The perils of losing control: Why self-control is not just another value-based decision

A widely-observed principle underlying behavior and cognition is that some processes are more automatic (e.g., those that are heavily trained) and others are less automatic (and therefore demand greater cognitive control in order to achieve similar levels of performance) (Posner & Snyder, 1975; Shiffrin & Schneider, 1977). The degree to which a process is automatic or controlled can be determined by several factors, including its speed, flexibility, and susceptibility to interference from other ongoing processes (Botvinick & Cohen, 2015). This well-characterized dichotomy between automaticity and control was encapsulated in one of the earliest and most popular forms of a dual-process theory. Over time, however, dual-process theories have taken on many different forms and have come to encompass an increasingly wide array of increasingly difficult to operationalize dichotomies (e.g., intuitive vs. analytic, System 1 vs. System 2, impulsive/reflexive vs. deliberative/reflective; Evans, 2008; Sloman, 1996). As a result of this process of agglutination and shape-shifting, the dual-process literature has grown branches that now appear untenable, such as its occasional mapping to the distinction between emotion and cognition (Phelps, Lempert, & Sokol-Hessner, 2014). Recent work has pointed to the self-control domain as one particular area that suffers from dual-process overreach (Buckholtz, 2015; Kable & Glimcher, 2007), culminating in a new set of proposals by Berkman and colleagues (2017b; 2017a) suggesting that dual-process theories¹ have no role to play in accounting for the mechanisms of self-control. Berkman and colleagues argue that self-control and self-regulation can instead be fully described by mechanisms of value-based decision-making (VBDM).

Before I describe points of departure with the VBDM account, it is important to note at least two areas where the authors and I are in broad agreement. First, like others (Evans, 2008; Kahneman & Frederick, 2009), I agree that we do not have one system for automatic processes and one for controlled processes, and I also agree that it is premature to conclude that there are dedicated systems for particular types of outcomes, for instance those that are future-oriented versus present-oriented (cf. Fudenberg & Levine, 2006; Thaler & Shefrin, 1981). Second, I also agree with the authors that the most successful strategies for achieving one’s long-term goals avoid relying on control as much as possible, for instance by avoiding situations that are rife with temptations (Duckworth, Gendler, & Gross, 2016; Fujita, 2011) or by developing habits that align with those goals (Wood & Rünger, 2016). However, these are successful precisely because they avoid or minimize control, so I will also try to argue that the term ‘self-control’ may be a misnomer for such strategies.

Like others, I view self-control as a special case of cognitive control and I view the process of allocating cognitive control as a form of value-based decision-making (Botvinick & Braver, 2015; Kool, Shenhav, & Botvinick, 2017; Shenhav, Botvinick, & Cohen, 2013; Shenhav et al., 2017; Westbrook & Braver, 2015). The VBDM account is founded on this view of self-control as a value-based decision, but it takes this theory one step further. According to this account, self-control is not just another decision about control – rather, it is just another value-based decision (cf. Buckholtz, 2015). In other words, Berkman and colleagues conclude that one general form of VBDM captures decisions about control as well as any other potential decision targets. While this proposal offer a valuable counterpoint to decades of theorizing and experimentation, here I examine the assumptions that support the author’s conclusion and suggest three points of clarification. I conclude that caution is warranted before casting self-control as just another value-based decision.

Clarification 1: Dual-process theories are about process not outcome

The VBDM account is proposed as an alternative to ‘hot’/’cold’ dual-process accounts of self-control, which the authors suggest are rigidly tied to the outcome of a particular choice (e.g., unhealthy vs. healthy food). For instance, the authors argue that “when a person with a dieting goal (‘cold’ process) is tempted by an unhealthy snack (‘hot’ process), dual-process models focus on the strength of the hot process and fatigue of the cold one” (2017b, p. 10), and “if a hungry participant does not eat a tempting food then his behavior is attributed to effective self-control, but if he eats it this is attributed to poor self-control and/or excessive impulsiveness” (2017a, p. 7). They go on to note that this property of dual-process accounts “ignores fluctuations in the goal’s value from choice anomalies and other dynamic processes, such as when framing alters an option’s salient attributes” (2017b, p. 10).

However, this set of arguments only applies to a narrow subset of the dual-process literature rather than to dual-process accounts writ large. The dual-process theories Berkman and colleagues refer to are tied to the distinction between automaticity and control, and this distinction depends critically on factors related to the strength and speed of information processing, most notably how susceptible a given task is to interference from other processes. For instance, naming the ink color of the word GREEN is affected by the content of the word itself, but reading the word is not so affected by the ink color (Cohen, Dunbar, & Mcclelland, 1990; MacLeod, 1991). The distinction between automaticity and control is therefore agnostic to the choice outcome. Just as food choice is influenced by the state and trait variables mentioned above, color-naming can be made less control-demanding with training (MacLeod & Dunbar, 1988) and word-reading produces less interference if the words are presented in an unfamiliar language (Dyer, 1971) or if multiple unrelated words are presented nearby, “diluting” attention to the content of the target word (Kahneman & Treisman, 1984) (Figure 1).

