Computational modelling of social feedback prediction errors on the neural level
Due to the consistent information carried by cue and outcome in the neutral condition, the prediction error for each trial was only computed for the social reward and social punishment conditions and employed as trial-wise parametric modulator on the neural level (66). The prediction error (PE) and expected value (EV) for each trial were estimated based on the Rescorla-Wagner model (66-68) according to
\begin{equation} PE_{t}=R_{t}-EV_{t}\nonumber \\ \end{equation}\begin{equation} EV_{t+1}=EV_{t}+\alpha\times PE_{t}\nonumber \\ \end{equation}
PE reflects the difference between the expectation and the actual outcome, and EV reflects the expectation of receiving a certain feedback on a given trail. R is the actual feedback received, t is the given trial, and \(\alpha\) is the learning rate. The initial EV for social reward anticipation and social punishment anticipation were set to 0.5 and -0.5 respectively. A learning rate α=0.7 was used across participants (see 66, 69). The EV for the next trial is updated based on the EV of the current trial and the prediction error of that trial multiplied by the learning rate. Altough previous studies suggests that a range of learning rates (0.2-0.7) does not affect computation (66, 67, 69-71), additional tests with learning rate = 0.2 were conducted and confirmed the primary findings (see supplementary results ).
For determining the social feedback PE on the neural level EV and PE were included on the first level as parametric modulators corresponding to anticipation and outcome phase respectively. On the first level the GLM models corresponded to the BOLD activation models and the EV and PE were included as parametric modulators. Based on previous studies (66, 69) outcome stage was modeled without taking specific feedback into account to avoid overfitting. Treatment effects on the second level were examined by means of directly comparing the two treatment groups by means of independent t tests.