Impaired adaptation of learning to contingency volatility in internalizing psychopathology

  1. Christopher Gagne
  2. Ondrej Zika
  3. Peter Dayan
  4. Sonia J Bishop  Is a corresponding author
  1. Department of Psychology, UC Berkeley, United States
  2. Max Planck Institute for Biological Cybernetics, Germany
  3. Max Planck Institute for Human Development, Germany
  4. University of Tübingen, Germany
  5. Wellcome Centre for Integrative Neuroimaging, University of Oxford, FMRIB, John Radcliffe Hospital, United Kingdom
  6. Helen Wills Neuroscience Institute, UC Berkeley, United States
17 figures, 5 tables and 1 additional file

Figures

Figure 1 with 2 supplements
Bifactor analysis of internalizing symptoms.

(a-b) Bifactor analysis of item-level scores from the STAI, BDI, MASQ, PSWQ, CESD, and EPQ-N (128 items in total) revealed a general ‘negative affect’ factor (xaxis) and two specific factors: one …

Figure 1—figure supplement 1
Scree plot for the eigenvalue decomposition of the covariance matrix of individual items from the battery of internalizing symptom measures.

Bifactor analysis was applied to the item-level scores from the STAI, BDI, MASQ, PSWQ, CESD, and EPQ-N (128 items in total) for participants in experiment 1. Prior to the estimation of the bifactor …

Figure 1—figure supplement 2
Correlation matrices for internalizing questionnaire scales and latent factors from the bifactor analysis.

(a) We calculated summary scores for each participant for each of the standardized questionnaires we administered. All participants who completed the full set of questionnaires were included: this …

Task.

(a) On each trial, participants chose between two shapes. One of the two shapes led to receipt of shock or reward on each trial, the nature of the outcome depending on the version of the task. The …

Cross-group results from experiment 1 for effects of block type (volatile, stable), task version (reward, aversive), and relative outcome value (good, bad) on learning rate (n = 86).

(a) This panel shows the posterior means along with the 95% highest posterior density intervals (HDI) for the group means (μs) for each learning rate component (i.e. for baseline learning rate and …

Figure 4 with 1 supplement
Experiment 1: Effect of general factor scores on learning rate (in-lab sample, n = 86).

Panel (a) shows posterior means and 95% highest posterior density intervals (HDI) for the effect of general factor scores (βg) on each of the learning rate components. General factor scores credibly …

Figure 4—figure supplement 1
Effect of depression-specific and anxiety-specific factors on learning rate and its components (data from experiment 1).

Panel (a) and panel (b) show the effects of depression-specific factor scores and anxiety-specific factor scores, respectively, on each of the learning rate components (e.g. αvolatilestable). The 95%-HDI’s for …

Figure 5 with 2 supplements
Experiment 2: Effect of the general factor scores on learning rate (online sample, n = 147).

Panel (a) shows posterior means and 95% highest posterior density intervals (HDI) for the effect of general factor scores (βg) on each of the learning rate components. Replicating findings from …

Figure 5—figure supplement 1
Cross-group results from experiment 2 for effects of block type (volatile, stable), task version (reward, aversive), and relative outcome value (good, bad) on learning rate.

(a) This panel shows the posterior means along with the 95% highest posterior density intervals (HDI) for the group means (μs) for each learning rate component (i.e. for baseline learning rate and …

Figure 5—figure supplement 2
Effect of depression-specific and anxiety-specific factors on learning rate and its components (data from experiment 2).

Panel (a) and panel (b) show the effects of depression-specific factor scores and anxiety-specific factor scores, respectively, on each of the learning rate components (e.g. αvolatilestable). The 95%-HDI’s for …

Appendix 4—figure 1
Recovery of individual-level learning rate parameters.

Posterior mean parameter estimates for participants from experiment 1 were used to simulate new choice data from the winning model (#11). The model was then re-fit to each of these simulated …

Appendix 4—figure 2
Recovery of other model parameters.

Posterior mean parameter estimates for participants from experiment 1 were used to simulate new choice data from the winning model (#11). As in Appendix 4—figure 1, we show the results of parameter …

Appendix 4—figure 3
Variability of population-level learning rate parameters across simulated datasets.

The robustness of the estimates for the population-level parameters (μ,  βg, βa, βd) was explored by examining the variability in parameter values across the 10 simulated datasets (blue data points). …

Appendix 4—figure 4
Variability of population-level parameters across simulated datasets for other parameters.