Figure 1.

A Stroop task can be stripped of its signature control-demanding properties by manipulating properties of the stimuli (e.g., blurriness) and/or the participant (e.g., their literacy, language fluency, or training). These manipulations can render correct choices (e.g., “red” in this example) undiagnostic of one’s ability to overcome a prepotent bias. Similarly, properties of food choice can make it difficult to diagnose whether control should be expected to play any role in guiding one’s behavior. These include factors related to the salience of the food’s tastiness (e.g., food concreteness) and healthiness (e.g., intrinsic preference for healthy foods), as well as whether a participant views tastiness as a distracting feature to be overridden (something that is unambiguous for word content in the Stroop task).

Properly grounded in the distinction between automaticity and control, these dual-process accounts therefore avoid the forms of outcome-focused rigidity that the VBDM account was designed to overcome. To the contrary, models of controlled processing exploit information about task context and the relative saliency and goal values associated with each stimulus feature to predict how fast and accurate an individual will be when responding to those stimuli (Botvinick, Braver, Barch, Carter, & Cohen, 2001; Cohen et al., 1990; Collins & Frank, 2013; Hübner, Steinhauser, & Lehle, 2010; Wiecki & Frank, 2013). However, as discussed next, the specific approach each model takes to generating those predictions does little to constrain theorizing about control.

Clarification 2: Process models of value-based choice do not constrain representation

The VBDM account proposes that self-control is best captured by a value-based accumulation process rather than by a process that involves control. However, this argument has two shortcomings. First, it conflates process and outcome. For instance, the authors suggest that “self-regulation is less the result of a battle between ‘hot’ impulses and ‘cold’ control…than it is an integration of value inputs” (2017a, p. 19); or, put another way, “self-control operates as a valuation process rather than a battle between different systems” (2017b, p. 8). But this is rather like saying that World War II operated as a process of competing strategies rather than a battle between Axis and Allies. Not only can accumulator models and valuation play nicely with control, they have already done so in important ways. Accumulator models formed the basis for classic computational models of control (Botvinick & Cohen, 2015; Hazy, Frank, & O’Reilly, 2007; Ratcliff & Frank, 2012), and they now sit at the core of value-based models of control allocation (Frank & Badre, 2012; Lieder, Shenhav, Musslick, & Griffiths, in revision; Musslick, Shenhav, Botvinick, & Cohen, 2015).

Second, the authors seek parsimony by proposing that the ventromedial prefrontal cortex (vmPFC) governs the value-based decisions about self-control (Kable & Glimcher, 2007; McGuire & Kable, 2013), just as this region governs evaluations for any other kind of decision. However, this assumption that vmPFC is responsible for all value-based decisions is at odds with available data. A growing body of evidence suggests that value-based decision-making is instead distributed across separate neural circuits (Cisek, 2012; Haber & Knutson, 2010; Rushworth, Kolling, Sallet, & Mars, 2012) – these circuits specialize in accumulating value-based information for different types of decision target, for instance choices between stimuli versus choices between actions or control states (Daw & O’Doherty, 2013; Padoa-Schioppa, 2007; Rushworth, Noonan, Boorman, Walton, & Behrens, 2011; Shenhav et al., 2013; van der Meer, Kurth-Nelson, & Redish, 2012). In fact, despite recent debates regarding the function(s) of dorsal anterior cingulate cortex (dACC: Kolling et al., 2016; Shenhav, Cohen, & Botvinick, 2016), there is broad agreement that the target of dACC evaluations is different than the target of evaluations that occur within vmPFC (Fellows, 2011; Rushworth et al., 2012; Wallis & Kennerley, 2011).

Having a valuation system for control, separate from a system that evaluates stimuli in the environment (e.g., food options) also addresses two mechanistic gaps in Berkman and colleagues’ theory. First, their account quite reasonably assumes that we weigh the values of our options according to how much attention is paid to each of the attributes of those options (e.g., the healthiness versus tastiness of a food). This begs the question of how this attention gets allocated in the first place. The fact that people are able to allocate attention in a deliberate and value-sensitive manner (Krebs, Boehler, & Woldorff, 2010; Padmala & Pessoa, 2011) suggests that attention must itself be partly determined by a value-based decision (Shenhav et al., 2017; Westbrook & Braver, 2015). It is difficult to imagine how this decision about what to attend (e.g., which food attribute) could be determined by the same evaluative system that is the beneficiary of attentional re-weighting (e.g., the one that determines which food to eat), without risking infinite regress.

