4 DISCUSSION
Hare et al. (2009) were among the first to provide empirical evidence on the neural correlates of self-control. Since then, this seminal study has had profound impact on theory and empirical work across multiple fields, but it has never been directly replicated. We performed a preregistered, direct replication of this experiment with two goals: (1) to further strengthen the evidence base for self-control theory and research, and (2) to test the robustness of the original results across analytical choices. The results of the four key hypothesis tests are summarized in Table 2.
TABLE 2. Hypothesis test overview| Hypothesis (quoted from Hare et al., 2009, p. 646) | Replication findings |
|---|
- [Activity] in vmPFC should be correlated with participants' goal values regardless of whether or not they exercise self-control
| Supported |
- [A]ctivity in the vmPFC should reflect the health ratings in the SC group but not in the NSC group.
| Supported, with reservations |
- [T]he dlPFC should be more active during successful than failed self-control trials.
| Not supported |
- dlPFC and vmPFC should exhibit functional connectivity during self-control trials.
| Mixed evidence |
- Abbreviations: dlPFC, dorsolateral prefrontal cortex; NSC, non-self-controllers; SC, self-controllers; vmPFC, ventromedial prefrontal cortex.
Our data provide further support for the now widely accepted notion that decisions are associated with a value signal in vmPFC, which integrates relevant choice attributes to inform a final decision (Hypotheses 1 and 2; Table 2). Specifically, like Hare et al. (2009), we found positive correlations between participants' goal values (choices for food items) and activity within vmPFC, regardless of whether participants exercised self-control. We were also able to replicate findings which were reported in the original study in support of the idea that vmPFC prioritizes choice attributes that are consistent with each individual's subjective values. Specifically, as in the original study, activity in vmPFC was associated with the perceived healthiness of food items in participants who were relatively more successful at exercising self-control in the experimental task but not in participants who were relatively less successful. However, we did not find evidence of significant differences between the two groups. Overall, these results are in line with a broader set of literature in neuroeconomics, which has described the role of vmPFC in valuation across diverse types of stimuli (e.g., money, consumer goods, etc., for a review see (Bartra et al., 2013)). The present study is the first to provide a direct replication of this effect in the context of food-related decision-making. Thus, this replication study increases the confidence in choice models of self-control which describe self-control as a value-based choice (Berkman et al., 2017).
In addition to the replication of the originally reported analyses, we added several analysis branches to further test the robustness of these results. First, in a follow-up analysis to the whole-brain search for brain regions associated with goal value (Figure 4), Hare et al. (2009) highlight the fact that individual scale points (−2 –2) of goal value are neatly distinguished in a step-wise pattern in their vmPFC ROI, suggesting that the ROI can be used to precisely distinguish and predict choices. However, the original analysis approach was optimized to demonstrate this effect and requires individual-level choice data to identify individual peak-voxels within a larger vmPFC ROI. In addition, this analysis supports the limited conclusion that, on average, most study participants show this step-wise encoding of goal value in at least one voxel within a larger vmPFC area. We added an alternative analysis approach by averaging signal extracted from all voxels within the vmPFC ROI in which activity was associated with goal value in our replication sample. We show that the step-wise encoding of choice behavior is largely preserved in this more general analysis, but that the effect size is substantially smaller. Similarly, when examining relationships between health and taste ratings and average signal within vmPFC, we do not find significant encoding of health ratings in the SC group despite the relatively large size of this replication sample. In other words, future studies that are interested in reusing these vmPFC ROIs as indicators of goal value without the luxury of an individual-level localizer task that allows them to identify individual peak voxels per person likely require a much larger sample to be appropriately powered than implied by the original publication.
Further, next to vmPFC and in contrast to the original study, we identified positive associations between goal value and activity in clusters within the striatum at a relatively lenient statistical threshold (p < .001, uncorrected) used in the original study. This discovery is likely a function of the increased power in the larger replication sample and largely in line with the neuroeconomics literature on subjective valuation which regularly identifies clusters in both vmPFC and striatum (Bartra et al., 2013). Following up on this finding, we found some evidence of differentiation between individual levels of goal value, even within our caudate ROI when applying the optimized analysis procedure reported in the original study. This adds to the findings in prior work suggesting that vmPFC is not the exclusive locus of goal value representation in the human brain.
