Table 1.
Participant characteristics (total N = 572): Mean (SD), [N] if missing data, unless otherwise stated.
Fig 1.
Task presentation and pooled task behavior.
(A) An example of several consecutive trials—on each trial, participants have to choose between two stimuli, presented pseudorandomly in two of the four possible locations. Feedback is received in the form of a smiling green face (positive) or a sad red face (negative) and is probabilistic, meaning that some is “misleading” (e.g., trial 3). Win-stay trials are those in which individuals repeat their stimuli choice following positive feedback (e.g., trials 2 and 3), and lose-shift trials are those in which individual change their stimuli choice following negative feedback (e.g., trials 4 and 5). (B) The structure of the task—the first stimuli chosen by each participant is correct in the acquisition phase (trials 1–40; here: yellow). Feedback was given with an 80:20 reward/punishment ratio; green blocks indicate reward and red blocks indicate punishment. In the reversal phase (trials 41–80), the true correct stimulus is reversed (here: blue) as is the contingency schedule. (C) Overall trial-by-trial behavior—All participants’ data, sorted by performance, with average performance overlaid (black line) regardless of diagnosis or age group. Compare to (B) to see how task structure is experienced in practice (see S1 Data).
Fig 2.
(A) Trial-by-trial data for each age group with diagnostic group averages overlaid. More evidence of task understanding in adults, as indicated by more correct task behavior and steeper shifts at reversal in comparison to children. (B) Task accuracy was greater (1) in the acquisition phase compared to the reversal phase, (2) in older age groups compared to younger, and (3) in TD individuals compared to ASD individuals. (C-E) Linear mixed-effects models showed a main effect of diagnosis for all three task performance measures (perseverative errors, win-staying, lose-shifting) and a main effect of age for win-staying (D) and lose-shifting (E) but not perseverative errors (C). For win-staying, a diagnosis × age group interaction was also found. Post hoc tests revealed ASD adolescents showed significantly reduced win-staying compared with TD adolescents (D), ***p < .001 (see S1 Data). ASD, autism spectrum disorder; TD, typical development.
Fig 3.
Model comparisons, validations, and parameters.
(A) Evidence (model weights) for models within each diagnostic and age group. Very similar patterns are observed for TD and ASD groups; winning models for children, adolescents, and adults are the CU, R-P, and EWA-DL, respectively. (B) One-step-ahead posterior predictions for each age and diagnostic group according to winning models. Colored lines indicate diagnostic-group-averaged trial-by-trial task behavior; shaded areas indicate 95% HDI of the one-step-ahead simulation using the entire posterior distribution. Compare with actual task data in Fig 2A. Posterior predictive accuracies are also indicated on each plot (ASD: red; TD: blue). (C) Model parameter comparisons. Within each winning model and thus age group, parameter estimates were compared between diagnostic groups: (1) ASD children showed a significantly higher learning rate (η) than TD children, in which simulations showed the optimal learning rate to be 0.18; (2) ASD adolescents showed a significantly lower reward learning rate than TD adolescents, but no difference between punishment learning rates was observed; (3) ASD adults showed significantly lower φ than TD adults, the optimal value was shown to be 0.85 in simulations, and ASD adults also showed significantly greater experience decay (ρ) than TD adults, suggesting great perseveration. (D) Learning rate simulations showing optimal learning rates for each model (Counterfactual update, compare to Fig 3C Children; Rew-Pun, compare to Fig 3C Adolescents—Learning rate; EWA, Experience-weighted attraction-dynamic learning rate, compare to Fig 3C Adults—Inverse learning rate). ***p < .001, **p < .01, *p < .05; Δ indicates group mean (see S1 Data). ASD, autism spectrum disorder; CU, counterfactual update; d, Cohen’s d model; EWA-DL, experience-weighted attraction–dynamic learning rate model; HDI, highest density interval; R-P, reward-punishment model; Rew-Pun, reward-punishment; RL, reinforcement learning; RW, Rescorla-Wagner; TD, typical development.
Table 2.
Model parameters for each age and diagnosis group’s winning model and within age-group comparisons.
Fig 4.
Symptomatology correlations in ASD.
(A) In ASD children, perseverative errors were significantly correlated with anxiety (r72 = 0.34, p = .0040). In ASD adults, (B) perseverative errors were significantly correlated with ADI-R RRB (r116 = 0.29, p = .0013). (C) Perseverative errors were further significantly positively related to parent-reported ADHD Hyperactivity/Impulsivity (r94 = 0.32, p = .0017). Win-staying was significantly negatively related to (D) ADI-R RRB (r116 = −0.31, p = .0007) and (E) RBS-R Ritualistic-Sameness (r91 = −0.30, p = .0004). In ASD adults, experience decay (ρ) was significantly positively associated with (E) RRB (ADI-R RRB r116 = 0.28, p = .0022) as was (F, G) value sensitivity (β; ADI-R RRB r116 = −0.29, p = .0019; RBS-R r91 = −0.30, p = .0040). (H, I) Value sensitivity (β) was also significantly negatively correlated with parent-reported ADHD symptomatology (ADHD hyperactivity/impulsivity r116 = −0.37, p = .0003; ADHD inattention r116 = −0.30, p = .0037). ADHD, attention-deficit hyperactivity disorder; ADI-R, Autism Diagnostic Interview-Revised; ASD, autism spectrum disorder; RBS-R, Repetitive Behavior Scale-Revised; RRB, restricted, repetitive behavior (see S1 Data).