Skip to main content
NCBI home page
Social Cognitive and Affective Neuroscience logoLink to Social Cognitive and Affective Neuroscience
. 2019 Oct 8;14(9):957–966. doi: 10.1093/scan/nsz068

Comparing two neurocognitive models of self-control during dietary decisions

Danielle Cosme 1,#,✉, Rita M Ludwig 1,#, Elliot T Berkman 1
PMCID: PMC6917023  PMID: 31593247

Abstract

Self-control is the process of favoring abstract, distal goals over concrete, proximal goals during decision-making and is an important factor in health and well-being. We directly compare two prominent neurocognitive models of human self-control with the goal of identifying which, if either, best describes behavioral and neural data of dietary decisions in a large sample of overweight and obese adults motivated to eat more healthfully. We extracted trial-by-trial estimates of neural activity during incentive-compatible choice from three brain regions implicated in self-control, dorsolateral prefrontal cortex, ventral striatum and ventromedial prefrontal cortex and assessed evidence for the dual-process and value-based choice models of self-control using multilevel modeling. Model comparison tests revealed that the value-based choice model outperformed the dual-process model and best fit the observed data. These results advance scientific knowledge of the neurobiological mechanisms underlying self-control-relevant decision-making and are consistent with a value-based choice model of self-control.

Keywords: self-control, dual-process, value-based choice, health, decision neuroscience

Introduction

Self-control is central to health and well-being and thus is a promising target for preventive interventions. Designing effective interventions requires a precise neurocognitive model self-control, but identifying a working model of self-control that maps onto underlying brain systems has proven challenging (Fujita et al., 2018). One barrier to progress is that neural models of self-control are almost always tested in isolation, not directly compared to one another. Here, we define self-control as the process of favoring abstract, distal goals over concrete, proximal goals during decision-making (Fujita, 2011) and compare two prominent neurocognitive models of self-control with the goal of identifying which, if either, best describes behavioral and neural data of dietary self-control-relevant decisions. The overarching purpose of this work is to develop a more-refined neurocognitive model of self-control to enable translational interventions to improve health outcomes.

We focus on two prominent models of self-control: dual-process models and value-based choice models. There are many variations on each class of models within the fields of social psychology and neuroeconomics; we draw hypotheses from the features that are most common within each model family. As such, it should be noted that the statistical models we compare in this paper are just several of many possible model instantiations. In general, dual-process models describe self-control as a battle between ‘hot’ affective states (e.g. craving) and ‘cold’ cognitive states (e.g. inhibitory control). In these models, self-control outcomes are the product of an antagonistic, seesaw relationship between affect and cognitive control (Kotabe and Hofmann, 2015). Though the precise neural location of these states differs substantially across studies (Berkman, 2017), dorsolateral prefrontal cortex (dlPFC) is posited as a key node for flexibly maintaining goal representations (Braver et al., 2009) and implementing cognitive control across a variety of cognitively demanding tasks (MacDonald et al., 2000; Miller and Cohen, 2001; Duncan, 2013; Shenhav et al., 2013; Buhle et al., 2014). On the other hand, regions in the mesolimbic dopamine system, such as ventral striatum (VS), are posited as key nodes for reward processing (McClure et al., 2007). According to these models, ‘cold’ regions inhibit activity in ‘hot’ regions (Figner et al., 2010), and therefore the balance, or difference, between lateral and subcortical activity should predict self-control-relevant outcomes (Heatherton and Wagner, 2011; Lopez et al., 2014; Lopez et al., 2017). This is because these regions are hypothesized to serve opposing functions during self-control-relevant choices.

In value-based choice models, subjective values from a range of choice attributes are integrated into a single accumulated expected value (or utility) for each choice (Berkman et al., 2017a). Attributes are not limited to just ‘hot’ and ‘cold’—they contribute separately to an accumulation process rather than compete with each other directly. Evidence for choice options is accumulated over time until one choice reaches the threshold for enactment, determining behavior and therefore self-control-relevant outcomes (Harris et al., 2018; Tusche and Hutcherson, 2018). Although it’s possible to classify outcomes as self-control ‘successes’ or ‘failures’ (c.f. Hare et al., 2009), value-based choice models often focus on the weights assigned to choice attributes (e.g. taste and health) during decision-making rather than the outcomes per se. This is because choices are a combination of both signal and noise, but with enough choices available, it is possible to estimate parameters for relevant choice attributes that are more stable and drive behavior in the long run. For example, even if an individual ‘fails’ to exert self-control by choosing to eat an unhealthy snack, that specific choice may occur in the presence of a relative shift in the value of healthiness or tastiness, which may accumulate over time to impact the probability of selecting healthy foods overall.

In value-based choice models, dlPFC and VS are also implicated as important nodes during dietary decisions, but their relationship is not expected to be antagonistic. Instead, these regions are hypothesized to independently encode the subjective value (i.e. goal value) derived from different choice attributes relevant to the decision-making process (e.g. health and taste of food options), which are ultimately integrated in ventromedial prefrontal cortex (vmPFC). Although dlPFC is expected to contain information about the subjective value of health while individuals with goals to eat healthfully make dietary decisions, it is unclear whether dlPFC computes goal values, implements cognitive control to modulate the value of health or plays some other role (Plassman et al., 2010; Shenhav, 2017). In any case, within the value-based choice framework, (a) dlPFC is expected to show increased activation to healthy foods, (b) VS is expected to show increased activation to tasty, unhealthy foods and (c) both regions are expected to positively correlate with vmPFC. Notably, options that are tasty and healthy are expected to elicit activity in both dlPFC and VS. In contrast to dual-process models, which predict that the balance between dlPFC and VS predicts self-control, value-based choice models theorize that self-control-relevant choices should be recoverable from activity in vmPFC. Although it remains unclear what precise role vmPFC plays in decision-making and whether it may serve as a final common pathway for choice, it has been implicated in choice regardless of stimulus type (e.g. food or money) or motivation (Rangel and Hare, 2010; Hare et al., 2010; Hare et al., 2009; for meta-analyses, see Clithero and Rangel, 2013; Bartra et al., 2013) and is thought to serve functions such as contributing to the integration or comparison of value signals (Lim et al., 2011; Padoa-Schioppa and Conen, 2017; Levy and Glimcher, 2012) or alternatively, the construction of integrated meaning of the self in context (Roy et al., 2012) or situational processing (Lieberman et al., 2019). Consequently, value-based choice models suggest that the choice that evokes the greatest activation in vmPFC is likely to be enacted, and activation in vmPFC is driven by inputs from regions such as dlPFC, VS and potentially their interaction. However, in contrast to dual-process models, which imply that lateral prefrontal and subcortical regions should be negatively correlated during decision-making, value-based choice models posit that there is not necessarily a directional relationship between neural activity in brain regions representing the subjective value of choice attributes, which are subsequently integrated in vmPFC (i.e. these value signal inputs might be positively, negatively or not related).

