On the dynamics of the responses in Frydman and Jin (2022): Nullius in verba

Frydman and Jin (2022) ["Efficient coding and risky choice," Quarterly Journal of Economics, 137, 161---213] present a model of efficient coding whereby decision makers are Bayesian learners of a stochastic distribution. The model predicts that decision makers will devote more cognitive resources to---and therefore be more sensitive to--values that appear more frequently. The authors conduct two experiments where subjects make binary choices between a certain amount and a lottery, where the trial-specific values are drawn from a stochastic distribution. While unknown to the subjects, the distribution can be learned over the course of the experiment. The authors conclude that the observations are consistent with efficient coding. However, we note that the authors do not examine observations across trials. When we examine the data from Experiment 1, we do not find evidence that the relationship between sensitivity and frequency increased across trials. When we include specifications that account for the parameters in the previous trial, the treatment interaction estimates are no longer significant. The effects identified by Frydman and Jin (2022) in Experiment 1 are simply a recency bias and not the result of Bayesian learning. We find that subjects in Experiment 2 are less---not more---sensitive to values they encounter more frequently. In summary, we do not find support for the central claims made by the authors. Finally, we describe some unreported details in the preregistration reports of Frydman and Jin (2022). We encourage economists to exercise more skepticism until convinced by the authors' arguments.