Autistic traits relate to reduced reward sensitivity in learning from social point-light displays (PLDs)
Data files
Jan 14, 2025 version files 336.27 KB
-
parameters.csv
19.64 KB
-
raw_choices.csv
309.68 KB
-
README.md
6.94 KB
Abstract
A number of studies have linked autistic traits to difficulties in learning from social (vs. non-social) stimuli. However, these stimuli are often difficult to match on low-level visual properties, which is especially important given the impact of autistic traits on sensory processing. Additionally, studies often fail to account for dissociable aspects of the learning process in the specification of model parameters (learning rates and reward sensitivity). Here, we investigate whether learning deficits in individuals with high autistic traits exhibit deficits when learning from facial point-light displays (PLDs) depicting emotional expressions. Social and non-social stimuli were created from random arrangements of the same number of point-lights and carefully matched on low-level visual properties. Neurotypical participants (N = 63) were assessed using the Autism Spectrum Quotient (AQ) and completed a total of 96 trials in a reinforcement learning task. Although linear multilevel modeling did not indicate learning deficits, preregistered computational modeling using a Rescorla-Wagner framework revealed that higher autistic traits were associated with reduced reward sensitivity in the win domain, demonstrating attenuated response to received social (compared to non-social) feedback during learning. These findings suggest that autistic traits can significantly impact learning from social feedback beyond a general deficit in learning rates.
README: Autistic traits and reinforcement learning
General
This repository contains data and scripts for the RSOS submission "Autistic Traits relate to reduced Reward Sensitivity in Learning from Social Point-Light Displays (PLD’s)". The data was collected to investigate the relationship between autistic traits and learning performance using point-light displays (PLDs) in a reinforcement learning task. Sixty-three participants categorized two-digit numbers while receiving probabilistic feedback. Social feedback was provided via PLDs depicting facial emotions (happiness or anger), while non-social feedback used PLDs forming checkmarks or crosses. Computational modeling (Rescorla-Wagner framework) was used to fit parameters such as learning rates and reward sensitivity to explore differences in social versus non-social learning and their associations with autistic traits, measured via the autism spectrum quotient (AQ). Results showed that higher autistic traits were associated with reduced reward sensitivity. Figure 1 in the supplementary material illustrates a single trial.
Description
The analysis is split in two parts: raw choice data (raw_choices.csv) and parameters of the reinforcement learning model (parameters.csv). Data files and scripts are named accordingly. Scripts have been uploaded to Zenodo for the reported analyses (raw_choices_GLM.R and parameters_glmmTMB.R) and for plotting (raw_choices_plots.R and parameters_plots.R). Figures included in the publication have been uploaded to Zenodo as well.
Codebook for raw_choices.csv
This file contains raw trial-by-trial choice data from participants in the reinforcement learning task.
- ParticipantID: Unique identifier for each participant.
- stim: The presented number stimulus for the trial.
- ParticipantAnswer: The choice made by the participant (
0
for "A",1
for "B"). - high_prob_choice: The option (
0
for "A",1
for "B") associated with a higher reward probability for the given trial. - Reward: Indicates whether a reward was delivered for the trial (
1
for reward,0
for no reward). - CorrectAns_participant: Whether the participant's answer was correct (
1
) or incorrect (0
) based on task contingencies, i.e. whether the participant's choice aligned with the option in high_prob_choice. - BlockType: Type of block in the trial, either "social" (PLDs of human faces) or "nonsocial" (checkmarks and crosses).
- TrialNumber: The sequential number of a trial within its respective block type (social or nonsocial). Each block type was presented twice in an interleaved manner during the experiment.
- RunNumber: The sequence number of the trial within a block.
- BlockNumber: The sequence number of the block within the experiment.
- PredictedChoice: The predicted choice made by the reinforcement learning model for the participant on this trial.
- CorrectPrediction: Whether the model’s predicted choice for the trial matched the correct choice (
1
for correct,0
for incorrect). - AQ_score: Autism Spectrum Quotient score for each participant. Measures autistic traits on a scale (range 0-33) with higher values indicating more autistic traits.
- AQ_group: A categorical variable splitting participants into three groups based on their AQ scores: "Low", "Medium", and "High".
- Med_Split: A binary variable splitting participants into two groups based on the median AQ score ("Low" and "High").
Codebook for parameters.csv
This file contains subject-level parameters derived from the Rescorla-Wagner reinforcement learning model applied to trial-by-trial choice data.
participant: Unique identifier for each participant.
AQ_score: Autism Spectrum Quotient score for each participant. Measures autistic traits on a scale (range 0-33), with higher values indicating more autistic traits.
Gender: Self-reported gender of the participant.
BlockType: Type of block in the trial, either "social" (PLDs of human faces) or "nonsocial" (checkmarks and crosses).
Alpha_Win: Learning rate parameter for reward feedback in win trials (0-1). Reflects how quickly participants update expectations based on positive feedback.
Theta_Win: Exploration parameter for win trials. Higher values indicate more randomness in choice behavior under uncertainty.
Rho_Win: Reward sensitivity parameter for win trials (0-1). Represents the subjective valuation of rewards received.