Despite these different predictions about how neural activity in dlPFC, VS and vmPFC relate to self-control-relevant decisions, these predictions are rarely compared within the same study. Researchers typically test the neural predictions of dual-process (Lopez et al., 2017) and value-based choice models (Hare et al., 2009; Hare et al., 2011) separately, with notably few attempts to directly compare the models (Hutcherson et al., 2012). This may be because traditional univariate analyses are not well suited to model comparison.

Here, we employ a trial-by-trial modeling approach that allows us to relate neural activity to self-control-relevant decisions in the dieting domain on a finer time scale than has previously been possible and provides relative model fit indices. We recruited a large sample of overweight and obese participants (N = 94; BMI M = 31.33, BMI SD = 3.95) who had explicit dieting goals and examined neural responses in three brain regions of interest (ROIs)—dlPFC, VS and vmPFC—while they made real dietary decisions about healthy and unhealthy snack foods during a food auction task (Hutcherson et al., 2012). Because participants enrolled in this study to improve healthy eating habits, there is an implicit self control dilemma between distal health goals and proximal hedonic goals during the task. On each trial, we extracted the average blood-oxygenation-level-dependent (BOLD) signal within each ROI, regressed bid value on these BOLD signal estimates and assessed the relative evidence for each model of self-control using multilevel modeling. Dual-process models predict an association between the balance between activity in dlPFC and VS with bid value, whereas value-based choice models predict that vmPFC should be associated with bid value, even when dlPFC and VS are included in the statistical model. Our primary hypotheses and related analytic decisions were preregistered and are available in Supplementary material and online: https://osf.io/8bvxe/registrations.

Methods

Participants

Participants were 94 adults aged 35–46, (77 females, 16 males, 1 declined to respond; age M = 39.2, age SD = 3.5) with goals to eat healthfully, recruited as part of a 6-month longitudinal intervention study to improve healthy eating habits during middle age. The present study is a secondary analysis of the data. The sample size was determined based on the power analysis accompanying the original grant application (7R21CA175241–03) and constrained by the grant budget and award period. The data were collected before participants were randomized into intervention groups. To ensure all participants shared a healthy eating goal, only interested participants who endorsed readiness to change their eating habits were enrolled. We excluded four participants: two due to excessive motion, one due to technical failure and one due to an incidental finding, yielding a total of 90 participants for statistical analyses. This study was approved by the University of Oregon Institutional Review Board; all participants gave written informed consent and were compensated for their participation.

Food auction task

To measure individual subjective value of healthy and unhealthy snack foods, participants completed a willingness-to-pay task (Hutcherson et al., 2012; http://github.com/UOSAN/WTP/tree/chives) while undergoing functional neuroimaging. The task is an incentive-compatible economic auction in which participants view images of thirty healthy and thirty unhealthy snack foods and choose how much they are willing to pay for each item. Foods that are energy-dense and high-sugar or contain processed or red meat were classified as unhealthy, whereas foods that are not energy-dense and are high-fiber, low-fat and low-sugar were classified as healthy. Participants were endowed with $2.00 to buy a snack and were told that one trial would be randomly selected and enacted. Bids greater than or equal to a randomly selected bid resulted in the participant getting the snack, whereas lower bids resulted in participants receiving the money, but not the snack. The optimal strategy is to bid the true amount one is willing to pay for each item. The task utilized an event-related design (Figure 1), and food image order was randomized for each subject.

Fig. 1.

Fig. 1

Open in a new tab

Task design. Each trial consisted of a 4 s snack food presentation, followed by a 4 s bid period. Snack foods were either healthy (e.g. carrot sticks) or unhealthy (e.g. candy). All trials ended with a jittered fixation cross presented for 3–6 s (M = 4.38 s).

Neuroimaging data acquisition

Data were acquired using a 3 T Siemens Skyra scanner at the University of Oregon’s Lewis Center for Neuroimaging. High resolution anatomical volumes were acquired using a T1-weighted MP-RAGE pulse sequence and functional volumes were acquired using a T2*-weighted echo-planar sequence (voxel size = 2 mm3). Scan parameters are listed in Supplementary material.

Neuroimaging data preprocessing and analysis

Neuroimaging data were pre-processed and analyzed using SPM12 (Wellcome Department of Cognitive Neurology; http://www.fil.ion.ucl.ac.uk/spm). For each participant, functional images were realigned, coregistered to the high-resolution anatomical image, unwarped to reduce susceptibility artifacts and smoothed using a 2 mm3 FWHM Gaussian smoothing kernel. First-level statistical analyses were conducted in native space. Each trial was entered in the model as a separate regressor (rather than grouped by condition). Trial duration was specified as 8 s from food image onset to fixation (Figure 1). Realignment parameters were transformed into five motion regressors, including absolute displacement from the origin in Euclidean distance and the displacement derivative for both translation and rotation and a single trash regressor for images with motion artifacts (e.g. striping) identified using automated motion assessment (Version v0.2-alpha; Cosme et al., 2018) and visual inspection. These regressors were included as covariates of no interest. Two participants were excluded for having >10% unusable volumes, which was more than three standard deviations from the median (Mdn = 1.57%, SD = 3.21%). The resulting statistical maps for each trial were concatenated to create a beta-series (Rissman et al., 2004). Preprocessing and analysis scripts are available online (https://osf.io/pevmy).

