Exploring randomness in autism

Dual-system model and arbitrary decision making

In our everyday life, when confronted with dilemmas that can be analysed with rational criteria, either one or both thinking systems can be recruited to reach a decision (Kahneman, 2003). What happens however, when arbitrary decisions need to be made? In situations where arbitrary decisions need to be made no clear reason exists as to why one choice should be preferred over the other one, since from all available choices none is more advantageous than the others. This scenario might seem extreme but almost every day we are faced with such choices. Which apple would you pick from the pile in the store, when all apples placed in front of you are of the same quality? When exactly will you choose to schedule your first appointment, during working hours, when your calendar is free, at 10:15 am or 10:20 am? Which pencil would you peak from a box of identical pencils? These and many other similar decisions raise the question of how a “thinking” system that is based on reason, can actually make such arbitrary choices?

Depending only on system 1, an arbitrary decision can be made automatically, without active consideration, and it could be ‘biased’ or not. By ‘biased’ we mean any decision that would not be considered random by the system. System 1, for example, may be guided from our habits to make such a decision. In that case and in the aforementioned example we would choose the apple that is on the left side of the pile, because we always do so but may not consciously remember. When not ‘biased’, our arbitrary decisions could be the result of a random function that could be expressed per se or influence system 1 (Hogarth, 2005).

When the slower system 2 is involved in arbitrary decisions, this can happen through aroused emotions, memories and attention to environmental nudges that are provoked by system 1 (e.g., recognize the outline of a beloved person on the surface of the apple), by substituting the question we are faced with, with a rational one that could be answered easily (e.g., which apple is more red), by consciously following or deviating from a set of rules (e.g., always choose apples from the bottom row or decide to actively avoid this pattern), or by depending on our environment either passively (e.g., pick the apples that will remain untouched by other customers) or actively (e.g., coin flipping or asking someone else to make a decision for us) (Hogarth, 2001; Moulton et al., 2007). Our efficiency and ability to make such decisions, can impact our everyday functioning. If we actively need to consider all our choices, that would require the involvement of system 2 and this could impact our performance especially in relation to the time and effort we spend in decision making (Croskerry & Nimmo, 2011; Moulton et al., 2007). Moreover, if the responsibility of making a decision has to be transferred to someone else, then that will result in a dependent behaviour. Alternatively, when certain rules need to be strictly followed that can lead to a stereotypical behaviour.

Randomness in arbitrary decision making

Randomness as a concept is by itself a complicated one, and the way humans perceive randomness might differ from the actual physical notion of the concept (Bar-Hillel & Wagenaar, 1991). Many attempts to define randomness have been made. Each one focuses on different concepts that events which could be considered random should entail, including equal probability, sequential independence and unpredictability (Nickerson, 2002). As the most widely accepted characteristic of a random sequence of events would be that of equal probability, for the purpose of this paper, we define a sequence of infinite events as random when all possible events have equal probability to occur at each time point of this sequence (Borowski, n.d).

Several theories have been stated and experiments conducted in order to examine how humans perceive randomness. Many of these theories and experiments support that humans are very poor in recognizing sequences that are the result of random processes (for review of earlier experiments Tune, 1964). When the production of random number sequences is concerned, literature findings are conflicting as some studies show that they produce number sequences that very well resemble those of a random process (Persaud, 2005) while other studies indicate that humans have a relatively poor performance in this task (e.g., Bains, 2008; Figurska, Stańczyk & Kulesza, 2008). Some of these findings could be due to the fact that our cognitive capacity is limited (Bains, 2008; Persaud, 2005). Moreover, when no instructions are provided to participants as to what is considered random or no indication is provided that a random approach would be needed to execute a task, participants can produce random sequences approaching them as arbitrary decisions and thus successfully engage in tasks where a random approach is the best strategy (Rapoport & Budescu, 1992; Sharifian, 2016).

Randomness exists in nature and can behaviourally be expressed as an ability to avoid predictability, thus linked to an evolutionary advantage (Riotte-Lambert & Matthiopoulos, 2020) . Experiments in animals have shown that their brain can turn on a ‘random mode’ when rules and settings change and there is no prior knowledge to guide decisions (Tervo et al., 2014). In the same context, when infants are placed in a novel environment, for example with toys they have never seen before, the decision around which one they would pick, is arbitrary. To reach that decision, they have to obviously recruit a randomness mechanism. The choice they make, under an evolutionary perspective, if proven a safe one, would shape their future preferences (Silver et al., 2020).

