With the release of The Diagnostic and Statistical Manual of Mental Disorders (5th ed.; DSM-5; American Psychiatric Association, 2013), “limited prosocial emotions” was introduced as a specifier to the diagnosis of conduct disorder. “Limited prosocial emotions” describe an individual with high levels of callous-unemotional, or psychopathic, traits. These traits are extremely important in conduct disorder, both clinically and forensically, as they demarcate a significant risk for antisocial behaviour that is serious, resistant to typical treatments, and ultimately an adult diagnosis of antisocial personality disorder. Despite some excellent theoretical advances in recent years, testable mechanistic accounts of psychopathic traits are lacking.
Psychopathic traits are associated with, amongst others, increased risk-taking behaviour, punishment insensitivity, and a diminished conditioned-threat response (Moul, Killcross, & Dadds, 2012). These deficits all point towards an underlying problem with accurately predicting outcomes. Here we propose a testable computational framework, which can explain prior data regarding the role of outcome expectancies in psychopathic traits. We have defined a heuristic risk factor, “Learning Window Width” (W) which refers to the number of trials, or learning episodes, that is used to calculate an average expected outcome for a given cue.
In devising how predictions are affected by a simple learning window, we applied a relatively straightforward formula. Every time a learning episode occurs, for a given context or cue, the outcome is added to the string of previous outcomes that have occurred in the same context or in response to the same cue. Thus, with every learning episode, the list of outcome events experienced grows. The content of that list is then used to form a prediction of the next outcome. However, to account for the influence of the history of recent outcome events, the expected outcome value is calculated from the content of a subset of those events (a window) that includes only the more recent of these past experienced outcomes. This learning window begins at the most recently experienced outcome and extends backwards through the history of experiences. The expected value of the outcome on trial t (EVt) is calculated as a simple mean of the contents of the window (i.e. the sum of all the outcome values divided by the number of observations in the window) as illustrated in Figure 1.
The width of the learning window varies between individuals. As such, one individual may have a narrow window (e.g. averaging the outcomes from the past 3 events to calculate EVt) whereas another individual may have a wide learning window (e.g. averaging the outcomes from the past 10 events to calculate EVt). Equation 1 shows how the expected outcome is generated by a learning window width function.
Here, W represents the learning window width (the number of trials included in the window). The expected value is equal to the sum of the experienced outcomes that fall within the window divided by the learning window width.
In this function W has a dual role. When W is small, the outcomes from fewer past trials are aggregated for the purposes of estimating the expected outcome by virtue of the role of W in the summation term. Concurrently, when W is small the relative influence of recent outcomes on the expected outcome is large. In other words, when W is small, fewer recent events are used to estimate the mean outcome and thus each of these events has a greater influence on the expected value than if W was large.
W is a simple functional estimate of the breadth of experience that an individual draws upon when making a prediction about a probabilistic event. In this manner, the wider the learning window the less influence each outcome has on the new expected value. In comparison to a narrower learning window, the expectancy generated by a wider window is more robust against stochastic fluctuations in outcomes given a stable base rate. Conversely, a wider learning window is less sensitive to genuine changes in the base rate and a greater number of experiences of the outcome will be required for the expected outcome to approximate the new base rate. The opposite pattern is true for a narrow learning window – expectancies will be influenced more by stochastic fluctuations but a stable change in base rate will be adapted to more quickly. W is an explicit instantiation of properties that are implicit in the operations of reinforcement learning models, which use learning rate and decay rate to moderate how quickly predictions change. Behrens, Woolrich, Walton, and Rushworth (2007), for instance, point out that a slow learning rate will result in a long decision history while a fast learning rate will result in a short decision history. They modelled differences in the rate of evidence accumulation by modifying the learning rate in a simple prediction error model (see Supplementary Material A for background information regarding prediction error models and the decay rule).