The example of attentional allocation exposes a second mechanistic gap in the current account, and with it another reason to favor a separate system for control allocation. The types of value-based accumulators Berkman and colleagues propose are well-suited for selecting between discrete options, like foods or trinkets. However, allocating attention requires more than a single discrete choice – it requires selecting both a target (e.g., an attribute or location) and an intensity (i.e., how strongly to attend) (Shenhav et al., 2013). The intensity of control also tends to be associated with increased feelings of cognitive effort, which can in turn inform the selection of how much control to allocate. The discrete choice accumulator model that Berkman and colleagues envision is not ideally suited for this more complex selection process. However, another system already exists that has had to solve this computational problem, and that is the system that governs motor control. Selecting motor control signals requires specifying both a target (e.g., specific muscle group) and an intensity (i.e., force or vigor) (Manohar et al., 2015; Trommershäuser, Landy, & Maloney, 2008). Motivated in part by this line of reasoning, my colleagues and I have proposed that cognitive control signals may be evaluated and selected within dACC, a region that shares functional and structural properties with more caudal regions of medial prefrontal cortex that are involved in the allocation of motor control (Shenhav et al., 2013; Shenhav et al., 2016).

Together, this suggests that self-control could indeed be the product of a value-based evidence accumulation process, without being just like any other value-based decision. In other words, the valuation process could be maximally general and housed in vmPFC, or it could be housed elsewhere and dedicated to control (or even self-control in particular). However, as I explain in the final section, whether or not such a dedicated system exists, we still wouldn’t want to blame it for all apparent failures of self-control.

Clarification 3: Apparent failures of self-control do not necessitate real failures of control

As suggested earlier, control is defined by process, not outcome. Ignoring this critical distinction would lead one to conclude that a pre-literate child has superior control to a literate one based on their Stroop performance alone. Applied to the self-control domain, this means that choices that appear impulsive can arise through processes that circumvent control. For instance, just as people vary in their preferences for sweet and salty, they can also vary in how much weight they come to place on the health and taste values of a food. These preferences could have been shaped by concern for one’s long-term health or by the properties of the foods themselves (e.g., the freshness of healthy foods), but the point is that they have taken shape prior to encountering a food choice. These naturally varying preferences for health and taste can be integrated into the evaluation process to produce a choice that looks more or less self-controlled without involving inhibitory control, consistent with the VBDM account. But the fact that such choices can circumvent control doesn’t mean that they always or even typically do.

Berkman and colleagues primarily focus on dietary choices to make a case for their VBDM model, but there are differences between how such choices are made in the real world versus the laboratory, and these differences may make the lab analogs demand less control (Figure 1). For instance, lab dietary choices are multi-shot, delayed, abstract (e.g., they lack the relevant smells), and only affect a single intake of calories (i.e., have a somewhat negligible effect on an individual’s long-term health). As a result, such ersatz self-control manipulations can produce choices that look impulsive or self-controlled without requiring the participant to recruit control.

The VBDM account suggests that a general attention-weighted decision-making process is sufficient to describe all choices that we think of as self-controlled – and that may be the case –but the evidence these authors present only argues that inhibitory forms of control are by themselves insufficient to explain all choices that prima facie appear to involve self-control. This evidence must also contend with positive evidence for inhibitory control involvement in mediating successful self-control (Heatherton & Wagner, 2011; Kober & Mell, 2015; Kool, McGuire, Wang, & Botvinick, 2013; Ochsner, Silvers, & Buhle, 2012). For instance, putatively inhibitory interactions between lateral PFC and ventral striatum have been shown to mediate the regulation of cravings (Hall, 2016; Kober et al., 2010; Volkow et al., 2010), and individual differences in ventral striatal excitation and LPFC inhibition predict successful self-control outside of the lab (Demos, Heatherton, & Kelley, 2012; Jansen et al., 2013; Lopez et al., 2017; Lopez, Hofmann, Wagner, Kelley, & Heatherton, 2014; Lopez, Milyavskaya, Hofmann, & Heatherton, 2016; Rapuano et al., 2017). Such out-of-sample evidence suggests that inhibitory control is in fact important for real-world self-control, even though it is not sufficient nor is it necessarily the most effective tool (Duckworth et al., 2016).

Berkman and colleagues conclude that “cognitive control networks…contribute to self-control by influencing the weights given to different attributes in the value integration process, rather than by inhibiting other regions” (2017b, p. 8, emphasis added). The work discussed in this section modifies this argument in two ways – it suggests that these two roles for control are not mutually exclusive, and that self-control is shaped by both.

Conclusion

The VBDM account joins a series of important recent challenges to dual-process theory, offering an alternative model whereby control is like any other value-based decision. However, this challenge may be premature. It addresses a problem that dual-process theory already overcomes, using a decision model that is compatible with (rather than an alternative to) dual-process, in order to account for phenomena that may only involve self-control in name, but not process. It is still possible that self-control and self-regulation will be best understood without reference to the dual-process distinction, but I hope the clarifications above ensure that the meanings of control and regulation are not lost along the way.