ROI definition and parameter extraction

We defined bilateral ROIs for dlPFC, vmPFC and VS (Figure 2) using the Desikan-Killiany (Desikan et al., 2006) and Destrieux (Destrieux et al., 2010) cortical parcellation atlases and the FreeSurfer segmentation atlas (Fischl et al., 2002) and mapped these ROIs to participants’ T1-weighted anatomical scans using FreeSurfer 6 (Fischl, 2012). To determine which cortical parcels to use, we inspected meta-analytic association test maps from NeuroSynth (Yarkoni et al., 2011) for the following terms: cognitive control, dlPFC, value and vmPFC and identified overlapping FreeSurfer parcels (Supplementary Figure S1). To create the dlPFC ROI, we concatenated the middle frontal gyrus and the inferior frontal sulcus parcels. We created the vmPFC ROI using the medial orbitofrontal cortex parcels and the VS ROI by concatenating the nucleus accumbens and putamen segments. All ROIs were concatenated and binarized using the fslmaths function in FSL 5.0.10 (Jenkinson et al., 2012) and resliced to 2 mm3 using SPM12. This process yielded individually defined dlPFC, vmPFC and VS ROIs for each participant. To calculate the mean BOLD signal across the voxels in each ROI, we use the 3dmaskave function in AFNI 18.2.04 (Cox, 1996). For each participant, we extracted the mean parameter estimate of BOLD signal within each ROI for each trial in the beta-series. To account for differences in variability between individuals and ROIs, parameter estimates were standardized within participant and ROI.

Fig. 2.

Fig. 2

Open in a new tab

ROIs in volumetric space.

Multilevel modeling

Evidence for the dual-process and value-based choice models of self-control was assessed by inspecting parameter estimates from a series of multilevel models. Statistical analyses were conducted in R 3.5.1. (R Core Team, 2018; https://www.r-project.org/) using the lme4 package (Bates et al., 2015). For each theoretical model, we compared a series of nested statistical models in which bid value was the criterion and neural predictors were added to a base model that included only the fixed effect of Food Type (healthy or unhealthy). For all models, only participant intercepts were treated as random effects. We compared nested models using chi-square difference tests; models were treated as significantly improving model fit if P < 0.05. To determine the best fitting model across theoretical models, we inspected the Akaike information criterion (AIC). Because AIC improves as the predictive value of a model increases (Aho et al., 2014), this comparison reveals which of the models maximizes accuracy in predicting bid value (where smaller AIC indicates better prediction accuracy). To estimate multilevel model effect sizes, we calculated R2 according to the guidelines in Lorah (2018). To estimate correlations between ROIs and account for the nested structure of trials within participants, we calculated repeated measures correlations using the rmcorr package (Bakdash and Marusich, 2017) in R. Because these models are only several of many possible ones, we also run additional, non-preregistered models and compared model fit using a specification curve (Simonsohn et al., 2015). The results of this analysis can be found in Supplementary material.

Dual-process model comparison

To characterize the competitive nature of dlPFC and VS posited by the dual-process model, we created a ‘balance’ score (Lopez et al., 2017) by subtracting estimates of activity in VS from dlPFC on each trial. Positive values indicate relatively greater dlPFC activity, whereas negative values indicate relatively greater VS activity. To test dual-process predictions, we compared model fit among the following models. As stated in our preregistration, if DP1 is the best fitting model and the balance score (dlPFC—VS) is significantly associated with bid value, we will interpret this as evidence for the dual-process model of self-control. However, if DP2 is the best fitting model, suggesting that vmPFC is significantly associated with bid value, we will interpret this as evidence for the value-based choice model, which is the only model that predicts a critical role for vmPFC in self-control.

First level models: Base model: Yij (Bid value of trial i by person j) = β0j + β1jFood Typeij + εijDP1: Yij (Bid value of trial i by person j) = β0j + β1jFood Typeij + β2j(dlPFCij − VSij) + εijDP2: Yij (Bid value of trial i by person j) = β0j + β1jFood Typeij + β2j(dlPFCij − VSij) + β3jvmPFCij + εij

Second level equations: β0j = γ00 + μ0j β1j = γ10 (In DP1 and DP2): β2j = γ20(in DP2): β3j = γ30

Value-based choice model comparison

To assess evidence for this theoretical model, we compared the following statistical models. As stated in our preregistration, if VB1—which specifies terms for dlPFC and VS to represent subjective value of relevant choice attributes and vmPFC as the value integrator—is the best fitting model and the neural predictors are significantly associated with bid value, we will interpret this as evidence for the value-based choice model. To mirror the model comparison for the dual-process models, we also planned to test whether adding the balance score (dlPFC—VS) to the model (VB2) would improve model fit. However, this model did not converge because the balance score is a linear combination of dlPFC and VS and was therefore inestimable.

First level equations:Base model: Yij (Bid value of trial i by person j) = β0j + β1j Food Typeij + εijVB1: Yij (Bid value of trial i by person j) = β0j + β1j Food Typeij + β2jdlPFCij + β3jVSij + β4jvmPFCij + εijVB2: Yij (Bid value of trial i by person j) = β0j + β1j Food Typeij + β2jdlPFCij + β3jVSij + β4jvmPFCij + β5j(dlPFCij − VSij) + εij

Second level equations: β0j = γ00 + μ0j β1j = γ10 (In VB1 and VB2): β2j = γ20 (In VB1 and VB2): β3j = γ30 (In VB1 and VB2): β4j = γ40 (In VB2): β5j = γ50

Value integration in vmPFC

The value-based choice model specifies that value signals from dlPFC and VS are integrated in vmPFC. To assess evidence for this hypothesis, we regressed trial-level vmPFC activity on dlPFC and VS activity and their interaction. Participant intercepts were modeled as random effects. We expected that if vmPFC integrates the value signals from dlPFC and VS, then the fixed main effects of each region on vmPFC activity would be significant and positive and that the interaction between these regions also would be significantly associated with vmPFC activity.