Simulations displayed in Figure 2 show that W provides similar predictions to both prediction error and decay rate models in estimating the expected value of an outcome given stochastic variation and changes in the base rate. This correspondence is expected given that W is simply modelling the rate of information loss. It is, however, indifferent to the manner of that loss – whether it is from the overriding effects of new information (prediction error) or from the loss of old information (decay rate), or both. We present the learning window width construct here as a simple computational device for understanding how variations in a fundamental cognitive factor - the breadth of the accumulated evidence that is used to make decisions - might explain clinically relevant variations in behaviour and psychopathology. Our intention at this stage is not to argue that it fits data better than other computational approaches (indeed, it makes similar predictions to delta and decay models), but rather that it provides a complementary, and in many ways simpler, way of conceptualising this cognitive factor. The utility of this approach is that it makes it clear that a single factor (breadth of decision-relevant learning history) can account for a surprising proportion of the variation in behaviour among individuals who differ in terms of their psychopathic traits. Variance in this factor is naturally a corollary of varying parameters related to learning and decay in other models and thus there are a class of models that could make broadly similar predictions at this simple level of analysis. Future work should aim to design tasks to specifically test the predictive value of this simple ‘learning window’ model relative to other more complex models. Indeed one advantage of the current ‘modelling’ approach is that we are explicit about our assumptions and they can be tested and falsified in future work.
Using W we can make clear predictions about behaviour. We propose that behaviours associated with psychopathic (callous-unemotional) traits can be modelled by a wide W. A wide W results in; 1) slower adaptation to a change in the base rate of a cue-outcome contingency, and 2) an enhanced ability to estimate a stable base rate of a cue-outcome contingency and to make predictions that adhere to it despite stochastic fluctuations.
Slower adaptation to a genuine change in a cue-outcome relationship would result in apparent punishment insensitivity. A greater number of punishing trials would be required for a previously rewarded cue to generate a negative expected outcome. This is a simple result of the greater number of trials used to generate the average expected value in a wider window than in a narrower window. Experimentally, people with psychopathic traits demonstrate this deficit in passive-avoidance and response-reversal tasks (Moul et al., 2012). Indeed Figure 3 shows close correspondence between a simulation of the response reversal deficit produced by a wide W and real experimental results from a forensic sample. Importantly for theories that argue for a cognitive, as opposed to an emotional, account of psychopathy, these simulated data were generated by differences in W alone; rewarding and punishing outcomes were treated equally by the simulation.
The model also predicts that a wide W will lead to cue-outcome contingencies appearing predictable even when they include stochastic fluctuations. Risk aversion is the preference to engage in a behaviour that results in a predictable outcome over a behaviour that may lead to a more favourable outcome but for which the likelihood of that outcome is uncertain. The fewer cue-outcome contingencies that are experienced as unpredictable the less uncertainty, and therefore less risk, is perceived. Thus, a wide W would be associated with reduced risk perception and increased behaviour that appears to the observer as risk-taking. Research has found both diminished risk-perception (Hosker, Molnar, & Book, 2016) and increased risk-taking behaviour (Maes, Woyke, & Brazil, 2018) as psychopathic traits increase.
As some researchers have pointed out, the functional window of experiences that people use to make decisions may vary with experience. For instance, Behrens et al. (2007; see also Behrens, Hunt Woolrich, & Rushworth, 2008; Behrens, Hunt, & Rushworth, 2009) hypothesised that people mentally track the volatility in an underlying relationship, for example how quickly and predictably the underlying relationship between a cue and an outcome changes over time. They present evidence that when encountering a more volatile set of associations, people will adjust by learning faster, effectively adjusting to using a narrower set of recent experiences to make their decisions about the likelihood of an upcoming event. Similarly, Brazil, Mathys, Popma, Hoppenbrouwers, and Cohn (2017) hypothesised that learners develop mental representations not only of the contingencies between events but also their uncertainty about these representations and hypothesise that differences in representational uncertainty may explain some of the characteristics of psychopathy.
Our approach here is different in that, at least in this first step in developing the model, we are not making any assumptions about the learner’s ability to track volatility or mentally represent uncertainty about the precision of their knowledge. Instead we make the simplifying assumption that W is relatively stable for an individual, or at least that it tends towards a default width that can be relatively wide or narrow. Our aim is not to argue that people are unresponsive to volatility and representational uncertainty, nor that their learning rate is completely fixed—although the simple analysis we present here does not speak to these effects, we do not deny that they are real and important. However, our hypothesis cuts across this issue; what we are asking is whether there is an underlying propensity for some individuals to rely upon a short decision history (fast updating) or a long decision history (slow updating) by default and whether this can explain certain deficits related to psychopathy. Variation in the flexibility of the learning window width (for instance, if some individuals are relatively inflexible, while others change too quickly) could be an avenue to explore.