Alpha_Loss: Learning rate parameter for punishment feedback in loss trials (0-1). Reflects how quickly participants update expectations based on negative feedback.
Theta_Loss: Exploration parameter for loss trials. Higher values indicate more randomness in choice behavior under uncertainty.
Rho_Loss: Sensitivity parameter for loss trials (0-1). Represents the subjective valuation of punishments received.
Neg_LL: Negative log-likelihood of the reinforcement learning model during the optimization process. Lower values indicate a better fit of the model to the participant's data.
BIC: Bayesian Information Criterion for the reinforcement learning model. Lower values indicate a better balance between model fit and complexity.
Code/Software
R is required to run the scripts. The following description gives an overview of the analyses performed in each file:
raw_choices_GLM.R
This script implements a single GLMM to investigate the influence of AQ scores, block type, and trial number on choice correctness. It calculates fixed effect sizes (odds ratios with confidence intervals) and conducts post-hoc comparisons using the emmeans package to analyze interactions and trends. Interaction effects are visualized for further interpretation.
raw_choices_plot.R
This script visualizes trial-by-trial choice data to highlight behavioral patterns. It includes histograms of AQ scores split into groups (low, medium, high), line plots showing learning curves across trials for different AQ groups and block types, violin and box plots comparing correctness across blocks, and scatter plots exploring relationships between AQ scores and choice behavior.
parameters_glmmTMB.R
This script fits Beta regression models using the glmmTMB package to analyze subject-level parameters derived from the reinforcement learning model. The models predict parameters such as Alpha_Win, Rho_Win, Alpha_Loss, and Rho_Loss based on AQ scores and block type. The script also evaluates model assumptions using the DHARMa package, generating diagnostic plots.
parameters_plots.R
This script creates visualizations of the reinforcement learning model parameters. It includes histograms to show parameter distributions by domain (win/loss), violin and box plots comparing parameters across social and nonsocial blocks, and scatter plots examining relationships between AQ scores and parameters.
Methods
Sample
The sample consisted of typically developing (TD) individuals, which were contacted via a participation recruitment platform of the University of Vienna. Participants were required to fulfill the following inclusion criteria: a) age between 18 and 65 years, b) heterosexual orientation, c) proficiency in English, d) no drug or alcohol addiction or regular drug use, e) no psychiatric or neurological condition. As the experiment was part of a larger project, some of the exclusion criteria are related to another task and not directly relevant for this study.
From the total of 74 recruited individuals, three participants were excluded due to missing data for AQ scores and five participants were dropped based on exclusion criteria. Three individuals were excluded from analyses because of missing data for the task, leading to a final sample of N = 63 participants for analysis. Participants received either a financial compensation of 10€ or 4 study credits.
Measures - AQ
To measure autistic traits, a German shortened version of the Autism Spectrum Quotient (AQ-k; Freitag et al., 2007), widely used in research and clinical practice, was used. AQ-k contains 33 items (e.g. “I prefer to do things on my own rather than with others.”) and is suitable for adults and adolescents aged 16 years and above with normal intellectual functioning. For screening purposes of ASD in clinical practice, a cut-off of 17 was proposed (Freitag et al., 2007). In the present study, item responses were scored using a binary system, where endorsement of an autistic trait is scored with one point, while the opposite response is scored with zero, resulting in a maximum score of 33 (Ruzich et al., 2015). In the present sample, we report reliability scores as Cronbach's alpha (α) = .83 and McDonald’s omega (ω) = .85, indicating a good scale reliability.
Social Reinforcement Learning Task
To assess learning, we used a social reinforcement learning task with PLDs as feedback (see Figure 1). Participants were required to categorize randomly generated two-digit numbers (e.g. 99) into arbitrary groups (“A” or “B”) via button press. Importantly, feedback was delivered probabilistically: in 85% of trials, correct responses were followed by rewarding feedback, while in 15% of trials correct responses were followed by non-rewarding feedback. This contingency was chosen to provide an appropriate level of difficulty for learning the underlying associations. Participants were informed that categories were arbitrary with no underlying rule, requiring them to learn via trial-and-error from feedback. The exact contingencies were not disclosed.
After each response, participants received PLD feedback. First, a random pattern of point lights was displayed, which transformed into either social or non-social feedback. In social blocks, point lights formed happy or angry human faces to indicate correct or incorrect responses. In non-social blocks, they formed check marks or crosses. Participants completed two blocks for each condition, each block consisting of six trials. Within each trial, 4 unique two-digit numbers were presented and repeated across subsequent trials, with the presentation order randomized. To avoid carry-over effects, even numbers were assigned to the social blocks, uneven numbers to the non-social blocks. Response options (“A” and “B”) were displayed to the left and right of the screen, with their positions randomly switched between trials to avoid simple motor learning.
To control for order effects, the first block type (social / non-social) was randomly selected, and subsequent blocks alternated between the two types. The dependent variable for the GLMM analysis was the proportion of correct answers, defined as the proportion of trials, in which participants chose the high probability option. We expected participants to perform around chance level (50%) in the first trial and improve in subsequent trials, as they learned the underlying reward contingencies. A German version of the task is available online (Chrome recommended): https://raimund-buehler.github.io/SOCIALRL_PLD/