Fig 1.
(A) The trial structure: On each trial, participants are presented with two pieces of information: a social cue (a recording of an adviser holding up one of the colors), a non-social cue (a pie chart) indicating true winning probabilities of the two options. (B) Probability of adviser’s fidelity (helpfulness) across the task.
Fig 2.
Graphical representation of the mean-reverting 3-level HGF.
The highlighted parameters are free parameters. m3 is volatility equilibrium point towards which the agent slowly drifts irrespective of empirical observations. μ3 and μ2 are the priors about advice volatility and fidelity, respectively, that the agent has before starting the task. κ2 is the phasic learning rate which regulates how much the inferences about the volatility (at the third level) modulate learning about adviser’s fidelity (at the second level). ω2 is the baseline leaning rate (evolution rate), which regulates the learning rate at the second level independent from the estimates at the third level. ζ parameter captures how much participants rely on the social cue relative to the non-social cue. Finally, ν captures intrinsic noise at the stage of decision-making that is unrelated to inferred volatility of the adviser.
Table 1.
Priors on free model parameters.
Prior means and their respective variances are denoted in brackets, and for some parameters followed by an upper bound: (mean, variance), upper bound. See Fig 2 for a detailed explanation of each parameter.
Fig 3.
Test-retest reliability of behavioral measures.
(A) Total accuracy. (B) Total advice taking. (C) Advice taking during the stable phase. (D) Advice taking during the volatile phase. (E) the probability of staying with the same advice taking strategy (trusting advice or not) after a win in a previous trial (F) the probability of changing the advice taking strategy (trusting advice or not) after a loss in a previous trial. ICC(A,1) and Pearson’s correlation coefficients are shown above each panel. The square brackets indicate 95% confidence intervals.
Fig 4.
Test-retest reliability of computational measures.
(A) Prior expectations about the adviser’s fidelity before starting the task (B) Prior expectations about the adviser’s volatility before starting the task. (C) Volatility equilibrium point (D) Phasic learning rate about the adviser’s fidelity (E) Baseline learning rate about the adviser’s fidelity (F). The relative weighing of the advice compared to the non-social cue. (G) Decision noise. ICC(A,1) and Pearson’s correlation coefficients are shown above each panel. The square brackets indicate 95% confidence intervals.
Fig 5.
Parameter recovery and test-retest reliability.
(A) Prior expectations about adviser’s fidelity before starting the task (B) Prior expectations about the adviser’s volatility before starting the task. (C) Volatility equilibrium point (D) Phasic learning rate about the adviser’s fidelity (E) Baseline learning rate about the adviser’s fidelity (F). The relative weighing of the advice compared to the non-social cue. (G) Decision noise. ICC(A,1) and Pearson’s correlation coefficients are shown above each panel. The square brackets indicate 95% confidence intervals. (H) Comparison of parameter recovery vs. test-retest reliability for each model parameter. Note, (A-G) results are based on one particular random number generator seed (used to generate synthetic data), while (H) shows results averaged across 20 simulations with different seeds. The error bars denote standard error.
Fig 6.
Practice effects in the behavioral measures.
(A) Total accuracy. (B) Total advice taking. (C) Advice taking during the stable phase. (D) Advice taking during the volatile phase. (E) the probability of staying with the same advice taking strategy (trusting advice or not) after a win in a previous trial (F) the probability of changing the advice taking strategy (trusting advice or not) after a loss in a previous trial. Red crosses indicate outliers. p-values and Bayesian factors for the null hypothesis (BF01) of paired t-tests are presented above each plot.
Fig 7.
Practice effects in the computational measures.
(A) Prior expectations about adviser’s fidelity before starting the task (B) Prior expectations about the adviser’s volatility before starting the task. (C) Volatility equilibrium point (D) Phasic learning rate about the adviser’s fidelity (E) Baseline learning rate about the adviser’s fidelity (F). The relative weighing of the advice compared to the non-social cue. (G) Decision noise. Red crosses indicate outliers. p-values and Bayesian factors for the null hypothesis (BF01) of paired t-tests are presented above each plot.
Fig 8.
Correlation between the changes in behavioral and computational measures.
(A-C) Prior expectations about adviser’s fidelity before starting the task vs. total advice taking (A), vs. advice taking during the stable phase (B), and vs. win-stay frequency (C). (D-F) The relative weighing of the advice compared to the non-social cue vs. total advice taking (D), vs. advice taking during the stable phase (E), and vs. win-stay frequency (F). Pearson’s correlation coefficients and corresponding p-values are presented above each panel.
Fig 9.
Face validity: The task and the model.
Reported adviser’s fidelity vs actual fidelity during test (A) and retest (C) at different points in the task. Model-derived estimates of adviser’s fidelity () broken down according to the explicitly reported fidelity at the corresponding points in the task during test (B) and retest (D).