The robustness of the estimates for the population-level parameters (μ, βg, βa, βd) was explored by examining the variability in parameter values across the 10 simulated datasets (blue data points). …

Appendix 4—figure 5
Recovery of individual-level learning rate parameters in the thee-way interaction model.

In experiment 1, we additionally fit a model that included the three-way interaction of block type (volatile, stable), relative outcome value (good, bad), and task version (reward, aversive) for …

Appendix 5—figure 1
Comparison of task performance between experiment 1 and experiment 2.

The four panels depict the performance of participants in each block (stable left column; volatile right column) and in each task (reward top row; punishment bottom row). Data from experiment 1 is …

Appendix 6—figure 1
Learning rate parameters for experiment 1 data as estimated using alternate population-level parameters for specific effects of anxiety and depression.

Two alternative models were fit to the behavioral data from experiment 1, in addition to the main bifactor model. For the first alternative model, population-level parameters entered comprised …

Appendix 6—figure 2
Learning rate parameters for experiment 2 data as estimated using alternate population-level parameters for specific effects of anxiety and depression.

In addition to the main bifactor model, an additional alternative model was also fit to the behavioral data from experiment 2. In this model (the second alternate model described in Appendix …

Appendix 7—figure 1
Comparison of actual and model generated numbers of switch trials in experiment 1.

For each participant, we calculated the number of trials on which they switched choice of shape. As described under parameter recovery, each participant’s posterior means for each of model #11 …

Appendix 7—figure 2
Comparison of actual and model generated numbers of switch trials in experiment 2.

For each participant, we calculated the number of trials on which they switched choice of shape. As described under parameter recovery, each participant’s posterior means for each of model #11 …

Author response image 1
Information captured by top five PCs: We conducted PCA on the item-level questionnaire responses from experiment 1 (n=86).

Scores on the first PC correlated highly with general factor scores (r=0.9). Scores on the second PC correlated strongly positively with PSWQ (r=0.59), and moderately negatively with MASQAD …