We did not find strong evidence in support of the second set of hypotheses (Hypotheses 3 and 4, Table 2) proposed by Hare et al. (2009), which highlight the role of left dlPFC in self-control. First, we examined average activity levels in left dlPFC. Even though there were clear (and replicated) behavioral differences between participants who were relatively more and those who were relatively less successful at exercising self-control in the scanner task, we did not find hypothesized, statistically significant group differences in dlPFC activity during successful self-control trials in a whole-brain analysis. Instead, we observed relative deactivation across multiple brain regions in NSC relative to SC, including, but not limited to, areas that are involved in processing of subjective values such as vmPFC. One possible alternative hypothesis supported by our data thus is that SC do not rely on more intensive executive processing indicated by higher dlPFC activity to downregulate subjective value in self-control situations, they simply perceive less intensive subjective value for “tempting” food items to begin with. Another alternative explanation is that this null finding is due to power limitations in our data, given that only 15 participants (compared to 19 in the original sample) qualified as SC. In other words, there is a possibility that positive activations in dlPFC during self-control are simply more subtle than the resulting deactivation in value-related areas. Although we cannot conclusively disentangle these contradictory ideas, note that we exclusively found negative (although nonsignificant) coefficients within dlPFC in this sample.
Next, we followed procedures reported by Hare et al. (2009) to examine the role of dlPFC in self-control in terms of its functional connectivity with brain activity in vmPFC. Since we were unable to identify a functionally defined dlPFC cluster in which average activity was involved in self-control in the replication sample, we relied on a meta-analytically defined map from www.neurosynth.org (Yarkoni et al., 2011) associated with the term “self-control” and intersected it with an anatomical, left dlPFC mask. Our analyses which fully replicated the original work by focusing exclusively on processes in participants who were relatively more successful SC during the scanner task did not replicate the original findings which suggested a negative indirect relationship between dlPFC and vmPFC activity through IFG/BA46 during self-control. We followed up on this null-result by rerunning the PPI on the full sample of participants who exercised any self-control in the scanner task (N = 59) to address concerns about statistical power. This path was chosen given the absence of strong theoretical arguments that the mechanisms that drive successful self-control differ qualitatively (rather than just in intensity) between people who are successful more often and those who are successful relatively less often. Indeed, in this larger sample, we do find some evidence of replication. Stronger still, we found evidence of direct, negative correlations between activity within our meta-analytic left dlPFC seed and an area within vmPFC, which was hypothesized, but not found by Hare et al. (2009). It is important to note, however, that we simultaneously found evidence for unexpected positive associations between activity in the left dlPFC ROI and another, more dorsal MPFC cluster. Of note here is that the whole-brain table for this analysis in the original publication revealed a similar positive association with an MPFC cluster in almost the exact same location (see Figure 11). While there was (minimal) overlap between the unexpected MPFC cluster that showed positive functional connectivity with left dlPFC and the vmPFC ROI that was associated with goal value in our sample, we did not find such overlap between the vmPFC cluster that showed the hypothesized negative association with dlPFC. In other words, the first PPI, at best, provides mixed evidence regarding the nature of the relationship between dlPFC and vmPFC activity during self-control. Hare et al. (2009) proceeded to follow-up on the lack of a negative direct association between dlPFC and vmPFC in their first PPI by identifying a cluster in BA46 that was negatively associated with dlPFC as the seed region for a second PPI. Following this analysis approach, we were able to replicate the original findings, identifying a cluster in vmPFC that was positively associated the BA46 seed identified in PP1 based on the full replication sample (N=59) and thus indirectly negatively associated with the meta-analytic dlPFC ROI. In sum, our replication data provides mixed evidence with regards to Hypothesis 4 regarding a negative relationship between dlPFC and dlPFC activity during self-control.
