Agency and reward across development and in autism: A free-choice paradigm

Our ability to perform voluntary actions and make choices is shaped by the motivation from control over the resulting effects (agency) and from positive outcomes (reward). The underlying action-outcome binding mechanisms rely on sensorimotor abilities that specialise through child development and undergo different trajectories in autism. The study aimed at disentangling the role of agency and reward in driving action selection of autistic and non-autistic children and adults, who were asked to freely select one of three candies and feed the animals appearing on a tablet. The candies were associated with different probabilities of delivering a neutral vs no effect (agency task), or a positive vs neutral effect (reward task). Choices and reaction times (RT) were measured to understand whether participants preferred and were faster at selecting options with higher probability of producing a neutral vs. no effect (agency) or a positive vs. neutral effect (reward). Participants’ choices and RT were not affected by agency, whereas a more frequent selection of the option with higher probability of a positive vs. neutral effect emerged across groups, thus suggesting a reward effect. Autistic participants selected less frequently the option with chance level of receiving a neutral or no effect, which could be interpreted as a sign of reduced tolerance of uncertainty. Across tasks, conditions and age groups, autistic participants presented shorter RT, which is a marker of reduced action planning and control. Future research should deepen how tolerance of uncertainty, action planning and control impact the way autistic individuals make choices in everyday life situations, potentially contributing to restricted and repetitive behaviours.

Back to the reviewer's suggestion, using age as continuous would result in estimating the effect of age on a year-to-year basis. We agree that in theory this is a more informative approach, however without a large sample the possibility of estimating the age effect so finely is limited. Following the reviewer's input, we've run the models using age in the continuous format. In this way, some of the models fail to converge, which is probably related to the insufficient data to model the effect of age as continuous.
R: Minor concern: I am not sure if this is because of the format of the journal, but aim and hypotheses are not materials and methods at all, so these descriptions have to be rather integrated in the introduction. A: We thank the reviewer for this suggestion, and we moved the aim and hypotheses section accordingly.
Reviewer #2: the manuscript "Agency and reward across development and in autism: a free-choice paradigm" is well-written and provides data and script. Would be neat to also provide the material, as this would facilitate replications (combining your task with eye-tracking for example). A: We thank the reviewer for this nice suggestion and great idea. We have now set the Labvanced experiment as open access design so that the template can be accessed and used by anyone on the Labvaced platform. The link is reported in the manuscript. https://www.labvanced.com/page/library/29586 R: I have only a few issues. I was surprised about the introduction not linking agency to ToM. Agency and intentional binding is seen as a prerequisite for ToM. It was new to me to link it to stereotypes (I am more familiar with stereotypes being a mean to express feelings). Since you did neither know about the participants repetitive movements / behaviours nor their social cognition, it will not matter. A: We thank the reviewer for giving us the opportunity to clarify this aspect. The interest in the link between agency and repetitive behaviours is raised by the authors of the theoretical model of reference (control based response selection framework). As Karsh & Eitam (2015) stated, "this 'reward from control' may explain everyday addictions such as prolonged engagement in arcade games and pathological behaviors, such as stereotypy." However, to the best of our knowledge this link has not yet been tested in the literature. Instead, the link between reward and stereotypies, often conceptualised as self-stimulation, is more studied. We chose autism as a condition of interest for this study, as the presence of repetitive behaviours and stereotypies are one of the two diagnostic macro-categories of the disorder. We agree that it would have been interesting to have a direct measure of the presence of stereotypies in our autistic participants. We now better discuss it as a limitation and future perspective for further studies.
R: There is also no information about co-morbidities or IQ. You do explain why, however, you also describe that the rational for using a non-verbal task is to include children but also those that do not poses sufficient verbal skills as adults (which often means low IQ). You do find differences in choice pattern (and RT) but I do not think they are due to differences in cognitive ability. Still, it would have been nice to know a bit more about your sample, not just age and gender. A: We agree with the reviewer that having a better characterisation of the sample would have been informative. Unfortunately, for privacy reasons we could not access the complete medical records and diagnostic information of the participants. In any case, the heterogeneity of the population on the autism spectrum is such that a much larger sample would be needed to investigate the role of individual differences. Fundamental questions remain unanswered and deserve further large-scale studies.
R: Notably, the RT in the reward task are lower than in the agency task. Here a correlation would be interesting. Is there a higher correlation in the non-autistic group than in the autistic group between RT_agency and RT_reward. The rational for that is smth we see in ASD and SCZ, there is more noise within the ASD and SCZ group, meaning lower correlation across tasks. This would (see below) support the uncertainty and volatility interpretation, i.e. they are less consistent. (..run..) correlation between RT_agency and RT_reward (using longformat and hence calculate Pearson's r per ID and then compare). A: We thank the reviewer for giving us the opportunity to further discuss this interesting aspect. As we disclose in the Procedure, "All participants firstly performed the agency task, and then the reward task, to avoid carryover effects (i.e., a potential reduction of the value of the neutral effect after receiving a positive effect in the previous block of trials)." Therefore, we cannot distinguish whether the different reaction times in the two tasks are due to the fact that the reward effect is greater (in terms of facilitating planning and execution of the choice) or whether it is a general learning effect of the task. As suggested, we did check whether the correlation between RT_agency and RT_reward is different in the two groups. Spearman's correlations reveal very close indices: rautistic = 0.72 (n=54); rnon-autistic= 0.73 (n=54).