First level equation: VMPFC: Yij (vmPFC on trial i by person j) = β0j + β1jdlPFCij + β2jVSij+ β3jdlPFCij ×VSij + εij

Second level equations: β0j = γ00 + μ0jβ1j = γ10β2j = γ20β3j = γ30

Results

Descriptives

Inspection of the data revealed a main effect of the Food Type (healthy vs. unhealthy) on bid value, such that participants were willing to pay more for healthy foods than unhealthy foods (Figure 3; Mhealthy = 0.96, SDhealthy = 0.65; Munhealthy = 0.66, SDunhealthy = 0.63). This is not unexpected given the dieting goals of participants in the sample. Visual inspection of the neural data revealed moderate positive correlations among the ROIs (Table 1). In terms of differential neural activation, healthy foods were associated with greater BOLD signal than unhealthy foods in dlPFC and VS, but not vmPFC (Figure 4). In terms of behavioral responses, higher bid values were associated with increased BOLD signal in all ROIs (Figure 5) and similar trajectories were observed for both healthy and unhealthy foods (Figure 6). The relevant inferential tests are reported in the following section.

Fig. 3.

Fig. 3

Open in a new tab

Mean bid value in dollars for healthy and unhealthy snack foods. Points represent mean bid value for individual participants. Error bars are 95% confidence intervals across trials.

Table 1.

Repeated measures correlations among ROIs

ROI M SD 1 2 3
1. VS 0.82 1.10 --
2. dlPFC 0.79 1.10 0.52 [0.50, 0.54] --
3. vmPFC 0.13 1.09 0.35 [0.33, 0.37] 0.50 [0.48, 0.52] --
Open in a new tab

Note. N = 5220 trials. All correlations are statistically significant, P < 0.001. 95% confidence intervals are bracketed. Correlations adjust for trials nested within participant using multilevel modeling.

Fig. 4.

Fig. 4

Open in a new tab

Mean parameter estimates of BOLD signal as a function of Food Type (healthy or unhealthy) and ROI. Error bars are 95% confidence intervals across trials.

Fig. 5.

Fig. 5

Open in a new tab

Mean parameter estimates of BOLD signal as a function of bid value and ROI. Points are scaled by the number of observations. Error bars are 95% confidence intervals across trials.

Fig. 6.

Fig. 6

Open in a new tab

Mean parameter estimates of BOLD signal as a function of bid value, Food Type (healthy or unhealthy) and ROI. Points are scaled by the number of observations. Error bars are 95% confidence intervals across trials.

Model comparison

In general, results of the multilevel modeling analyses did not support the dual-process hypothesis. Compared to the base model that included only the fixed effect of Food Type, adding the difference term representing the relative activation of dlPFC and VS to the model DP1 did improve model fit as indicated by a statistically significant change in chi-square, χ2(1) = 8.42, P = 0.004 (Table 2 for a summary of model fit results). However, the model that included an additional term for vmPFC activity (DP2) significantly improved fit over the basic dual-process model (DP1), χ2(1) = 31.26, P < 0.001. In DP2 (Table 3), each one standard deviation increase in vmPFC activity was associated with a 5.0 cent increase in bid value (b = 0.05, 95% CI = [0.03, 0.06], P < 0.001), while the difference between dlPFC and VS activity was associated with a 1.8 cent increase for each SD change (b = 0.02, 95% CI = [0.00, 0.03], P = 0.046).

Table 2.

Summary of model comparison results

Model df χ2 (df) AIC Deviance P
Dual-process models
Base model 4 9753.22 9745.22
DP1 5 8.42 (1) 9746.80 9736.80 0.004
DP2 6 31.26 (1) 9717.54 9705.54 <0.001
Value-based choice models
Base model 4 9753.22 9745.22
VB1 7 53.73 (3) 9705.49 9691.49 <0.001
Open in a new tab

Note. Models were compared sequentially in t he order presented here using a change in chi-square difference test. A statistically significant P value (i.e. P < 0.05) indicates that that model fits the data better than the model above. The bolded model is the best fitting model.

Table 3.

Results of the multilevel models DP1, DP2 and VB1

Fixed effects
DP1 b SE t (df) P
Intercept 0.96 0.03 34.08 (106.04) <0.001
Food Type −0.30 0.02 18.43 (5219.69) <0.001
dlPFC—VS 0.02 0.01 2.90 (5305.03) 0.004
Random effects
Variance SD
Participant 0.06 0.24
Fixed effects
DP2 b SE t (df) P
Intercept 0.95 0.03 33.78 (106.18) <0.001
Food Type −0.30 0.02 18.44 (5218.74) <0.001
dlPFC—VS 0.02 0.01 2.00 (5305.17) 0.046
vmPFC 0.05 0.01 5.60 (5305.69) <0.001
Random effects
Variance SD
Participant 0.06 0.24
Fixed effects
VB1 b SE t (df) P
Intercept 0.92 0.03 31.34 (125.43) <0.001
Food Type −0.30 0.02 18.25 (5218.31) <0.001
dlPFC 0.04 0.01 3.83 (5304.49) <0.001
VS 0.00 0.01 0.06 (5301.72) 0.950
vmPFC 0.03 0.01 3.00 (5305.26) 0.003
Random effects
variance SD
Participant 0.06 0.24
Open in a new tab

Note. DP1 is the model representing the core dual-process theoretical model; DP2 adds a term for vmPFC to DP1; VB1 is the model representing the core value-based choice theoretical model. The reference group for Food Type is healthy. Bolded values indicate statistical significance at P < 0.01. Degrees of freedom (df) were calculated using the Satterthwaite approximation.