Author response image 2

Tables

Appendix 1—table 1
Basic demographic details for participants in Experiment 1.
Participant recruitment groupMajor Depressive Disorder (MDD)Generalized Anxiety Disorder (GAD)Healthy ControlsUnselected Community Sample
Participants
Total N
(N for reward task, N for aversive task)
20
(19, 17)
12
(10, 11)
24
(22, 19)
30
(29, 30)
Female10
(10, 8)
11
(9, 10)
16
(14, 13)
13
(12, 13)
Age mean ± sd
(for reward task, for aversive task)
31 ± 10
(31 ± 10, 28 ± 10)
32 ± 9
(31 ± 9, 32 ± 9)
27 ± 6
(27 ± 6, 28 ± 6)
27 ± 5
(27 ± 5, 27 ± 5)
STAI mean ± sd
(for reward task, for aversive task)
59 ± 6
(59 ± 7, 58 ± 6)
58 ± 9
(60 ± 9, 57 ± 9)
40 ± 12
(41 ± 12, 41 ± 12)
36 ± 12
(36 ± 11, 36 ± 12)
BDI mean ± sd
(for reward task, for aversive task)
24 ± 9
(25 ± 9, 23 ± 9)
20 ± 11
(22 ± 10, 20 ± 11)
7 ± 7
(7 ± 7, 7 ± 8)
6 ± 8
(5 ± 7, 6 ± 8)
MASQ-AD mean ± sd
(for reward task, for aversive task)
80 ± 10
(81 ± 10, 79 ± 10)
74 ± 16
(75 ± 17, 73 ± 16)
55 ± 18
(56 ± 18, 56 ± 19)
50 ± 20
(48 ± 18, 50 ± 20)
MASQ-AA mean ± sd
(for reward task, for aversive task)
28 ± 7
(28 ± 7, 28 ± 7)
33 ± 10
(34 ± 11, 34 ± 10)
21 ± 4
(21 ± 4, 20 ± 2)
22 ± 6
(22 ± 6, 22 ± 6)
PSWQ mean ± sd
(for reward task, for aversive task)
62 ± 14
(61 ± 14, 60 ± 14)
76 ± 9
(75 ± 9, 75 ± 9)
52 ± 13
(54 ± 12, 51 ± 14)
42 ± 15
(42 ± 15, 42 ± 15)
CESD mean ± sd
(for reward task, for aversive task)
30 ± 9
(30 ± 9, 28 ± 8)
30 ± 14
(32 ± 14, 30 ± 14)
12 ± 8
(12 ± 8, 11 ± 8)
10 ± 11
(9 ± 10, 10 ± 11)
EPQ-N mean ± sd
(for reward task, for aversive task)
18 ± 3
(18 ± 3, 17 ± 3)
19 ± 4
(19 ± 3, 18 ± 4)
10 ± 6
(11 ± 6, 10 ± 6)
10 ± 6
(10 ± 6, 10 ± 6)
General Factor mean ± sd
(for reward task, for aversive task)
1.1 ± 0.8
(1.1 ± 0.9, 1.0 ± 0.8)
1.3 ± 1.0
(1.5 ± 1.0, 1.4 ± 1.0)
−0.3 ± 0.8
(−0.2 ± 0.8,–0.4 ± 0.7)
−0.2 ± 0.8
(−0.3 ± 0.7,–0.2 ± 0.8)
Depression-Specific Factor mean ± sd
(for reward task, for aversive task)
0.8 ± 1.0
(0.9 ± 1.0, 0.8 ± 1.1)
−0.0 ± 0.8
(−0.1 ± 0.8,–0.1 ± 0.7)
0.1 ± 1.2
(0.1 ± 1.2, 0.3 ± 1.2)
−0.2 ± 0.9
(−0.2 ± 0.9,–0.2 ± 0.9)
Anxiety-Specific Factor mean ± sd
(for reward task, for aversive task)
−0.5 ± 1.1
(−0.5 ± 1.1,–0.6 ± 1.2)
0.8 ± 0.9
(0.5 ± 0.8, 0.7 ± 0.9)
−0.2 ± 0.9
(−0.1 ± 0.8,–0.1 ± 0.9)
−0.3 ± 1.1
(−0.3 ± 1.1,–0.3 ± 1.1)
  1. STAI = Spielberger State-Trait Anxiety Inventory (form Y; Spielberger et al., 1983) BDI = Beck Depression Inventory (Beck et al., 1961) MASQ-AD/MASQ AA=anhedonic depression and anxious arousal subscales for the Mood and Anxiety Symptoms Questionnaire (Clark and Watson, 1995; Watson and Clark, 1991) PSWQ = Penn State Worry Questionnaire (Meyer et al., 1990) CESD = Center for Epidemiologic Studies Depression Scale (Radloff, 1977) EPQ-N = the Neuroticism subscale for the 80-item Eysenck Personality Questionnaire (Eysenck and Eysenck, 1975). The two healthy control participants whose data were excluded from both tasks are omitted from this table.

Appendix 1—table 2
Basic demographic details of participants who provided questionnaire data for the confirmatory factor analysis.
Number of participants (Total N)199
Female (N)120
Age (mean ± sd)21 ± 4
STAI (mean ± sd)44 ± 9
BDI (mean ± sd)7 ± 6
MASQ-AD (mean ± sd)54 ± 15
MASQ-AA (mean ± sd)24 ± 8
PSWQ (mean ± sd)57 ± 13
CESD (mean ± sd)24 ± 8
EPQ-N (mean ± sd)6 ± 4
General Factor (mean ± sd)−0.1 ± 1.0
Depression-Specific Factor (mean ± sd)0.4 ± 1.0
Anxiety-Specific Factor (mean ± sd)−0.1 ± 1.0
  1. STAI = Spielberger State-Trait Anxiety Inventory (form Y; Spielberger et al., 1983) BDI = Beck Depression Inventory (Beck et al., 1961) MASQ-AD/MASQ AA=anhedonic depression and anxious arousal subscales for the Mood and Anxiety Symptoms Questionnaire (Clark and Watson, 1995; Watson and Clark, 1991) PSWQ = Penn State Worry Questionnaire (Meyer et al., 1990) CESD = Center for Epidemiologic Studies Depression Scale (Radloff, 1977) EPQ-N = the Neuroticism subscale for the 80-item Eysenck Personality Questionnaire (Eysenck and Eysenck, 1975).

