Explaining human sampling rates across different decision domains

Abstract

Undersampling biases are common in the optimal stopping literature, especially for economic full choice problems. Among these kinds of number-based studies, the moments of the distribution of values that generates the options (i.e., the generating distribution) seem to influence participants’ sampling rate. However, a recent study reported an oversampling bias on a different kind of optimal stopping task: where participants chose potential romantic partners from images of faces (Furl et al., 2019). The authors hypothesised that this oversampling bias might be specific to mate choice. We preregistered this hypothesis and so, here, we test whether sampling rates across different image-based decision-making domains a) reflect different over- or undersampling biases, or b) depend on the moments of the generating distributions (as shown for economic number-based tasks). In two studies (N = 208 and N = 96), we found evidence against the preregistered hypothesis. Participants oversampled to the same degree across domains (compared to a Bayesian ideal observer model), while their sampling rates depended on the generating distribution mean and skewness in a similar way as number-based paradigms. Moreover, optimality model sampling to some extent depended on the the skewness of the generating distribution in a similar way to participants. We conclude that oversampling is not instigated by the mate choice domain and that sampling rate in image-based paradigms, like number-based paradigms, depends on the generating distribution.