2), ΔM = 24.6, 95% CI [6.68, 42.46], t(144) = 2.71, p = .007, d = 0.23, 95% CI [0.06, 0.37]. Therefore, Hypothesis 5 was supported. Importantly, the self-reported measure of dietary footprint in the follow-up was significantly correlated with the real-life buffet food choice, rho = 0.36, p < .0001, which supports the validity of the self-report measure of real-life food behaviour.
We found no difference between conditions in terms of dietary carbon footprint, F(1,142) = 0.41, MSE = 7,519.10, p = .521, η ̂_G^2 = .003; therefore, Hypothesis 6 was not supported. See Fig. 3D. Nevertheless, the change in VR choice (difference between the first and second selections in the VR simulation) predicted the change in dietary footprint from pre- to post-test for the participants in the VR experience condition, b = 54.46, 95% CI [10.66, 98.27], t(73) = 2.48, p = .016, but not those in the VR information condition, b = -27.10, 95% CI [-69.00, 14.80], t(68) = -1.29, p = .201, suggesting that the VR experience condition influenced the participants’ behaviour one week after the intervention.
To gain a better understanding of the psychological mechanisms behind the behavioural change intervention, we tested various mediation models using the PROCESS macro51. All reported indirect effects were tested for significance using bootstrapping procedures, with the indirect effects being computed using 10,000 bootstrapped samples and 95% confidence intervals being computed by determining the indirect effects at the 2.5th and 97.5th percentiles. As preregistered, real-life food choice was treated as a dichotomous variable (meat versus vegetarian choice), as PROCESS uses logistic regressions only in the case of dichotomous outcome variables and, therefore, is not suitable for using ordinal variables as an outcome.
The first model (preregistered) assumed a direct effect on the part of self-efficacy and response efficacy on the real-life buffet choice. As hypothesised, the preregistered model revealed indirect effects on the part of self-efficacy (indirect effect: -0.15, 95%CI [-0.44, -0.003]) and response efficacy (indirect effect: -0.18, 95%CI [-0.446, -0.022]). Therefore, Hypothesis 7A was confirmed (see Fig. 4a).
To provide a more comprehensive model in line with PMT, we included intentions as a predictor of behaviour. Mediation analyses showed that the VR experience condition indirectly influenced the real-life buffet choice via increases in self-efficacy and response efficacy, which led to an increase in intentions and, consequently, more sustainable food choices. That is, those exposed to the VR experience scenario were more likely to improve their self-efficacy, which increased their intentions to reduce meat consumption, and they were, therefore, more likely to choose more sustainable food (indirect effect: -0.14, 95% CI [-0.336, -0.0106]). Furthermore, they were also more likely to exhibit increased response efficacy, which, again, improved intentions and, consequently, led to more sustainable food choices (indirect effect: -0.08, 95% CI [-0.179, -0.009]). See Fig. 4b.
We tested the same model for dietary carbon footprint, which was measured one week after the intervention. The preregistered model confirmed the indirect effect of the VR experience scenario via increased response efficacy (partially standardised indirect effect: -0.09, 95% CI [--0.204, -0.003]) but not via self-efficacy (partially standardised indirect effect: -0.08, 95% CI [--0.191, 0.0002]). Therefore, Hypothesis 7B was only partially confirmed. Nevertheless, exploring the comprehensive model, including intentions, we found the same indirect pathways as for food choice. We found a significant indirect effect on the part of self-efficacy on intentions (partially standardised indirect effect: -0.041, 95% CI [-0.105, -0.001]), as well as an indirect effect on the part of response efficacy on intentions (partially standardised indirect effect: -0.029, 95% CI [-0.068, -0.002]). See Figs. 5a and 5b for a graphic depiction of the models.
Similarly, exploring the same comprehensive model regarding beef consumption in grams, we also found a significant indirect effect on the part of self-efficacy on intentions (partially standardised indirect effect: -0.05, 95% CI [-0.1128, -0.004]) and indirect effects on the part of response efficacy and intentions on beef consumption (partially standardised indirect effect: -0.03, 95% CI [-0.066, -0.003]). That is, again, the experience VR scenario increased participants’ self-efficacy and response efficacy, which, in turn, enhanced intentions to reduce meat consumption and resulted in lower consumption of beef in grams in the buffet. See Fig. 6 for details.
Furthermore, as the VR buffet was intentionally designed to be an exact visual copy of the real-life buffet, we explored the relationship between the participants’ behaviour in VR and their real-life behaviour. We found a strong correlation between the second VR food choice (after receiving all the information) and food choice in real-life, rho = 0.6, p < .0001, and we also found a correlation between the second VR choice and dietary footprint in the follow-up, rho = 0.34, p < .0001. On the other hand, the reported intentions show a significant but lower correlation with the real-life food choice than the VR food choice, rho = − 0.48, p < .0001, indicating that VR behaviour may be a better indicator of PEB in the real world as compared to reported intentions.
Lastly, as preregistered, we investigated the impact of indulgent labelling on taste expectations. The indulgent label group did not differ from the explicit label group in terms of the taste expectations, M = -0.18, 95% CI [-0.59, 0.24], t(161.91) = -0.84, p = .400, d = -0.13, 95% CI [-0.45, 0.18]. Therefore, Hypotheses 8 and 9 were not supported. Finally, despite the taste expectations predicting the food choice in the buffet, we did not find any evidence for the indirect effect of labelling via taste expectations (indirect effect: -0.11, 95% CI [-0.396, 0.139]). See Fig. 7 for details.
[1] In the preregistration, we specified that we would analyse the impact of the VR condition and the labeling condition on food choice in the buffet using the Kruskal-Wallis test; however, as the Kruskal-Wallis test does not have an option for two independent factors, we opted to use the Wilcoxon Rank Sum Test.