Appendix 1—table 3
Basic demographic details for participants in Experiment 2.
Participants (Total N)147
Female (N)64
AgeNot recorded; required to be 18 years or older.
STAI (mean ± sd)43 ± 13
BDI (mean ± sd)11 ± 12
MASQ-AD (mean ± sd)63 ± 18
MASQ-AA (mean ± sd)23 ± 8
General Factor (mean ± sd)−0.1 ± 0.9
Depression-Specific Factor (mean ± sd)−0.2 ± 0.9
Anxiety-Specific Factor (mean ± sd)0.1 ± 1.0
  1. STAI = Spielberger State-Trait Anxiety Inventory (form Y; Spielberger et al., 1983); BDI = Beck Depression Inventory (Beck et al., 1961) MASQ-AD/MASQ AA=anhedonic depression and anxious arousal subscales for the Mood and Anxiety Symptoms Questionnaire (Clark and Watson, 1995; Watson and Clark, 1991).

Appendix 3—table 1
Model comparison table for Experiment 1.

Thirteen models were fit to participants’ choice data from experiment 1. Models were fit hierarchically and compared using leave-one-out cross validation error approximated by Pareto smoothed …

Model NumberParameters# of Parameter
Components
PSIS-LOO
Model #1α, γ,ω11227,801
Model #2α, λ,ω1226,164
Model #3αgb, λ,ω1525,550
Model #4α, λgb,ωgb1826,042
Model #5αgb, λgb,ωgb2125,462
Model #6αgb, λgb,ωgb2425,486
Model #7αgb, λgb,ωgb, rra only2325,154
Model #8αgb, λgb,ωgb, rra only, ϵra only2525,185
Model #9αgb, λgb,ωgb, rra only, δ2725,377
Model #10αgb, λgb,ωgb, rra only, δ2725,325
Model #11 **αgb, λgb,ωvgb,rra only, ωkra only, ηbaseline2625,037
Model #12αgb, λgb,ωvgb, rra only, ωkra only, ηbaseline, δ3225,216
Model #13αgb, λgb,ωvgb, rra only, ωkra only, ηbaseline, δ3225,181
  1. 1:Unless otherwise stated, each parameter is divided into four parameter components: a shared baseline parameter across blocks and tasks, and differences in the parameter between stable and volatile blocks (volatile-stable), between different task versions (reward-aversive) and an interaction of those differences (reward-aversive)x(volatile-stable).

    gb: For each parameter with this superscript, three additional parameter components were added for the relative value of previous outcome (good-bad) and the interactions of relative outcome value with block type (volatile-stable)x(good-bad) and task version (reward-aversive)x(good-bad).

  2. ra only: For each parameter with this superscript, only differences in the parameters between the reward and aversive task versions (reward-aversive) were included.

    baseline: For each parameter with this superscript, only one single baseline parameter was used, across both task versions and volatile and stable blocks.

  3. **Indicates best fitting model.

Appendix 3—table 2
Model comparison table for Experiment 2.

The same 13 models fit to participants’ choice data from experiment 1 were also fit to participants’ choice data from experiment 2. Models were fit hierarchically and compared using leave-one-out …

Model NumberParameters# of Parameter
Components
PSIS-LOO
Model #1α, γ,ω11257,520
Model #2α, λ,ω1254,002
Model #3αgb, λ,ω1552,918
Model #4α, λgb,ωgb1853,755
Model #5αgb, λgb,ωgb2152,758
Model #6αgb, λgb,ωgb2452,769
Model #7αgb, λgb,ωgb, rgl only2352,139
Model #8αgb, λgb,ωgb, rgl only, ϵgl only2552,136
Model #9αgb, λgb,ωgb, rgl only, δ2752,083
Model #10αgb, λgb,ωgb, rgl only, δ2752,169
Model #11αgb, λgb,ωvgb,rgl only, ωkgl only, ηbaseline2652,048
Model #12αgb, λgb,ωvgb, rgl only, ωkgl only, ηbaseline, δ3252,005
Model #13αgb, λgb,ωvgb, rgl only, ωkgl only, ηbaseline, δ3252,084
  1. 1:Unless otherwise stated, each parameter is divided into four parameter components: a shared baseline parameter across blocks and tasks, and differences in the parameter between stable and volatile blocks (volatile-stable), between different task versions (gain-loss) and an interaction of those differences (gain-loss)x(volatile-stable).

    gb: For each parameter with this superscript, three additional parameter components were added for the relative value of previous outcome (good-bad) and the interactions of this difference with block type (volatile-stable)x(good-bad), and task version (gain-loss)x(good-bad). gl only: For each parameter with this superscript, only differences in the parameters between the reward gain and reward loss task versions (gain-loss) were included.

  2. baseline: For each parameter with this superscript, only one single baseline parameter was used, across both task versions and volatile and stable blocks.

Additional files

Download links