Future work should test these predictions by creating a reliable and valid test of individual differences in learning window width. Then the specific relationships between W and the behavioural features and empirical findings of psychopathy can be explored, and model comparison/simulations can identify whether learning rate or decay rate, or a combination of the two, best explain the behaviour. Future research should also explicitly test the learning window width model against other models that would make similar predictions. Importantly, these tests would need to be designed to be sensitive to the differences between models so that true comparisons can be made. For example, is psychopathy related to differences in learning rate in volatile situations or differences in the representation of the uncertainty of associations in volatile environments? W makes no claim as to the source of learning window width variation. In other words, processes such as attention to, or the encoding of, outcomes or other latent processes involved in associative learning may have a role to play. Investigating these possibilities is an exciting direction for future work but beyond the scope of this paper.
Importantly, there is no reason to suggest that W is uniquely relevant to psychopathy. Given the use of forensic samples in the majority of psychopathy research, it is possible that W could also be useful in understanding conduct problems more generally. It could also, for example, characterise anxiety, which has often been argued to be the mirror image of psychopathy (Cleckley, 1941). Indeed, enhanced risk-aversion but not loss aversion, has been demonstrated in people with pathological anxiety (Charpentier, Aylward, Roiser, & Robinson, 2017) – the inverse to the psychopathy findings highlighted above. In addition to uncovering the role of individual differences in the generation of outcome expectancies in psychopathy, W may provide insights into any psychiatric condition in which aberrances in perceived unpredictability, risk-perception, and punishment-sensitivity feature. In all, W provides an intuitive, accessible and testable mechanistic account for some of the more challenging and immutable features of psychopathic traits.
Caroline Moul reports no competing interests.
Oliver J. Robinson is funded by an MRC senior fellowship partially in collaboration with Cambridge Cognition and he is running an investigator-initiated trial with medication donated by Lundbeck. He also holds an MRC-Proximity to discovery award with Roche for work regarding work on heart rate variability and anxiety. He has also completed consultancy work for Peak and IESO digital health. He is on the committee for the British Association of Psychopharmacology.
Evan J. Livesey reports no competing interests.
American Psychiatric Association. (2013). Diagnostic and statistical manual of mental disorders (5th ed.). Washington, DC: Author. DOI: https://doi.org/10.1176/appi.books.9780890425596
Behrens, T. E., Hunt, L. T., & Rushworth, M. F. (2009). The computation of social behavior. Science, 324(5931), 1160–1164. DOI: https://doi.org/10.1126/science.1169694
Behrens, T. E., Hunt, L. T., Woolrich, M. W., & Rushworth, M. F. (2008). Associative learning of social value. Nature, 456(7219), 245–249. DOI: https://doi.org/10.1038/nature07538
Behrens, T. E., Woolrich, M. W., Walton, M. E., & Rushworth, M. F. (2007). Learning the value of information in an uncertain world. Nature Neuroscience, 10(9), 1214–1221. DOI: https://doi.org/10.1038/nn1954
Brazil, I. A., Mathys, C. D., Popma, A., Hoppenbrouwers, S. S., & Cohn, M. D. (2017). Representational uncertainty in the brain during threat conditioning and the link with psychopathic traits. Biological Psychiatry: Cognitive Neuroscience and Neuroimaging, 2(8), 689–695. DOI: https://doi.org/10.1016/j.bpsc.2017.04.005
Budhani, S., Richell, R. A., & Blair, R. J. (2006). Impaired reversal but intact acquisition: probabilistic response reversal deficits in adult individuals with psychopathy. Journal of Abnormal Psychology, 115(3), 552–558. DOI: https://doi.org/10.1037/0021-843X.115.3.552
Charpentier, C. J., Aylward, J., Roiser, J. P., & Robinson, O. J. (2017). Enhanced Risk Aversion, But Not Loss Aversion, in Unmedicated Pathological Anxiety. Biological Psychiatry, 81(12), 1014–1022. DOI: https://doi.org/10.1016/j.biopsych.2016.12.010
Hosker-Field, A. M., Molnar, D. S., & Book, A. S. (2016). Psychopathy and risk taking: Examining the role of risk perception. Personality and Individual Differences, 91, 123–132. DOI: https://doi.org/10.1016/j.paid.2015.11.059
Maes, J. H. R., Woyke, I. C., & Brazil, I. A. (2018). Psychopathy-related traits and decision-making under risk and ambiguity: An exploratory study. Personality and Individual Differences, 122, 190–194. DOI: https://doi.org/10.1016/j.paid.2017.10.017
Moul, C., Killcross, S., & Dadds, M. R. (2012). A model of differential amygdala activation in psychopathy. Psychological Review, 119(4), 789–806. DOI: https://doi.org/10.1037/a0029342