These mixed results highlight the need for additional work to fully understand the role of the dlPFC in food-related decision-making and in theories of self-control more generally. Overall, our findings are most in line with a conceptualization of self-control as a simple form of value-based decision-making in which different choice attributes (here health and taste considerations) are encoded and integrated in vmPFC according to subjective values of the decision-maker (Berkman et al., 2017). This contrasts with the model that the findings of Hare gave rise to, wherein longer-term goals (here health considerations) required dlPFC involvement in order to be effectively integrated into subjective value judgments (Hare et al., 2009).
A frequently voiced explanation for failed replications is that the (cultural) context differed between the original and replication study (Zwaan et al., 2018). In our case, the original study was performed in the United States before 2009 and the replication in the Netherlands, approximately 10 years later. Thus far, we are not aware of any strong theoretical or empirical claims that the brains or fundamental psychological processes surrounding self-control of US subjects are different from those of Dutch study participants or that the basic neural processes of valuation and self-control have changed over the past decade. However, what could differ between US and Dutch individuals and what could have changed over the past decade is the role of food and dieting in society, and more specifically, to what extent food choices can generate a self-control conflict and how people cope with that. This may—in theory—influence the way in which people respond to the task and stimuli. Naturally, for a self-control dilemma to occur one should have the goal to diet or eat healthy. It could be argued that stronger goal commitment may strengthen attempts of overruling impulses and therefore amplify control-related responses. Observational studies showed that the prevalence of dieting is higher in Europe than in the United States (Santos et al., 2017) and a large proportion of the Dutch population self-reports to diet or actively restrain their food intake (de Ridder et al., 2014). This would speak against this being an explanation for the null finding. It should however be noted that self-reports of dieting and dietary restraint have been shown to be unrelated or weakly related to actual intake (de Ridder et al., 2014; Stice et al., 2004) which casts doubt on this measure being a reliable proxy of goal strength. We cannot rule out but we also cannot support that goal commitment was stronger for the successful SC in the original study compared to the current replication study.
Another important conclusion from this project is that analytical flexibility can influence fMRI results. Specifically, for H1 and H2 we presented two sets of results produced using two different analysis strategies. While the overall patterns of results remained similar, increasing confidence in the directionality of effects, effect sizes differed significantly. This has important implications for follow-up research which may rely on existing work for power calculations. Previous work has shown that not only analytical flexibility but also different preprocessing approaches to the fMRI data (e.g., different software packages and varying parameters) may affect task-based fMRI results (Bowring et al., 2022; Mikl et al., 2008; Triana et al., 2020). In this replication study we employed a state-of-the-art, standardized, and optimized preprocessing pipeline provided by fMRIprep, which was not available to the authors of the original study (Esteban, Markiewicz, et al., 2018). As much as possible, we chose parameters similar to those used in the original study (e.g., the same smoothing kernel). Though submitting the data through different preprocessing pipelines was outside of the scope of the current study, we acknowledge that doing so could potentially further inform the field about the (in)variability of individual results to specific choices made by the researchers. Unpreprocessed data for this project is available on OpenNeuro and would support such an investigation for those interested.
4.1 Impact on theory
Our findings are relevant for future theorizing on self-control. Specifically, this replication data set supports the conceptualization of self-control as either a very simple form of value-based decision-making (Berkman et al., 2017) or as automatic “effortless” self-control (Gillebaart & de Ridder, 2015) rather than a dual-system which involves conscious effortful control.
In psychology, self-control has traditionally been explained with dual-system theories (e.g., Hofmann et al., 2008; Metcalfe & Mischel, 1999). These theories are characterized by the notion of two (competing) systems for processing information, namely a “hot”/automatic/impulsive system and a “cold”/rational/reflective system. According to these dual-system models, self-control is successful when the impulses arising from the “hot” system are overcome and, consequently, behavior is in line with long-term goals. In this traditional approach, the dilemma first must be identified and, subsequently, effortful and conscious inhibition is required to overcome it (Fujita, 2011). A neurobiological parallel to these dual-system models has been proposed in which self-control involves a balance between brain regions representing the reward, salience and emotional value of a stimulus and prefrontal regions associated with (effortful) inhibition and cognitive control (Heatherton & Wagner, 2011). In this traditional perspective, effortful and conscious impulse inhibition is a necessary or defining feature of (successful) self-control.
