Longing to Act: Bayesian Inference as a Framework for Craving in Behavioral Addiction

Abstract

Traditionally, craving is considered a defining feature of drug addiction. Accumulating evidence suggests that craving can also exist in behavioral addictions (e.g., gambling disorder) without drug-induced effects. However, the degree to which mechanisms of craving overlap between classic substance use disorders and behavioral addictions remains unclear. There is, therefore, an urgent need to develop an overarching theory of craving that conceptually integrates findings across behavioral and drug addictions. In this review, we will first synthesize existing theories and empirical findings related to craving in both drug-dependent and -independent addictive disorders. Building on the Bayesian brain hypothesis and previous work on interoceptive inference, we will then propose a computational theory for craving in behavioral addiction, where the target of craving is execution of an action (e.g., gambling) rather than a drug. Specifically, we conceptualize craving in behavioral addiction as a subjective belief about physiological states of the body associated with action completion and is updated based on both a prior belief (“I need to act to feel good”) and sensory evidence (“I cannot act”). We conclude by briefly discussing the therapeutic implications of this framework. In summary, this unified Bayesian computational framework for craving generalizes across addictive disorders, provides explanatory power for ostensibly conflicting empirical findings, and generates strong hypotheses for future empirical studies. The disambiguation of the computational components underlying domain-general craving using this framework will lead to a deeper understanding of, and effective treatment targets for, behavioral and drug addictions.

Craving as a transdiagnostic concept

In the latest iteration of the Diagnostic and Statistical Manual of Mental Disorders (DSM-5), craving has been included as an essential diagnostic criterion for substance use disorders (SUDs), reflecting its central role in addiction. For individuals struggling with SUDs, craving is a highly intrusive and pervasive occurrence. Craving is commonly associated with unpleasant withdrawal symptoms or relapse (Filbey & DeWitt, 2012; Haughey et al., 2008; Norberg et al., 2016; Singleton et al., 2002; Tiffany & Wray, 2012) and interferes with cognitive processes that are employed in day-to-day life, such as working memory and decision-making (Ray & Roche, 2018). In turn, dealing with the stress and negative outcomes associated with craving may reciprocally drive addictive behaviors. Indeed, ecological momentary assessment (EMA) and other self-report studies have demonstrated that high levels of craving are predictive of increased usage across many substances (Koob & Volkow, 2010). Importantly, craving often persists after long periods of abstinence, a phenomenon known as “incubation of craving” (Gawin & Kleber, 1986; Grimm et al., 2001; Parvaz, Moeller, & Goldstein, 2016). Treatments and interventions that can curb drug-taking behavior may not necessarily reduce subjective craving (Tiffany & Wray, 2012). These observations suggest that craving, while heavily intertwined with overt behaviors associated with drug-taking, is a distinct subjective experience that requires its own mechanistic explanation.

Similar to SUDs, behavioral addictions such as gambling disorder (the only form of behavioral addiction diagnosis acknowledged by DSM-5), involve repetitive and impulsive actions that prove to be detrimental to the long-term wellbeing of the individual (Robbins & Clark, 2015). Across both categories, affected individuals can exhibit diminished control and perseveration of maladaptive behaviors (Grant et al., 2010), features that overlap with characteristics of other distinct diagnoses, including obsessive-compulsive disorder (OCD), tic disorders, and eating disorders (Robbins & Clark, 2015). In contrast to SUDs, behavioral addictions are directed toward one specific behavioral target (e.g. gambling) instead of a stimulus; and the execution of the desired behavior can be followed by feelings of relief or pleasure (Bonny-Noach & Gold, 2021; Grant et al., 2010). Despite this difference, clinical manifestations of behavioral addiction mimic those of SUDs, including the impairment of normal daily functioning, loss of control, increased tolerance, and withdrawal (Rash et al., 2016; Tiffany & Wray, 2012). These empirical findings lead to a longstanding question - can one crave performing certain actions such as gambling, in a way similar to drug craving?

Although craving is not officially listed as a core symptom of behavioral addiction in DSM-5, accumulating evidence suggests that craving can exist in the context of behavioral addiction, whether the individual craves a drug-induced high, relief from tension, attenuation of intrusive thoughts, or some other reward associated with a specific action or set of behaviors (Grasman et al., 2016). Craving, then, can be considered a unifying component of impulsive-compulsive disorders, and should be evaluated as a promising potential avenue for mechanistic and therapeutic research. In this article, we aim to present a succinct overview of the literature on craving in behavioral addiction, with a focus on providing a formal, computational account for action craving.

