Individuals with anxiety and depression use atypical decision strategies in an uncertain world
- Shanghai Mental Health Center, School of Medicine, Shanghai Jiao Tong University, China
- School of Psychology, Shanghai Jiao Tong University, China
- School of Psychology, South China Normal University, China
- The Center of Psychosomatic Medicine, Sichuan Provincial Center for Mental Health, Sichuan Provincial People's Hospital, University of Electronic Science and Technology of China, China
- Centre of Centre for Cognitive and Brain Sciences, Institute of Collaborative Innovation, University of Macau, China
- Department of Psychology, University of Arizona, United States
- School of Humanities and Management, Guangdong Medical University, China
- Tianqiao and Chrissy Chen Institute for Translational Research, China
- Shanghai Key Laboratory of Mental Health and Psychological Crisis Intervention, School of Psychology and Cognitive Science, East China Normal University, China
Figures
Figure 1
Schematic diagram of the experimental task in Gagne et al., 2020.
(A) In each trial, participants were presented with two stimuli associated with their potential feedback magnitude. They were instructed to choose one of the two stimuli to receive feedback, but only one stimulus would result in feedback. Participants were required to complete tasks across four experimental contexts. (B) Each run consisted of 90 trials in the stable context and 90 trials in the volatile context. In the stable context, the true environmental probability remains unchanged, while in the volatile context, the probability flips every 20 trials.
Figure 2
Task performance comparison between healthy control participants and patients diagnosed with major depressive disorder (MDD) and generalized anxiety disorder (GAD).
Significance symbols: *p<0.05; **p<0.01; ***p<0.001; n.s.: non-significant. Abbreviations: HC, healthy controls; PAT, patients. (A) Comparison of hit rates for healthy controls and patients in stable and volatile contexts. Error bars represent the standard deviation of the estimated mean across 86 participants. (B) Learning curves for healthy controls and patients throughout the learning process. The dashed line represents the exemplar feedback probability sequence. For runs that do not follow this exemplar sequence (e.g. starting with volatile and then moving to stable conditions), responses were converted to match the exemplar sequence. The learning curves for both groups were then generated by averaging these converted responses across participants within each group. For better visualization, these curves were then smoothed using a Gaussian kernel with a standard deviation of two trials. The blue arrows indicate the apparent deviation between the true feedback probability and the patients’ asymptotic performance. (C) Hit rate differences for healthy controls and patients and their relationship with participants’ symptom severities. Error bars represent the standard deviation of the estimated mean across 54 healthy controls and 32 patients, respectively.
Figure 3 with 3 supplements
Models’ quantitative and qualitative fit to human behavioral data.
Significance symbols: *p<0.05; **p<0.01; ***p<0.001; n.s.: non-significant. Abbreviations: HC, healthy controls; PAT, patients. (A) Relative performance of models compared to the MOS6 model, as measured by the Akaike Information Criterion (AIC), and Bayesian Information Criterion (BIC). Each dot represents a model’s fit for an individual participant, with error bars showing the standard deviation of the estimated mean across 86 participants. (B) Group-level Bayesian model selection as indicated by Protected Exceedance Probability (PXP). (C–E) Models' predicted hit rate (C) hit rate differences (D) and learning curves (E) for healthy controls and patients, respectively. Error bars denote the standard deviation of the estimated mean across 54 healthy controls and 32 patients, respectively.
Figure 3—figure supplement 1
The fit to the human data of different mixture-of-strategies (MOS) variants.
We compared the target MOS6 model (red) with its 13 alternatives in three categories: lesion models generated by removing strategies from the MOS6 (orange), models with components replaced (green), and extensions generated by adding a new strategy (blue). Note that the PF here is the abbreviation of Probability of Feedback, which constructs a policy only depends on the estimated feedback probability, the RD here is the abbreviation of Random strategy, and the probability of choosing each stimulus is 1/2. The target EU, MO, and HA strategies construct a simple decision pool but not too simpler compared to the other alternatives. First, the three ‘lesion’ variants, EU + MO, EU + HA, and MO + HA, all exhibit a lower model fittings performance compared to the MOS6 model, with the EU + HA variant being marginally closer. However, the context-independent EU + HA and its context-dependent version (EU + HA18) provide conflicting interpretations of the behavioral differences between healthy controls and patients, suggesting a potential oversimplification of human behavior. Second, replacing the EU strategy with PF or the MO strategy with RD adversely affects fitting performance. The PF strategy’s failure indicates that a mere linear combination of feedback probability and potential magnitude does not account for human decision-making behavior. The RD strategy’s failure confirms that participants were actively using the MO strategy, rather than making random choices. Lastly, the two extension models do not significantly improve the fitting performance, suggesting that the current MOS6 model is good enough describing human behaviors. There is no need to involve additional components.