A major criticism of this traditional perspective is that successful self-control does not always require effortful inhibition or conscious control. It has been proposed that there are many different routes to self-control, only some of which involve effortful inhibition (Fujita, 2011). Research has indicated that people can automate goal-striving behaviors in response to contextual cues (Bargh et al., 2001; Chartrand & Bargh, 1996). For instance, providing cues related to the long-term goal (e.g., dieting cues) promotes goal-congruent choices through goal priming (Fishbach et al., 2003; Papies, 2016; Van der Laan et al., 2017), which is thought to occur without requiring conscious deliberation or effort. Further, by systematically repeating (healthy) behaviors (healthy) habits can be created. It has been shown that successful SC do not necessarily exert more effort; they perform healthy behaviors automatical because of healthy habits (Galla & Duckworth, 2015; Gillebaart, 2018).
This has led to alternative conceptualizations of self-control which do not include or at least attenuate the role of effortful inhibition. As mentioned, recently, successful self-control has been conceptualized as being at least partly an automatic process in which responses to environmental cues that are routinized (or automatically triggered) in the direction that is in line with their long-term goals (Fujita, 2011; Gillebaart, 2018). A second theory, which recently has gained more traction, is to consider self-control as a simple value-based choice (Berkman et al., 2017). Value-based decision-making involves choosing an option from a set based on its relative subjective value. This process involves calculating a value for each option by evaluating various attributes—gains (e.g., improved health) and costs (e.g., less food enjoyment), assigning weights to these attributes, and enacting the most valued option. It should be noted that this is a dynamic process. That is, the weight of each attribute is sensitive to attentional shifts (e.g., being explicitly guided toward certain attributes like health), contextual effects and framing of the choice set. Within this conceptualization of self-control, there is nothing special about long-term goals: attributes related to short- and long-term goals treated similar in this equation though the relative weights may be different based on the aforementioned factors. This discussion in psychology intersects with the ongoing debate in decision neuroscience and temporal discounting where Kable and Glimcher (2007) suggested there is one common valuation in vmPFC while McClure et al. (2004) suggested that separate neural systems encode value for immediate versus longer-term attributes.
The study of Hare conceptualizes self-control as a value-based decision (H1, H2) but in line with traditional dual-system models it still posi that there are dual motives and that the future part is “special”: integrating longer-term considerations into the value system, that is, changing the weight of long-term attributes, requires involvement from control-related areas (i.e., the dlPFC; H3, H4). Their hypothesis about the role of the dlPFC had its basis in the role of dlPFC in cognitive control and emotion regulation. The authors speculated that vmPFC originally evolved to predict the short-term value of stimuli and that humans developed the ability to incorporate long-term considerations into values by giving structures such as the dlPFC the ability to modulate this value.
Our mixed findings regarding dlPFC involvement highlight the need for more research to understand the role of dlPFC in assigning weight to these longer-term consequences. The replication results rather point to the conceptualization of self-control as either automatic and “effortless” or as a (simple) form of value-based decision-making. At a minimum, our results support the idea that that it is not the dlPFC that is responsible for increasing the weight of the longer-term attributes into the choice. In support of the latter: when comparing successful to unsuccessful trials that required self-control in all participants, we observed a deactivation of vmPFC, which suggests that successful self-control in this sample may be driven by a weaker subjective value for a given food item rather than by more intensive control driven by dlPFC. The finding, that in successful SC, vmPFC reflects health ratings, even though dlPFC is not active, suggests that dlPFC activation is not needed to incorporate health into the vmPFC value signal. Thus indeed, in line with the proposition of self-control as a simple form of value-based decision-making (Berkman et al., 2017), decisions may just be the result of multiple single value-calculations.
Posts