Existing paradigms and frameworks of craving in addictive disorders

Research on craving has primarily relied on cue-reactivity paradigms (Filbey & DeWitt, 2012), which stems from the Pavlovian conditioning literature and suggests that repeated pairing of a cue (e.g., a bar sign) with a rewarding unconditioned stimulus (e.g., alcohol) increases the value of those cues which become conditioned stimuli that can elicit the same subjective states such as craving. Although action craving is relatively under-investigated compared to drug craving, initial evidence based on similar cue-elicited paradigms has shown that individuals with gambling disorder exhibit increased activity in distributed areas, including the insular cortex, anterior cingulate, ventral striatum, and ventromedial prefrontal cortex (vmPFC) when presented with a gambling cue versus a non-gambling cue (Limbrick-Oldfield et al., 2017). These neural signatures parallel those found to be highly involved in SUD craving (Kühn & Gallinat, 2011; Naqvi et al., 2015).

Pavlovian conditioning, however, does not fully capture the complexity of either SUDs or behavioral addiction. Instrumental or reinforcement learning (RL) theories, which suggest that behaviors can be strengthened or weakened by their outcomes, have become highly successful in accounting for the development of addictive behaviors. A rich behavioral psychology and neuroscience literature (see (Everitt & Robbins, 2005) for a comprehensive review) has demonstrated the importance of learning mechanisms and associated neural circuits (e.g., a ventral corticostriatal loop). In this view, instrumental learning dominates during the goal-oriented phase, where an addictive behavior (e.g., consuming a drug or gambling) is reinforced by positive outcomes (e.g., feeling high or winning money). Later on, a habitual circuit takes over, resulting in persistent repetition of the reinforced actions even when outcomes are neutral or negative (Vanderschuren & Everitt, 2004; Voon et al., 2015). Recent work has also examined other processes such as model-based (MB) learning in addictive behaviors (Groman et al., 2019) and how increasing MB learning might prevent habit formation (Gillan et al., 2015). However, this work still focuses on explicit choice behaviors per se, leaving the subjective phenomenology of craving unexplained. The numerous empirical findings reviewed above, such as incubation of craving (Grimm et al., 2001) and resistance of craving to treatment (Tiffany & Carter, 1998), are not explained by existing learning models and highlight the dissociability between craving and choice behaviors. We aim to address this significant gap in the field by proposing a distinct framework for the subjective experience of craving.

An interoceptive inference view of craving

Modeling subjective states is a difficult topic and not unique to craving. Recent advances in computational neuroscience and computational psychiatry literatures, however, have presented the Bayesian brain framework as a likely candidate theory (see (Friston, 2010; Gu et al., 2019; Owens et al., 2018). In the past decade, Bayesian models have become a widely used framework to account for various mental processes including perception (Knill & Pouget, 2004), emotion (L. F. Barrett & Simmons, 2015; Seth & Friston, 2016), and decision-making (Schwartenbeck et al., 2015). Briefly, the Bayesian brain hypothesis proposes that the brain updates its beliefs about the world or itself by combining prior knowledge or expectation, and sensory evidence, based on Bayes’ rule:

$$\mathit{\text{posterior}}\mathit{\text{probability}}=\frac{\mathit{\text{likelihood}}\times \mathit{\text{prior}}\mathit{\text{probability}}}{\mathit{\text{marginal}}\mathit{\text{likelihood}}}$$

Here, a “belief” in the Bayesian sense is not necessarily conscious as in its colloquial use (e.g., a political or religious belief); instead, it is defined as a mental state based on the neural computation of probabilistic distributions of prior expectations and sensory evidence. Amongst the types of Bayesian beliefs the brain can develop, we (Gu et al., 2013; Gu & FitzGerald, 2014) and others (L. F. Barrett & Simmons, 2015; Seth & Friston, 2016) have previously proposed that the brain actively makes predictions and form beliefs about interoceptive states (i.e. “interoceptive inference”). This proposal accounts for a wide range of emotional phenomena (e.g., increased heart rate can be interpreted by the brain as anxiety). Crucially, under this framework, craving can be considered a special case of Bayesian belief about bodily states associated with availability of addictive substances (Gu & Filbey, 2017).