We used Spearman's rather than Pearson's correlation as RT is non-normally distributed.
R: Indeed, reporting the SD per group might be informative, I expect the autistic group to be more variable (as Table 2 indicates) -please report the statistical test (if it would be two groups it would be Levene) A: We are grateful to the reviewer for this suggestion. We have run the Levene test on RT_agency (F=61.75; p<.001) and RT_reward (F=6.36; p=.01) by group and reported the results in the relevant section of the manuscript (results and discussion). This evidence indeed suggests that autistic participants have more variable RTs.
R: It is an interesting but not surprising finding that persons with ASD least choose the condition where they can least predict the outcome, i.e. the 50% condition. If you also find more "noisy" responding (use in a GLM the SD per participant as outcome and group (ASD vs non-ASD) and age group as predictor) ….. and you have three independent indications of the ASD group differing from the non-ASD group that align with predictive coding ideas of ASD. A: We thank the reviewer for the insights and practical suggestions. We would first like to clarify how our choice variable is calculated. This is the probability that each participant chooses each of the 3 response keys. As such, we rely on the individual observations (choices) (namely, the dataset in the long format with one row per trial) to calculate this probability. The resulting probability variable, on which the statistical models are based, is thus a single value for each key-participant combination (3 values per participant, which are not independent but sum = 1). What we then do is calculate the difference in the probability of choosing the key with a medium (or high) probability of effect versus the key with a low probability of effect. The result of the reduced probability of the autistic group to choose the medium probability of neutral effect key is based on this variable. We take the reviewer's suggestion and further investigate the homogeneity of the variability between groups in this specific effect of interest (and significant from the models). To do so, we conducted the Levene's test on the aforementioned variable (difference in the probability of choosing the medium probability of effect key versus the low probability of effect key). The result is not significant (F= 3.8383; p= 0.052) and does not allow us to reject the null hypothesis of equivalent variance between groups.
R: Since you provided the data I had a quick look at the elephant in the room, namely whether the difference in RT can be explained by the gender differences. Your ASD sample has mostly males whereas your control group is more even, and not as skewed in the gender distribution as the ASD sample is. In a model with RT as outcome variable and gender as predictor I got a significant effect (can't nest within group as one group had no females, e.g. older children). I would be interested to see a model with gender and whether it wins over model 1 and 2. A: We are aware that our sample is not balanced by gender and we have now explicitly mentioned this as a limitation in representing the general population. In the case of autism, this reflects the male prevalence of the condition, and we have given preference to recruiting male participants for the nonautistic group as well. We did not control for gender in our models not only in light of the unbalanced sample, but also because we had no gender-related research objectives and hypotheses. We preferred to keep the models as simple as possible, prioritising the factors of interest for our research questions. However, as pointed out in the discussion, the models explain a limited amount of variability in the data, and additional intervening variables (including individual differences that might include gender) will need to be explored. However, we do not feel comfortable including analysis on gender effects post hoc, which could lead to misleading conclusions.