In this commentary, we aim at extending Cesario's critique on the “use of experimental social psychology to explain real-world group disparities.” While we agree that focusing on average bias may impair external validity in significant ways, we disagree with the lesson Cesario draws from the use of experimental paradigms: Using such paradigms should be a tool of last resort to explain real-world disparities. Indeed, situated cognition experiments have already demonstrated their usefulness in shedding light upon the decision dynamics involved in bias, and not only on the presence of bias, which is Cesario's focus. We will develop why drawing Cesario's lesson would amount to throwing the baby out with the bathwater and will propose alternative solutions to his.

First, experimental paradigms in social psychology serve as models of the real world and as such (a) are by definition a simplification (parsimony in modelling); but (b) are still useful to understand the psychological processes (Smaldino, Reference Smaldino, Vallacher, Read and Nowak2017) that drive behaviour in the “real world.” This concern is not new and is aptly illustrated by George Box's aphorism (Reference Box1979, p. 202) “All models are wrong, but some are useful.” What Cesario is essentially saying is that current models in experimental social psychology are wrong and not even useful to explain real-world disparities. The three flaws accurately identified by Cesario are actually three types of missing variables – moderators – whose absence is involved in deterioration of external validity. One can reformulate this criticism as suggesting that the studied effect sizes are smaller than the smallest effect size of interest when more ecological variables – absent in most experimental models – are taken into account. Appraising the problem through this lens leads us to disagree with Cesario's idea that the “research is fundamentally flawed.” Still, the question remains: How to design experiments that yield more robust and meaningful effect sizes while accounting for these missing variables and which are applicable to real-life situations? The three identified flaws raise questions to which answers can help incrementally elaborate models that include crucial moderators.

Second, and related, we acknowledge that demonstrating average bias in “the general population” (often undergraduate, non-expert psychology students) with indirect measures like the implicit association test (IAT) may fall short in accounting for real-world group disparities. Nonetheless, using such paradigms to understand differential processes in real-world social groups has proven valuable. We were surprised that Cesario's section on implicit bias and science, technology, engineering, and mathematics (STEM)-related IAT did not refer to lines of research grounded in situated social cognition (Smith & Semin, Reference Smith and Semin2007) or Freeman's work on social perception (Freeman, Reference Freeman2014; Freeman, Pauker, & Sanchez, Reference Freeman, Pauker and Sanchez2016). For instance, using a mouse-tracking adapted gender-math IAT, completed by female and male STEM and non-STEM majors, research (Smeding, Quinton, Lauer, Barca, & Pezzulo, Reference Smeding, Quinton, Lauer, Barca and Pezzulo2016) has shown meaningful group differences in decision-making dynamics (i.e., attractions) and their early emergence in time (around 300 ms). Millisecond differences – or deviations in mouse trajectories while decision-making is unfolding – may thus represent one of those real-world forces that characterise real-world group disparities. These can be measured with experimental paradigms. Also, Cesario frames the interest about bias in decision-making in the minds of “gatekeepers” discriminating against ambiguous candidates. However, candidates themselves will be the first depleting link in the decision-making chain if they are biased by held stereotypes (e.g., Shapiro and Williams, Reference Shapiro and Williams2012) or because of imperfect inferences drawn from observed regularities (Kutzner & Fiedler, Reference Kutzner and Fiedler2017); leading attrition to occur way before formal selection by potential gatekeepers. By comparing engineering and humanities female students, Study 3 in Smeding et al. (Reference Smeding, Quinton, Lauer, Barca and Pezzulo2016) has shown that self-congruency trumps the role of stereotype-congruency in a “Math versus Language” IAT. Self-congruency would here be categorised by Cesario as a “missing force.” This moderator could not have been identified with a convenience sample. However, it can still be studied through an IAT paradigm, reconsidered from a situated social cognition approach. To explain underrepresentation of women in STEM majors, “men and women differ[ing] in their interest” (sect. 1, para. 6) can be considered as a mere demographic difference across groups, discarding the role of decision bias. However, the key role of self-congruency shows that the underrepresentation phenomenon can still be explained by decision bias (in the minds of candidates), which an IAT can meaningfully investigate while providing leads to implement change (e.g., impact on self-domain related associations). Still, such investigation requires participants who are involved in the actual field of application and not convenience samples, as called by Cesario. Besides, mouse-tracking-based paradigms, anchored in dynamical systems theory (which has nurtured many real-life applications spanning from physics to cognitive science, Krpan, Reference Krpan2017) represent a user-friendly tool (Rivollier, Quinton, Gonthier, & Smeding, Reference Rivollier, Quinton, Gonthier and Smeding2020) that allows understanding (a) continuity and (nonlinear) competition in decision-making (beyond discrete judgments) and (b) the influence of social category triggers specifically when people are ambiguous on a relevant real-world characteristic (Freeman, Reference Freeman2014; Freeman et al., Reference Freeman, Pauker and Sanchez2016).

Finally, experiments – as one of the methods available to psychologists – in combination not solely with in-depth (qualitative) field research (as suggested by Cesario), but also computational modelling have the potential to provide insight into real-world human behaviour, including group disparities. In Smeding et al. (Reference Smeding, Quinton, Lauer, Barca and Pezzulo2016), results for simulated social groups and real-world social groups were compared. While the former provided proof of concept regarding the (hypothesised) psychological processes, the latter sustained their real-world validity. Both contributed to a finer-grained understanding of sex differences in STEM engagement which, admittedly, seemed to be less related to average stereotypic bias than to differential associations related to the self. But an experimental paradigm like the IAT happened to be of paramount importance in such findings. Mixed-methods (including qualitative, experimental, correlational, observational, computational, but also emerging real-world data-driven approaches such as machine learning) would all greatly benefit the study of bias and group disparities in social psychology. However, their use in a research programme is certainly nonlinear and more dynamic than the fixed sequence depicted by Cesario in his suggested new/rehashed approach.