Accumulating evidence from the addiction neuroscience literature now supports this Bayesian view of craving. For instance, several studies have shown that craving not only depends on the availability of the addictive substance in the body, but also depends on smokers’ beliefs about the presence of nicotine (Gu et al., 2015; Juliano et al., 2011; Kelemen & Kaighobadi, 2007; McBride et al., 2006). Using the same Bayesian framework, we were also able to account for incubation of craving, which refers to the exacerbating rather than diminishing effect of craving during early abstinence (Bedi et al., 2011; Conrad et al., 2008; Grimm et al., 2001; Gu, 2018; Lu et al., 2004; Parvaz, Moeller, Malaker, et al., 2016). Taken together, this Bayesian framework has been proven powerful in accounting for craving in drug addiction.

Behavioral addiction: craving as a Bayesian belief about action

Here, we extend the Bayesian view of drug craving to action craving (Fig. 1). In this view, the Bayesian prior represents the agent’s belief about the physiological state associated with performing an action (e.g., feeling good from gambling); the likelihood represents the probability distribution of whether the action was indeed completed; the posterior is the new belief about the physiological state of the body updated based on the prior and the observed likelihood of action completion. Compared to our previous Bayesian model for drug craving, this model is based on the assumption that action completion itself can be a desirable target. Consequently, the discrepancies between the action probability distribution and linked outcomes such as “feeling good” and the posterior belief will be experienced by the brain as a mismatch signal corresponding to perceived action craving (Fig. 1). The prior distribution can be conceptualized as the “internal” pathway of craving, aggregating the internal states relevant to action completion, including memory of prior action-related experience, as well as interoceptive representations of relevant addictive cues. In contrast, the likelihood, or the probability distribution of action completion and associated outcomes, can be conceptualized as the “external” pathway. As such, this framework takes into consideration both components that determine an agent’s current belief: the prior (internal, based on previous experience about performing an action) and the likelihood (external, dependent on the current environment and feasibility of performing the desired action).

FIGURE 1. Proposed Bayesian framework for action craving.

When beliefs about action completion are normative, the Bayesian prior can be depicted using a probability distribution of the belief, or expectation, associated with the desirable action. Upon execution of the action, represented as a probability distribution of Bayesian likelihood, perceived action craving is resolved to some extent (left panel). In behavioral addiction, prior beliefs about action completion are pathological and the Bayesian prior becomes hyper-precise. This change diminishes the resolution of craving after action completion and leaves a high level of perceived craving (right panel).

This framework provides a computational account for subjective states related to compulsive behaviors in general; importantly, it accounts for the pathological version of these phenomena in behavioral addiction. Take gambling addiction as an example. Here, repeated execution of an action (gambling) leads to a hyper-precise prior belief (Fig. 1; “I must gamble to feel good” instead of “I might feel good by gambling”). Specifically, the level of precision represents the subjective certainty (or uncertainty) around the association between obtaining an addictive target (e.g., gambling) and bodily states, a computation that is distinct from whether one’s basic interoceptive function (ATEŞ ÇÖL et al., 2016) or general self-relevant metacognitive knowledge (Brevers et al., 2014; Goldstein et al., 2009; Moeller et al., 2010). In other words, it is learned interoceptive associations through the initial instrumental learning phase of addiction that drives hyper-precision in this pathological gambling example. In turn, this hyper-precision of the prior causes the resultant posterior belief to become more resistant to the influence of the likelihood signal even when the new evidence (i.e., “I got to gamble”) suggests that craving should be at least partially resolved. This framework also explains the “incubation of craving” effect in behavioral addictions (Fig. 2); during periods of abstinence (absence of action), the likelihood distribution is right-shifted, and subsequent updates to the resultant posterior distribution cause it to become hyper-precise, which is perceived as increased and more intense craving.

FIGURE 2. Incubation of craving.

When desired actions are unable to be executed, perceived craving increases systematically over time, referred to as the incubation of craving effect. A Bayesian view of incubation of craving depicts it as a series of Bayesian updates to the posterior belief about physiological states. When actions cannot be executed, the likelihood distribution is highly right-shifted, and consequently, posterior action craving increases. In subsequent updates, if actions still cannot be executed, the posterior belief continues to shift right and increase in precision, manifesting as incubation of craving.