Randomness in decision making is a rather difficult concept to examine experimentally. We can identify three ways randomness can be expressed in arbitrary decision making. True randomness when uncontrolled and unpredictable phenomena are utilized (e.g., coin flipping, rolling dice), imitated randomness when decisions are made based on a set of rules (e.g., making a conscious unprecedented choice to avoid following a pattern) and innate randomness through a possible innate mechanism of randomness production. Finally, a combination of the last two ways of randomness expression could serve arbitrary decision making, especially when a set of rules is not fully descriptive. For example, a decision to produce a number that cannot be a repetition of the last two numbers produced, can comply both with imitated and innate randomness. The first two ways, rely heavily on system 2 while the last mechanism would be part of system 1. The concept of innate randomness is one proposed by the authors of this paper, based on previous research findings and observations (for example, Wong, Merholz & Maoz, 2021; Tervo et al., 2014; Schulz et al., 2012; Laskaris, Zafeiriou & Garefa, 2009; Tune, 1964). The characteristics of an innate randomness mechanism remain largely unknown, although some observations have been made in animal and human studies (Tervo et al., 2014). In this study, we try to examine the possibility of innate randomness existence by trying to “force” its expression by favoring system 1.

Autism and the dual-system thinking model

Autism is one of the most common neurodevelopmental disorders (CDC, 2016). Its symptomatology includes persistent impairments in reciprocal communication and social interactions as well as restricted, repetitive patterns of behaviours, interests or activities (APA, 2013). Different autism phenotypes are characterized by a diversity of traits (Geschwind, 2009).

As far as decision making is concerned, autistic individuals are faced with difficulties, and are generally slower in taking decision and show less intuitive and more deliberate reasoning (Brosnan, Lewton & Ashwin, 2016; Farmer, Baron-Cohen & Skylark, 2017; Luke et al., 2012; Vella et al., 2018). To account for those characteristics and based on experimental knowledge, De Martino et al. (2008) proposed a model according to which autism would more heavily rely on system 2 when decision making is concerned.

Although helpful to interpret certain autistic behaviours during decision making, the De Martino model, does not offer any insight concerning the role of system 1. It is possible that autistic individuals heavily rely on system 2 as a counterbalancing strategy to a compromised system 1. Alternatively, system 1 could be intact but system 2 is overactive and dominant. Additionally, autistic individuals experience difficulties in using emotional or attentional cues in order to make decisions (Gaigg, 2012), indicating that the way by which system 1 could influence system 2 is even further limited and this could lead to stereotypical behaviour (Cunningham & Schreibman, 2008; Farmer, Baron-Cohen & Skylark, 2017) and dependency and would point towards autism being characterized by difficulties in abstract decision making. Reports of autistic individuals and their caregivers highlight the difficulties of these individuals to decide, especially when new choices need to be made or when decisions need to be taken on the spot (Frith, 2001; Vella et al., 2018). A recent study has indeed shown that autistic individuals that have high functioning ASD experience difficulties in routine everyday decisions, including what clothes to wear and when to shower but these difficulties are not present when crucial life decisions are involved (Levin et al., 2015). This difficulty could be a limiting factor to their everyday functionality which is of paramount importance and a major factor to consider in the diagnosis of autism.

The Random Number Generation (RNG) task

The RNG task is commonly used to examine the concept of randomness in humans (Matsukawa et al., 2006; Rosenberg et al., 1990; Shinba et al., 2000; Spatt & Goldenberg, 1993; Williams et al., 2002). During the task, participants are asked to produce a sequence of numbers that adheres to predefined rules that would make a sequence random, such as avoid patterns, repetitions of the same digits or use all digits available before recycling them. This could lead to responses that would imitate randomness. All the different ways that the RNG task has been employed, are based on the definition of what constitutes a random number sequence, that is either provided to participants or is based on how they perceive randomness themselves.

In that sense, the RNG and similar tasks, become heavily dependent on cognitive and attentional processes (Towse, 1998; Williams et al., 2002). Each choice of number is not an entirely arbitrary choice but a well thought decision which would depend on processes that are employed by system 2. In these types of experiments, participants’ performance improves with training and feedback, however, these experiments deviate from a concept of innate randomness and treat randomness as a skill that could be improved with practice (Biesaga, Talaga & Nowak, 2021; Gauvrit et al., 2017). Indeed, a recent experiment by Biesaga, Talaga & Nowak (2021) has shown that when cognitive demands and contextual task conditions are optimal, participants can produce sequences that could be considered more random, highlighting the importance of providing task instructions to participants to achieve that effect. The cognitive demands of such tasks become apparent however, as participants tire with time and the randomness of the sequences they produce decreases. The significant contribution of cognitive processes to the typical RNG task, also becomes apparent when the task is performed by individuals who experience cognitive difficulties. Patients with Alzheimer’s disease, schizophrenia as well as autistic individuals (low functioning) tend to produce shorter number sequences with more digit cycling and repetitions than the comparison groups and these differences have been attributed to attentional deficits, difficulties in working memory and executive functions (Brugger et al., 1996; Proios, Asaridou & Brugger, 2008; Williams et al., 2002).