In contrast, results generally supported the hypotheses of the value-based choice model. Fit improved significantly from the base model when individual terms for dlPFC, vmPFC and VS activity were added, according to the chi-square difference test, χ2(3) = 53.73, P < 0.001. Furthermore, directly comparing the canonical dual-process and value-based choice models (DP1 and VB1, respectively) revealed the value-based choice model as the better fitting, AICDP1 = 9746.80, R2DP1 = 0.19, AICVB1 = 9705.49, R2VB1 = 0.20. Critically, VB1 was also the best fitting model compared to other, non-preregistered potential specifications of dual-process and value-based choice models (see specification curve in Supplementary material). Inspection of the fixed effects of the canonical value-based choice model (VB1) revealed that bid value was positively associated with dlPFC activity (b = 0.04, 95% CI = [0.02, 0.06], P < 0.001) and vmPFC activity (b = 0.03, 95% CI = [0.01, 0.05], P = 0.003). Each one standard deviation increase in dlPFC activity was associated with a 4.4 cent increase in bid value, whereas it was associated with a 3.1 cent increase for vmPFC. VS activity was not significantly associated with bid value, b = 0.00, 95% CI = [−0.02, 0.02], P = 0.950. See Table 3 for VB1 parameter estimates and statistics.

We also observed qualified support for vmPFC integrating responses from dlPFC and VS, as hypothesized by the value-based choice model (Table 4). The results of this multilevel model showed that both dlPFC (b = 0.43, 95% CI = [0.40, 0.46], P < 0.001) and VS (b = 0.12, 95% CI = [0.09, 0.14], P < 0.001) were positively associated with vmPFC activity, but their interaction was not (b = 0.01, 95% CI = [0.00, 0.02], P = 0.123).

Table 4.

Results of the multilevel model regressing vmPFC activity on dlPFC and VS.

Fixed effects b SE t (df) P
Intercept −0.32 0.04 7.69 (101.75) <0.001
dlPFC 0.43 0.01 29.75 (5306.36) <0.001
VS 0.12 0.01 8.08 (5306.96) <0.001
dlPFCInline graphicVS 0.01 0.01 1.54 (5262.86) 0.123
Random effects variance SD
Participant 0.14 0.37
Open in a new tab

Note. This model represents an ancillary hypothesis of value-based choice models that the interaction between dlPFC and VS is associated with vmPFC activity. Bolded values indicate statistical significance at P < 0.001. Degrees of freedom (df) were calculated using the Satterthwaite approximation.

Discussion

We used a novel trial-by-trial statistical modeling approach to compare two prominent neurocognitive models of self-control. Analyses focused on the three ROIs (dlPFC, VS and vmPFC) commonly implicated in dual-process and value-based choice models of self-control. We preregistered and tested competing hypotheses about these regions posed by the theoretical models and then compared the models based on their fit to the data.

Our results did not support the core hypothesis posed by the dual-process model—that the relative activation between dlPFC and VS is what drives self-control-relevant outcomes (Lopez et al., 2017). If this were the case, DP1 should have been the best fitting model, but it was not. In addition, the consistent improvement in fit when including vmPFC is inconsistent with dual-process theory. Though chi-square statistics are sensitive to the number of free parameters, the lack of evidence for the dual-process model is not merely a function of reduced degrees of freedom. Because we preregistered our models, parameters were not included or excluded based on chance variation in the data. Also, VB1 had the lowest AIC, which penalizes additional parameters to reduce overfitting, despite having the most model parameters. Therefore, these results are inconsistent with the hypothesis that self-control-relevant decisions result from antagonism between dlPFC and VS.

In contrast, our results support the value-based choice hypothesis that vmPFC activity is associated with self-control-relevant decisions. Including vmPFC improved model fit, and the value-based choice model, VB1, was the best fitting model overall. Further, activations in the two regions that putatively represent the value of relevant choice attributes in this context—health in dlPFC and taste in VS—were both positively related to vmPFC activation. The observed positive correlation between VS and dlPFC is more consistent with the value-based choice model, where multiple attributes can contribute to the value of an option simultaneously.

Together, these results have implications for translational interventions to improve self-control. For example, they suggest that interventions seeking to amplify the subjective value of food health and/or decrease the value of food taste (e.g. via cognitive reappraisal) may be more effective than interventions targeting inhibitory control. Additionally, because the value-based choice model isn’t limited to ‘hot’ and ‘cold’ choice attributes, this model suggests that other sources of value, such as social norms or identity (Nook and Zaki, 2015; Berkman et al., 2017b; Pfeifer and Berkman, 2018), may be useful intervention targets.

However, our results did not support one secondary hypothesis of the value-based choice model. The model posed by Berkman et al. (2017a) specifies that vmPFC integrates value signals from dlPFC and VS, but it is unclear exactly how this ‘integration’ happens. If vmPFC merely serves to sum the weighted inputs, then an additive, ‘main effects only’ model might be possible. If vmPFC performs a more complex calculation (e.g. input-contingent) then an interaction model might also be possible. Thus, we tested both possibilities. Contrary to our hypothesis, while VS and dlPFC were positively associated with vmPFC activity, their interaction was not. Though there is no consensus across the various formulations of value-based choice models as to whether or not there are interactions among the inputs to the value accumulation (Berkman et al., 2017a; Hare et al., 2011; Lim et al., 2018; Sullivan et al., 2015), these data indicate that an interaction might not be present, at least in this task.