The key distinction between this Bayesian view and previous theories about habits and compulsion is the emphasis on subjective beliefs (and the uncertainty around them). Habits are typically conceptualized as a “model-free” mode of behavior resulting from conditioned stimulus-action associations (Watson et al., 2018) and thought to dominate the later stage of addiction (Everitt & Robbins, 2005). Here, we propose that subjective beliefs derived from these explicit actions – and what they entail for the agent – require their own distinct explanation scientifically, and matter for humans who suffer from behavioral addictions clinically. It follows that treatment strategies should consider not only the extinction of the action, but also associated subjective experience – such as craving for the actions.

Implication for treatment and intervention

Based on the literature reviewed so far, we consider a computational formalization of action craving (along with drug craving) to be important for not only obtaining a deeper and mechanistic understanding of craving per se, but also generating testable predictions that have therapeutic implications. For example,

• Abstinence alone might not be a sustainable treatment approach, as prolonged abstinence from the desired action without other interventions (e.g., CBT) will increase craving (i.e., incubation of craving).

• Interventions that can reduce the precision of the prior belief (e.g., from “I must gamble to feel good” to “I might or might not feel good by gambling”) will effectively reduce action craving.

• Substituting the old prior belief (e.g., “I must gamble to feel good”) with a new belief (e.g., “I also feel good by going to the gym”) can also change craving for the old action.

Many existing treatments, such as contingency management (Lamb et al., 2004), motivational interviewing (Heckman et al., 2010), and community reinforcement (Meyers et al., 2011), are long-standing and effective psychotherapeutic tools for treatment and management of addiction. Nevertheless, these approaches primarily focus on reducing addiction-associated behaviors rather than craving as a subjective experience. In line with our model’s focus on the subjective experience of craving, we are most interested here about therapeutic approaches to alleviate craving directly, though purely behavioral approaches may still achieve this indirectly by modifying the precision of the prior belief of how obtaining the addictive target can lead to bodily states. Several highly effective craving reduction therapies already exist, including cognitive behavioral therapies (Kober et al., 2010; Lopez et al., 2022; Naqvi et al., 2015), mindfulness (Li et al., 2018), and dialectical behavioral therapy (Rezaie et al., 2021), and our framework synergizes with these existing therapeutic models by providing mechanistic explanations for their success. Indeed, such collaborative efforts between experimentalists, computational modelers, and clinicians are essential for gaining valuable insights into the mechanisms underlying successful treatment outcomes, hopefully leading to more refined and personalized therapies in the future.

Finally, brain-based approaches such as transcranial magnetic stimulation (Mishra et al., 2010), transcranial direct current stimulation (Boggio et al., 2010), and deep brain stimulation (Voges et al., 2013) have been shown to be effective in reducing craving. Combined with recent evidence that hyper-precision of priors may be altered with brain stimulation (Avenanti et al., 2018; H. H. Barrett et al., 2016; Riva et al., 2021), our framework provides an explanation at the algorithmic level for these findings. Our model also provides justification for novel neuropharmacological interventions, especially dopaminergic drugs (Rossi et al., 2020), since dopamine has long been theorized to influence precision of Bayesian beliefs (Adams et al., 2013; FitzGerald et al., 2015; Friston et al., 2014). However, further empirical work would be required to determine the exact effects of these neuromodulatory approaches, including directionality (hyper- or hypo-precision), specificity to beliefs about craving, and long-term efficacy. Additionally, a detailed comparison of the different intervention alternatives (e.g., brain-based, mindfulness, psychotherapy) will be necessary to compare their relative effectiveness in various clinical contexts, groups, and levels of severity.

Conclusion

We hope that this new view of craving is useful for refining and clarifying the role of craving in behavioral addictions, which remains poorly characterized in terms of diagnostic and prognostic criteria. Our Bayesian framework proposes unifying substrates of craving common to both SUDs and behavioral addictions, which provides strong explanatory power for the highly similar empirical results shared by the two. Moreover, while SUDs have historically been studied in much detail, behavioral addictions have only recently been defined and acknowledged in the DSM-5. The shared framework we propose for behavioral addictions and SUDs suggests that we may be able to use similar tasks to explore, diagnose, and provide prognostic measures for both diagnostic categories.

Highlights

• Craving exists in both behavioral and drug addictions

• In behavioral addiction, the target of craving is the completion of a certain action

• Craving as a belief about physiological states associated with action completion

• Both prior expectation and new evidence contribute to current craving for an action