Human aging alters Bayesian social inference about others’ changing intentions

Abstract

Decoding others’ intentions accurately in order to adapt one’s own behavior is pivotal throughout life. Yet, it is a process that is imbued with uncertainty since others’ intentions are not directly observable and may change over time. In this study, we asked the question of how younger and older adults deal with uncertainty in dynamic social environments. We used an advice-taking paradigm together with biologically plausible hierarchical Bayesian modelling to characterize effects and mechanisms of aging on learning about others’ time-varying intentions. We observed age differences when comparing learning on two levels of social uncertainty: the fidelity of the adviser and the stability of intentions. We found that, prior to having any experience with the adviser, older adults expected the adviser to change his/her intentions more frequently. They also showed higher confidence in such beliefs and were less willing to change their beliefs over the course of the experiment. This led them to update their predictions about observable outcomes (i.e., advice correctness) more quickly. Potentially indicative of stereotype effects, we also observed that older advisers were perceived as more volatile, but at the same time, more faithful than younger advisers. Together these findings offer new insights into the behavioral and algorithmic mechanisms underlying adult age differences in response to social uncertainty, putatively driven by aging-related changes in neuromodulation to be tested in future studies.

A Social Inference Model of Idealization and Devaluation

“…To study idealization and devaluation, we implement a probabilistic model for how other peoples' internal states combine with external circumstances to give rise to their behaviour. In keeping with contemporary approaches (e.g., Barnby et al, 2020;Moutoussis, Fearon, et al, 2014;Moutoussis, Trujillo-Barreto, et al, 2014;Reiter et al, 2019;Wellstein et al, 2020), we suggest that humans use such a model to infer others' internal states in social interactions. We focus on representations of others, however as we address in the Discussion, the formal principles are easily generalized to beliefs about the self.…”

A social inference model of idealization and devaluation.

“…In this article, we model "splitting," or dichotomous thinking, from a Bayesian perspective. In keeping with previous approaches (e.g., Ajzen & Fishbein, 1975;Diaconescu et al, 2020;Moutoussis, Fearon, et al, 2014;Reiter et al, 2019;Siegel et al, 2018), a subject learns about others' dispositions by accruing information about their behavior across time. A novel feature is the addition of latent, split representations of others' dispositions as either extremely good or extremely bad, whose likelihood is increased following "good" or "bad" observations, respectively.…”

“…We formalize splitting within a probabilistic model wherein subjects infer dispositional (internal) and situational (external) causes of another person’s behavior. Our model follows myriad theories proposing that our brain evaluates probabilistic hypotheses about the hidden causes of its inputs by approximating Bayesian inference (Chater & Oaksford, 2008; Dunsmoor et al, 2015; Friston et al, 2016, 2017; Gershman, 2017; Gershman et al, 2013, 2015; Gershman & Blei, 2012; Gershman & Niv, 2012; Glimcher, 2004; Noorani & Carpenter, 2016; Tomov et al, 2018), which have previously been applied to social inference (e.g., Ajzen & Fishbein, 1975; Barnby et al, 2020; Diaconescu et al, 2020; Moutoussis, Fearon, et al, 2014; Moutoussis, Trujillo-Barreto, et al, 2014; Redcay & Schilbach, 2019; Reiter et al, 2019; Wellstein et al, 2020). Formally, Bayes’ theorem states how new information can be optimally combined with the prior knowledge to update belief in particular hypothesis as follows:…”