Figure 3—figure supplement 2
Hit rates (A) and hit rate differences (B) for all models.
Significance symbols: *p<0.05; **p<0.01; ***p<0.001; n.s.: non-significant.
Figure 3—figure supplement 3
Simulated learning curves for the healthy control (HC) and patient (PAT) groups, each averaged from 100 simulations within the group and were smoothed with a Gaussian kernel (standard deviation of two trials).
Figure 4 with 1 supplement
Parameter analyses of the MOS6 model and simulated behaviors for all three strategies.
Significance symbol conventions are: *p<0.05; **p<0.01; ***p<0.001; n.s.: non-significant. Abbreviations: HC, healthy controls; PAT, patients. (A) The fitted weighting parameters and learning rate of the MOS6 model. The y-axis means averaged preference over different volatile contexts (volatile/stable) and feedback contexts (reward/aversive). indicates the averaged weighting parameters for each participant group. Error bars denote the standard deviation of the estimated mean across 54 healthy controls and 32 patients, respectively. (B) Simulated hit rates for the three decision strategies. Error bars represent the standard deviation across 200 simulations. The 200 simulations were evenly divided between groups using parameters similar to the healthy control group and the patient group. The groups differed only in their strategy preference (differences in ) while all other parameters remained constant. For more simulation details, refer to Materials and methods, Simulation details. (C) The average simulated learning curve for each strategy across 200 simulations, was smoothed with a Gaussian kernel (standard deviation of two trials). (D) Simulated hit rate differences between volatile and stable for the three decision strategies. Error bars represent the standard deviation across 200 simulations. (E) Simulated learning curves for the healthy controls and patients, each averaged from 100 simulations within the group and smoothed with a Gaussian kernel (standard deviation of two trials).
Figure 4—figure supplement 1
Parameter analyses of the MOS22 model.
Significance symbol conventions are: *p<0.05; **p<0.01; ***p<0.001; n.s.: non-significant. Abbreviations: HC, healthy controls; PAT, patients. Error bars denote the standard deviation of the estimated mean. (A) The fitted weighting parameters of the MOS22 across participant groups. (B–D) The fitted log learning rate of the MOS22 model across participant groups (B) volatile contexts (C) and outcome valence (D).
Figure 5 with 1 supplement
Predict participants’ symptom severity (g score) using strategy preferences of the MOS6 model.
Each dot represents one participant. The shaded areas reflect 95% confidence intervals of the regression prediction.
Figure 5—figure supplement 1
Strategy preferences predict participants' general factor score (g score) in the bifactor analysis reported by Gagne et al., 2020.
The y-axis indicates the averaged preference over different volatility levels (volatile and stable) and feedback types (reward and aversive). This average operation is permitted here because the logit of the weight is normally distributed. The shaded areas reflect 95% confidence intervals of the regression prediction.
Figure 6
Reproduction of the two learning rate adaptation effects using the MOS6 model.
Significance symbol conventions are: *p<0.05; **p<0.01; ***p<0.001; n.s.: non-significant. HC represents the healthy-control-like agent; PAT represents the patient-like agent. (A) Simulated learning curves for the healthy controls and patients generated by the MOS6 model. Both curves are averaged over 80 runs of tasks (4 task sequences × 20 experiments) and are smoothed with a Gaussian kernel (standard deviation of two trials). (B) The fitted FLR22 learning rate parameters are for the stable context and the volatile context. Error bars denote the standard deviation across 40 synthesized datasets. (C) Learning rate adaptations, calculated by log volatile learning rate – log stable learning rate, for the healthy control-like agent and for the patient-like agent. Error bars stand for the standard deviation across 20 synthesized datasets.
Figure 7
Parameter and model recovery analyses.
(A) Parameterrecovery for the MOS6 model. (B) Model recovery analysis, showing the performance of models as evaluated by averaged relative Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC), as well as Protected Exceedance Probability (PXP) scores for synthesized data generated from each of the eight models. Darker tiles indicate better fits to the synthesized data.
Author response image 1
Thorough model comparison.
Author response image 2
Author response image 3
Tables
Table 1
Model’s parameters.
| Model | Context-free parameters | Context-dependent parameters |
|---|---|---|
| MOS6 | ||
| MOS22 | ||
| FLR6 | ||
| FLR22 | ||
| RS3 | ||
| RS13 | ||
| PH4 | ||
| PH17 |
Additional files
Download links
A two-part list of links to download the article, or parts of the article, in various formats.
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Individuals with anxiety and depression use atypical decision strategies in an uncertain world
eLife 13:RP93887.
https://doi.org/10.7554/eLife.93887.3