In light of the vmPFC model results, it is notable that this study design did not permit us to test directional relationships among ROIs. Though it would be possible to assess directionality using structural equation modeling, this method requires large samples (Kline, 2016) and we were underpowered to utilize it. Examining the structural relationship among these ROIs is an important avenue for future research. Previous tests of the directional relationship between vmPFC and dlPFC using a similar food task indicated that dlPFC moderated activity in vmPFC, which in turn influenced choice (Hare et al., 2011). The influence of dlPFC on vmPFC value signals has also been observed in other contexts (Hare et al., 2009; Hare et al., 2014).

The pattern of activity in VS is noteworthy in two ways. This region has been implicated in reward motivation (Schultz et al., 1992; Kelley, 2004), providing the basis of the dual-process prediction that VS activation would be more closely linked with bids for hedonically rewarding, unhealthy foods compared to healthy ones. In contrast, activity in VS was positively correlated with bid value regardless of stimulus type (Figure 5). It is possible that VS can represent non-hedonic types of reward (such as health) when stimuli come to be associated with those rewards for some other reason (such as a dieting goal). At the same time, we observed a drop in the magnitude of the positive relation between bid value and VS activity when dlPFC activity was entered into the model. This reduction may be due to the collinearity between VS and dlPFC, which in turn might be attributable to the participants’ dieting goals. Cases where participants choose between one option that has (mostly) hedonic value and another that has both abstract goal value and (at least some) hedonic value are understudied in the research literature but might more realistically reflect how self-control-relevant decisions operate. Compared to dual-process models, value-based choice models can more flexibly account for these cases because the value-integration process is agnostic about the number and sign of value inputs to a choice. Only in value-based choice models can VS, presumably representing hedonic or immediate reward value of some kind, contribute positively toward both options in a choice.

This study has several limitations. First, we did not collect independent liking, health or taste ratings. These ratings would be necessary to make claims about the engagement of self-control or the relative contributions of taste and health on any single trial. Instead, our approach was to induce self-control goal dilemmas by sampling dieters with healthy eating goals and looking at average effects across participants to test neurocognitive models of self-control. Second, we used a bid increment of $0.50, which may limit power in studies with smaller sample sizes and/or fewer experimental trials (Simms et al., 2019). Finally, region-to-region differences in signal-to-noise ratio (SNR) can make it difficult to compare the relative contributions of various brain regions to a statistical model. Indeed, there were region-wise differences in SNR (Supplementary material), but vmPFC had the lowest SNR of the three focal regions, suggesting that SNR alone cannot account for the observed effects.

The primary contribution of this research is the first direct comparison of two neurocognitive models of self-control, but several other features of the study are also noteworthy. The ecological validity of the task was high because our participants were overweight dieters who were bidding on food they would actually receive. This feature of the study is in contrast to many studies in the self-control literature that use a convenience sample without necessarily verifying that they have goals (e.g. dieting) that would confer subjective value to the healthiness of a food (Milyavskaya et al., 2018). The translational value of the results stems in part from the fact that the most common target for weight-reducing interventions is precisely the population from which the sample is drawn—overweight people who want to diet.

This study also highlights the usefulness of studying self-control within the context of an actual goal, dieting, that has strong translational potential. Dieting is a promising model for self-control because it unfolds across a longer time span than a typical laboratory study—extending weeks or months as opposed to an hour—and yet consists of a series of individual decisions (i.e. food choices) that can be investigated with a brief laboratory session. The context of dieting can also be fruitful for informing models of self-control because it can imbue different types of stimuli with value by increasing the importance of different attributes; for example, a slice of cake holds both (positive) hedonic value and, simultaneously (negative) value with respect to the dieting goal. Though our study design precluded a test of whether participants’ valuation of snack foods changed across contexts, our data are consistent with the possibility that the healthy foods might have accumulated some hedonic value in addition to their health goal value. In this way, dieting provides a more nuanced, complex test of self-control theories by shifting and broadening the set of food attributes that are relevant to participants’ multiple and (sometimes) competing goals.

Conclusions

We found that the value-based choice model of self-control better described the observed data. Our results neither prove nor disprove the theories of self-control, but instead provide evidence in support of the predictions of this and similar value-based choice models that activity in vmPFC is related to decisions requiring self-control and is positively associated with dlPFC and VS. Contrary to predictions, we found evidence suggesting that dlPFC and VS are not interactively associated with vmPFC, and our study design could not clarify the directionality of the relationships among these regions. Nevertheless, our unique approach to modeling neuroimaging data on a trial-by-trial basis and comparing theoretical models has helped refine our understanding of the neurobiological mechanisms underlying self-control and may, in turn, help inform the development of translational interventions to aid those who struggle with self-control.

Author contributions

All authors developed the study concept and analytic plan. D.C. and R.M.L. performed the data analysis and interpretation under the supervision of E.T.B. D.C. and R.M.L. drafted the manuscript, and E.T.B. provided critical revisions. All authors approved the final version of the manuscript for submission.

Funding

This work was supported by the NIH (CA211224 and CA17524 to E.T.B. and CA232357 to D.C.).

Supplementary Material

scan-19-009-File012_nsz068
Click here for additional data file. (1.8MB, docx)

Acknowledgements

We thank Nicole Giuliani for helping design the original study, Junaid Merchant and Bryce Dirks for data collection and Dagmar Zeithamova for methodological consultation. E.T.B. is manager of Berkman Consultants, LLC.