Study aims

The primary aim of this exploratory study is by silencing system 2, to show the contribution of system 1 and innate randomness in arbitrary decision making in autism. In order to achieve that, we modified the RNG task to limit the contribution of cognitive processes, such as working memory and consequently the contribution of system 2 to the task. That would allow us to examine the role of system 1 in arbitrary decision making in autism. The RNG task has not, to our knowledge, been administered to autistic participants with high functioning ASD before. However, given that autistic individuals rely heavily on system 2 to make decisions (De Martino et al., 2008; Levin et al., 2015) and considering that in our task the contribution of system 2 is reduced, we hypothesize that the autism and the comparison group would differ in both the quality of the randomness features and the quantity of the numbers produced.

Participants

This study recruited autistic adults (36 adults, 30 male) and an age and IQ-matched comparison group (40 adults, 34 male). For details see Table 1.

The autistic participants were recruited from a larger cohort of volunteers (Pehlivanidis et al., 2020) who participated in a research project on the de novo diagnosis of adults with neurodevelopmental disorders. ASD diagnosis was based on DSM-5 criteria and everyone in the autism group had fluent phrase speech and more than 12 years of education. Exclusion criteria included the presence of acute psychopathology, systematic psychopharmacological treatment up to 30 days prior to taking part in the study, current substance abuse disorder, IQ<70 assessed with Wechsler Adult Intelligence Scale (WAIS-IV), any known genetic condition (Wechsler, 2012).

The comparison group was recruited via advertisement and word of mouth and included neurotypical individuals with IQ>70 (WAIS-IV) and more than 12 years of education. Exclusion criteria included the presence of acute psychopathology.

Written informed consent was obtained from all participants and the study was approved by the Ethics Committee of the Department of Psychiatry, National and Kapodistrian University of Athens (IRB 12/7/2018 #517). All experiments were performed in accordance to relevant guidelines and regulations.

Modified random number generation task

Participants were instructed to pronounce a sequence of numbers, as comes to mind, with each number ranging from 1 to 10. The task duration was 1 min. Participants were advised but not obliged to produce numbers at a steady pace of approximately 1 number per second. No example sequence was given to participants to avoid unwanted influences from the experimenter.

To ensure minimal distractions and keep the experimental conditions as similar as possible between participants, individuals were placed in a dark room. The only source of light was a computer screen which was displaying a progress bar used to count down the remaining task time and helping participants to keep the pace. Throughout the task, participants were listening to a track consisting of nature sounds (forest, rain sounds etc.) at a loud volume through a headset. Nature sounds were chosen as they already are familiar to individuals. The sound volume was loud enough to block any surrounding noise as well as the sound of their own voice and this was confirmed before the start of the task. The purpose of this modification was to limit the contribution of auditory working memory—a process mainly employed by system 2, in other words limit the ability of individuals to choose a number based on the sequence of numbers that have already been used. The sequence was recorded automatically. The experimenter was present in the room to ensure proper execution of the task but was out of the participants’ field of view and did not interact with them in any way. Each participant produced a single random number sequence and at the end of each session the number sequences were transcribed by the investigator and the participants’ recording were deleted.

Data analysis

The task data were processed using in house code in R and Python programming languages. Two groups of variables were calculated using the RandSeq (Oomens et al., 2021) package in R. The first group of variables included standard measures of randomness: Redundancy (R), Random Number Generation (RNG), RNG2, Null Score Quotient (NSQ), Adjacency (see Table S1). The second group included measures that describe the characteristics of the sequence itself: response frequencies (RF), first order difference (FOD) (Towse, 1998) as well as the correlation frequency index (Cf) (Barbasz et al., 2008) (Table 2). Given that in our task sequence lengths were different between participants we computed the relative equivalent of the second group of variables (Table 2). These are standard variables that have been widely used in the literature to examine human randomness as expressed through the RNG task. In order to observe how randomness measures might change over time, an overlapping moving window of 20 secs in length (33% of total task time) was also applied to the sequences and the above variables were calculated for each step. In order to focus more on any “pattern” characteristics that the task sequences might have we also performed a recurrence quantification analysis (RQA) in R using casnet package (Hasselman, 2022). All the variables that were calculated are defined in Table 1 and for the values of each variable for both groups are summarised in Table 2.

A classification approach was used to examine whether we can distinguish the two groups based on the characteristics of the sequences that each group produced. The sklearn Python package for machine learning was used and a decision tree method was followed. The GridSearchCV was consulted for the hyperparameters selection. The model was tested for all possible combinations of features, up to three features and leave-one-out cross validation was applied. For the replication of the results the ‘random_state’ variable in both splitting the data and the decision tree functions, was set to 1.