Table 1.
Age, sex, ethnicity, mood and WASI total IQ distributions.
Fig 1.
Study task and longitudinal structure.
A. Go-NoGo task. Modified with permission from (26). Participants are presented with four stimuli for 36 trials each in fully randomized order. After a short ‘wait’ interval, they implement a decision of either to press or not press a button. Subjects discover by trial and error the two possible outcomes for choice of each stimulus (win/nothing or nothing/loss) and across trials learn which decision, ‘Go’ or ‘NoGo’, most often (with probability 0.8) leads to the best outcome. B. Longitudinal study structure, summarizing and illustrating the stages described in Methods.
Fig 2.
Parameter time dependency for the short follow-up study, baseline vs. 6 months later.
A. Go Bias B. (log) Pavlovian bias. Statistically significant correlation is observed for the Go-bias but not for the Pavlovian bias.
Fig 3.
Raw performance in the all conditions.
The middle two quartiles are in solid color, with the ‘avoid loss’ conditions in pink/purple and ‘win’ in gold/yellow. Just non-overlapping notches represent p = 0.05 for the uncorrected difference between two medians. One star denotes p <0.05 corrected for 8 comparisons; Five stars denote pcor < 1e-10. Long follow-up is shown next to baseline. A. Performance weighted towards early trials, the weighing decreasing linearly to 0 for the middle of the task. No-Go to win (NG2W) showed median success rate slightly below chance, in line with the literature. B. Late trials. Performance reaches maximum for at least a quarter of participants in both appetitive conditions, but a quarter (long follow-up) or more (baseline) participants still perform below chance in NG2W. Long follow-up shows better performance than baseline in all except the easy Go-to-Win (G2W) conditions.
Fig 4.
Comparison of fit quality for 6-parameter (left three) and 7-parameter (right three) models, using the baseline data.
The ‘valenced learning’ (leftmost) model performs best, i.e. had the lowest integrated Bayesian Information Criterion score. Adding forgetting parameters (right three bars) worsened the fits due to the complexity penalties involved.
Fig 5.
Model comparison based on Mean Prediction probability per trial (Ppt) in the long follow-up sample, showing that the difference in Ppt between the two best models is similar if one uses model-fitting vs. out-of-sample based methods A. ΔPpt estimated through a model fit measure, namely mean integrated likelihood per trial, N2 = 556. Both models have mean Ppt about 0.64. B. ΔPpt estimated by out-of-sample prediction of the 48th and 96th trials for each participant on a test subsample of N = 255. This out-of-sample comparison is more variable, but the resampling-based 95% confidence interval (CI) of the median difference is -0.0012 to 0.0026, consistent with A. If it were desirable to further reduce this CI, the estimate could be averaged over rotated out-of-sample trials, at the very considerable computational cost of re-estimating the entire model fit for each left-out sample.
Table 2.
Spearman correlation coefficients between the peak posterior parameter estimates at the 0.05 level, corrected for 6*5/2 comparisons, for the baseline sample (N = 817).
Correlation values below the diagonal, corrected p-values above.
Fig 6.
Stability in descriptive estimates of Pavlovian bias, assessed by the interaction between the fractions of total correct answers in the four conditions, ((G2W-NG2W)+(NG2AL-G2AL))/2.
A. Stability assessed by baseline vs. long follow-up estimates. A positive correlation is detectable, rho~0.15, p = 5.4e-4. B. Difference in performance between the appetitive (two left) and aversive (two right) conditions. The horizontal lines show the median appetitive bias at baseline (G2W-NG2W; first violin plot; salmon) vs. long follow-up (second plot, in green). This significant difference (p = 0.0027) drives an overall reduction in the estimate of Pavlovian bias (p = 0.0019). The aversive context, NG2AL-G2AL, shows no significant change on its own (blue and mauve; p = 0.35). White boxes are interquartile ranges.
Fig 7.
Parameter time dependency, long follow-up study, baseline vs. 18 months (mean) later.
A. Pavlovian Bias. The most prominent feature was a reduction in the group mean. B. Motivational exchange rate (log beta). Here there is little shift in the modal tendency. Note that log-units are used and the contours are estimated by axis-aligned, bivariate normal kernel density estimation.
Fig 8.
The distribution of model fit over the population is bimodal and best described by four Gaussian components.
A. Kernel density contours are shown, while the inset plots peaks of the landscape in 3D. B. 4-component mixture-of-Gaussians fitted to this joint distribution of integrated likelihoods, with participants clustered according to the Gaussian they are most likely to belong to. As can also be seen in the 3D rendering. Apart from the prominent peaks (cluster3, green, Nc3 = 176 participants, and 4, blue, Nc4 = 171) there are somewhat less prominent concentrations (1, black, Nc1 = 94 and 2, red, Nc2 = 113).