Overestimation of volatility in schizophrenia and autism? A comparative study using a probabilistic reasoning task

Background and objectives A plethora of studies has investigated and compared social cognition in autism and schizophrenia ever since both conditions were first described in conjunction more than a century ago. Recent computational theories have proposed similar mechanistic explanations for various symptoms beyond social cognition. They are grounded in the idea of a general misestimation of uncertainty but so far, almost no studies have directly compared both conditions regarding uncertainty processing. The current study aimed to do so with a particular focus on estimation of volatility, i.e. the probability for the environment to change. Methods A probabilistic decision-making task and a visual working (meta-)memory task were administered to a sample of 86 participants (19 with a diagnosis of high-functioning autism, 21 with a diagnosis of schizophrenia, and 46 neurotypically developing individuals). Results While persons with schizophrenia showed lower visual working memory accuracy than neurotypical individuals, no significant group differences were found for metamemory or any of the probabilistic decision-making task variables. Nevertheless, exploratory analyses suggest that there may be an overestimation of volatility in subgroups of participants with autism and schizophrenia. Correlations revealed relationships between different variables reflecting (mis)estimation of uncertainty, visual working memory accuracy and metamemory. Limitations Limitations include the comparably small sample sizes of the autism and the schizophrenia group as well as the lack of cognitive ability and clinical symptom measures. Conclusions The results of the current study provide partial support for the notion of a general uncertainty misestimation account of autism and schizophrenia.

Nevertheless, to obtain an approximate estimate of "random" or "noisy" behavior in the beads task, an additional measure was constructed based on all those occurrences where when a bead was of the same color as the previous two, the belief was updated into the opposite direction, i.e. the belief in the currently presented colors was decreased.
Example: a participant sees three white beads in a row and indicates a probability for them to originate from the bag with more white beads as 0.7 and 0.8 for the first two trials. On the third trial, they then decrease their belief to 0.7 again when actually, given the evidence, they should keep increasing their belief certainty about the beads to originate from the bag with more white beads.
Such "random belief updating" was calculated as the mean change in belief across all occurrences of this kind for each sequence, averaged over number of sequences for each participant.
A non-parametric Kruskal-Wallis test (due to the high positive skewness in random belief updating) revealed no significant group difference, χ 2 (2) = 3.32, p = 0.19, ε 2 = 0.04. Across groups, random belief updating was strongly and positively associated with volatility, ρ = .63, p < .001. While this might suggest that estimated volatility largely reflected noise or random behavior, it is important to consider that a conceptual distinction between both concepts may not fully be valid. After all, "random" belief changes may indeed be caused by an increased belief about the frequency with which the bag of origin is secretly changed (volatility), even in the absence of obvious evidence for an occurred change.
Importantly, volatility was also strongly related to disconfirmatory belief updating. Here, the conceptual relationship between both variables is slightly more obvious: in an unstable environment, disconfirmatory evidence might suggest an occurred changeso the larger one thinks the probability is for a change to occur, the more one will react to disconfirmatory evidence in terms of belief updating.
An additional analysis was conducted to gauge to what extent both random and disconfirmatory belief updating contributed to estimated volatility. Participants were divided into groups with high (above the median) or low (below or equal to the median) volatility estimates. A logistic regression was conducted on volatility group membership (0 = low, 1 = high), including main effects of both random and disconfirmatory belief updating, both standardized. McFadden's R 2 of this model was .40, and the Odds Ratio was 10.12 for (standardized) random belief updating [CI 2.5%: 2.64, 97.5%: 53.23] and 1.92 for (standardized) disconfirmatory belief updating [CI 2.5%: 1.92,97.5%: 10.62]. This demonstrates that even if random belief updating was interpreted as a pure measure of "noise" caused by different factors than an overestimation of volatility, when accounting for its contribution to volatility there remains a significant contribution of disconfirmatory belief updating, a variable which is clearly also conceptually related to volatility.

Volatility change throughout the task
Since feedback was provided after every completed sequence in the beads task, learning processes may have caused a decrease in subjective volatility over time. In the original volatility model, subjective volatility was estimated based on all sequences. To explore whether volatility estimates might have decreased over time, the model was refitted to the first two and the last two sequences, respectively. Volatility change was then calculated by subtracting volatility estimated for the first two sequences from volatility estimated for the last two sequences for each participant, with values below zero indicating a decrease of volatility towards the end of the task.
A one-sided one-sample Wilcoxon signed-rank test (due to the non-normality of the volatility change variable) on data of the whole sample confirmed that indeed, this change was significantly below zero across participants, Md = -0.01, V = 1260, p < .01.
To assess whether groups differed in terms of this volatility change, a Kruskal-Wallis test was applied. This did not reveal any significant group differences, χ 2 (2) = 0.77, ε 2 = 0.06, p = .68, indicating that groups learned similarly from feedback.