References

  1. Aho K., Derryberry D., Peterson T. (2014). Model selection for ecologists: the worldviews of AIC and BIC. Ecology, 95(3), 631–6. [DOI] [PubMed] [Google Scholar]
  2. Bakdash J.Z., Marusich L.R. (2017). Repeated measures correlation. Frontiers in Psychology, 8, 456. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Bartra O., McGuire J.T., Kable J.W. (2013). The valuation system: a coordinate-based meta-analysis of BOLD fMRI experiments examining neural correlates of subjective value. NeuroImage, 76, 412–27. [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Bates D., Mächler M., Bolker B., Walker S. (2015). Fitting linear mixed-effects models using lme4. Journal of Statistical Software, 67(1), 1–48. [Google Scholar]
  5. Berkman E.T. (2017). The neuroscience of self-control In: Ridder D., Adriaanse M., Fujita K., editors. Handbook of Self-Control in Health and Wellbeing, Routledge: Abingdon-on-Thames, pp. 112–24. [Google Scholar]
  6. Berkman E.T., Hutcherson C.A., Livingston J.L., Kahn L.E., Inzlicht M. (2017a). Self-control as value-based choice. Current Directions in Psychological Science, 26(5), 422–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Berkman E.T., Livingston J.L., Kahn L.E. (2017b). Finding the “self” in self-regulation: the identity-value model. Psychological Inquiry, 28(2–3), 77–98. [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Braver T.S., Paxton J.L., Locke H.S., Barch D.M. (2009). Flexible neural mechanisms of cognitive control within human prefrontal cortex. Proceedings of the National Academy of Sciences, 106(18), 7351–6. [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Buhle J.T., Silvers J.A., Wager T.D., et al. (2014). Cognitive reappraisal of emotion: a meta-analysis of human neuroimaging studies. Cerebral Cortex, 24(11), 2981–90. [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Clithero J.A., Rangel A. (2013). Informatic parcellation of the network involved in the computation of subjective value. Social Cognitive and Affective Neuroscience, 9(9), 1289–302. [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Cosme D., Flournoy J.C., DeStasio K.L. (2018). Auto-motion [code repository] . Retrieved from doi. doi: 10.5281/zenodo.1240528. [DOI]
  12. Cox R.W. (1996). AFNI: software for analysis and visualization of functional magnetic resonance neuroimages. Computers and Biomedical Research, 29(3), 162–73. [DOI] [PubMed] [Google Scholar]
  13. Desikan R.S., Ségonne F., Fischl B., et al. (2006). An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest. NeuroImage, 31(3), 968–80. [DOI] [PubMed] [Google Scholar]
  14. Destrieux C., Fischl B., Dale A., Halgren E. (2010). Automatic parcellation of human cortical gyri and sulci using standard anatomical nomenclature. NeuroImage, 53(1), 1–15. [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Duncan J. (2013). The structure of cognition: attentional episodes in mind and brain. Neuron, 80(1), 35–50. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Figner B., Knoch D., Johnson E.J., Krosch A.R., Lisanby S.H., Fehr E., Weber E.U. (2010). Lateral prefrontal cortex and self-control in intertemporal choice. Nature Neuroscience, 13(5), 538. [DOI] [PubMed] [Google Scholar]
  17. Fischl B. (2012). FreeSurfer. NeuroImage, 62(2), 774–81. [DOI] [PMC free article] [PubMed] [Google Scholar]
  18. Fischl B., Salat D.H., Busa E., et al. (2002). Whole brain segmentation: automated labeling of neuroanatomical structures in the human brain. Neuron, 33(3), 341–55. [DOI] [PubMed] [Google Scholar]
  19. Fujita K. (2011). On conceptualizing self-control as more than the effortful inhibition of impulses. Personality and Social Psychology Review, 15(4), 352–66. [DOI] [PubMed] [Google Scholar]
  20. Fujita K., Carnevale J.J., Trope Y. (2018). Understanding self-control as a whole vs. part dynamic. Neuroethics, 11(3), 283–96. [Google Scholar]
  21. Hare T.A., Camerer C.F., Rangel A. (2009). Self-control in decision-making involves modulation of the vmPFC valuation system. Science, 324(5927), 646–8. [DOI] [PubMed] [Google Scholar]
  22. Hare T.A., Camerer C.F., Knoepfle D.T., O'Doherty J.P., Rangel A. (2010). Value computations in ventral medial prefrontal cortex during charitable decision making incorporate input from regions involved in social cognition. Journal of Neuroscience, 30(2), 583–90. [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Hare T.A., Malmaud J., Rangel A. (2011). Focusing attention on the health aspects of foods changes value signals in vmPFC and improves dietary choice. Journal of Neuroscience, 31(30), 11077–87. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Hare T.A., Hakimi S., Rangel A. (2014). Activity in dlPFC and its effective connectivity to vmPFC are associated with temporal discounting. Frontiers in Neuroscience, 8, 50. [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Harris A., Clithero J.A., Hutcherson C.A. (2018). Accounting for taste: a multi-attribute neurocomputational model explains the neural dynamics of choices for self and others. Journal of Neuroscience, 38(37), 7952–68. [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Heatherton T.F., Wagner D.D. (2011). Cognitive neuroscience of self-regulation failure. Trends in Cognitive Sciences, 15(3), 132–9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Hutcherson C.A., Plassmann H., Gross J.J., Rangel A. (2012). Cognitive regulation during decision making shifts behavioral control between ventromedial and dorsolateral prefrontal value systems. Journal of Neuroscience, 32(39), 13543–54. [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Jenkinson M., Beckmann C.F., Behrens T.E., Woolrich M.W., Smith S.M. (2012). FSL. Neuroimage, 62(2), 782–90. [DOI] [PubMed] [Google Scholar]
  29. Kelley A.E. (2004). Ventral striatal control of appetitive motivation: role in ingestive behavior and reward-related learning. Neuroscience & Biobehavioral Reviews, 27(8), 765–76. [DOI] [PubMed] [Google Scholar]
