Task design.

On each trial, each participant received a thermal stimulus lasting 2s from a sequence of intensities. This was followed by a perception (A) or a prediction (B) input screen, where the y-axis indicates the level of perceived/predicted intensity (0-100) centred around participant’s pain threshold, and the x-axis indicates the level of confidence in one’s perception (0-1). The inter-stimulus interval (ISI; black screen) lasted 2.5s (trial example in C). D: Example intensity sequences are plotted in green, participant’s perception and prediction responses are in red and black. E: Participant’s confidence rating for perception (red) and prediction (black) trials.

Participant’s model-naive performance in the task. Violin plots of participant Root Mean Square Error (RMSE) for each condition for A: rating and B: prediction responses as compared with the input.

Expectation weighted models.

Computational models used in the main analysis to capture participants’ pain perception and prediction ratings. Both types of ratings are affected by confidence rating (Ct) on each trial. A) In the Reinforcement Learning model, participant’s pain perception (Pt) is taken to be weighted sum of the current noxious input (Nt) and their current pain expectation (Et). Following the noxious input, participant updates their pain expectation (Et+1). (B) In the Kalman Filter model, a generative model of the environment is assumed (yellow background) - where the mean pain level (xt) evolves according to a Gaussian random walk (volatility v2). The true pain level on each trial (πt) is then drawn from a Gaussian (stochasticity s2). Lastly, the noxious input, Nt, is assumed an imperfect indicator of the true pain level (subjective noise ε2). Inference and prediction steps are depicted in a blue box. Participant’s perceived pain is a weighted sum of expectation about the pain level (mt) and current noxious input (Nt). Following each observation, Nt, participant updates their expectation about the pain level (mt+1).

Confidence scaling factor demonstration.

A-F: For a range of values of the confidence scaling factor C, we simulated a set of typical responses a participant would make for various levels of confidence ratings. The belief about the mean of the sequence is set at 50, while the response noise at 10. The confidence scaling factor C effectively scales the response noise, adding or reducing response uncertainty. G-L: The effect of different levels of parameter C on noise scaling. As C increases the effect of confidence is diminished.

Model comparison for each sequence condition (A-D). The dots indicate the ELPD difference between the winning model (eKF) every other model. The line indicates the standard error (SE) of the difference. The non-winning models’ ELPD differences are annotated with the ratio between the ELPD difference and SE indicating the sigma effect, a significance heuristic.

(A-D): The effect of the confidence scaling factor on noise scaling for each condition. Each coloured line corresponds to one participant, with the black line indicating the mean across all participants. The mean slope for each condition is annotated.