  30. Kline R.B. (2016). Methodology in the Social Sciences. Principles and Practice of Structural Equation Modeling, 4th edn, New York, NY: Guilford Press. [Google Scholar]
  31. Kotabe H.P., Hofmann W. (2015). On integrating the components of self-control. Perspectives on Psychological Science, 10(5), 618–38. [DOI] [PubMed] [Google Scholar]
  32. Levy D.J., Glimcher P.W. (2012). The root of all value: a neural common currency for choice. Current Opinion in Neurobiology, 22(6), 1027–38. [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Lieberman M.D., Straccia M.A., Meyer M.L., Du M., Tan K.M. (2019). Social, self,(situational), and affective processes in medial prefrontal cortex (MPFC): causal, multivariate, and reverse inference evidence. Neuroscience & Biobehavioral Reviews.. [DOI] [PubMed] [Google Scholar]
  34. Lim S.L., O'Doherty J.P., Rangel A. (2011). The decision value computations in the vmPFC and striatum use a relative value code that is guided by visual attention. Journal of Neuroscience, 31(37), 13214–23. [DOI] [PMC free article] [PubMed] [Google Scholar]
  35. Lim S.L., Penrod M.T., Ha O.R., Bruce J.M., Bruce A.S. (2018). Calorie labeling promotes dietary self-control by shifting the temporal dynamics of health-and taste-attribute integration in overweight individuals. Psychological Science, 29(3), 447–62. [DOI] [PubMed] [Google Scholar]
  36. Lopez R.B., Hofmann W., Wagner D.D., Kelley W.M., Heatherton T.F. (2014). Neural predictors of giving in to temptation in daily life. Psychological Science, 25(7), 1337–44. [DOI] [PMC free article] [PubMed] [Google Scholar]
  37. Lopez R.B., Chen P.H.A., Huckins J.F., Hofmann W., Kelley W.M., Heatherton T.F. (2017). A balance of activity in brain control and reward systems predicts self-regulatory outcomes. Social Cognitive and Affective Neuroscience, 12(5), 832–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Lorah J. (2018). Effect size measures for multilevel models: definition, interpretation, and TIMSS example. Large-Scale Assessments in Education, 6(1), 8. [Google Scholar]
  39. MacDonald A.W., Cohen J.D., Stenger V.A., Carter C.S. (2000). Dissociating the role of the dorsolateral prefrontal and anterior cingulate cortex in cognitive control. Science, 288(5472), 1835–8. [DOI] [PubMed] [Google Scholar]
  40. McClure S.M., Ericson K.M., Laibson D.I., Loewenstein G., Cohen J.D. (2007). Time discounting for primary rewards. Journal of Neuroscience, 27(21), 5796–804. [DOI] [PMC free article] [PubMed] [Google Scholar]
  41. Miller E.K., Cohen J.D. (2001). An integrative theory of prefrontal cortex function. Annual Review of Neuroscience, 24(1), 167–202. [DOI] [PubMed] [Google Scholar]
  42. Milyavskaya M., Berkman E.T., De Ridder D.T. (2018). The many faces of self-control: tacit assumptions and recommendations to deal with them. Motivation Science, 5(1), 79. [Google Scholar]
  43. Nook E.C., Zaki J. (2015). Social norms shift behavioral and neural responses to foods. Journal of Cognitive Neuroscience, 27(7), 1412–26. [DOI] [PubMed] [Google Scholar]
  44. Padoa-Schioppa C., Conen K.E. (2017). Orbitofrontal cortex: a neural circuit for economic decisions. Neuron, 96(4), 736–54. [DOI] [PMC free article] [PubMed] [Google Scholar]
  45. Pfeifer J.H., Berkman E.T. (2018). The development of self and identity in adolescence: neural evidence and implications for a value-based choice perspective on motivated behavior. Child Development Perspectives, 12(3), 158–64. [DOI] [PMC free article] [PubMed] [Google Scholar]
  46. Plassmann H., O'Doherty J.P., Rangel A. (2010). Appetitive and aversive goal values are encoded in the medial orbitofrontal cortex at the time of decision making. Journal of Neuroscience, 30(32), 10799–808. [DOI] [PMC free article] [PubMed] [Google Scholar]
  47. Rangel A., Hare T. (2010). Neural computations associated with goal-directed choice. Current Opinion in Neurobiology, 20(2), 262–70. [DOI] [PubMed] [Google Scholar]
  48. Rissman J., Gazzaley A., D'Esposito M. (2004). Measuring functional connectivity during distinct stages of a cognitive task. NeuroImage, 23(2), 752–63. [DOI] [PubMed] [Google Scholar]
  49. Roy M., Shohamy D., Wager T.D. (2012). Ventromedial prefrontal-subcortical systems and the generation of affective meaning. Trends in Cognitive Sciences, 16(3), 147–56. [DOI] [PMC free article] [PubMed] [Google Scholar]
  50. Schultz W., Apicella P., Scarnati E., Ljungberg T. (1992). Neuronal activity in monkey ventral striatum related to the expectation of reward. Journal of Neuroscience, 12(12), 4595–610. [DOI] [PMC free article] [PubMed] [Google Scholar]
  51. Shenhav A. (2017). The perils of losing control: why self-control is not just another value-based decision. Psychological Inquiry, 28(2–3), 148–52. [DOI] [PMC free article] [PubMed] [Google Scholar]
  52. Shenhav A., Botvinick M.M., Cohen J.D. (2013). The expected value of control: an integrative theory of anterior cingulate cortex function. Neuron, 79(2), 217–40. [DOI] [PMC free article] [PubMed] [Google Scholar]
  53. Simms L.J., Zelazny K., Williams T.F., Bernstein L. (2019). Does the number of response options matter? Psychometric perspectives using personality questionnaire data. Psychological Assessment, 31(4), 557. [DOI] [PubMed] [Google Scholar]
  54. Simonsohn U., Simmons J.P., Nelson L.D. (2015). Specification curve: descriptive and inferential statistics on all reasonable specifications. SSRN Electronic Journal. doi: 10.2139/ssrn.2694998. [DOI] [Google Scholar]
  55. Sullivan N., Hutcherson C., Harris A., Rangel A. (2015). Dietary self-control is related to the speed with which attributes of healthfulness and tastiness are processed. Psychological Science, 26(2), 122–34. [DOI] [PMC free article] [PubMed] [Google Scholar]
  56. Tusche A., Hutcherson C.A. (2018). Cognitive regulation alters social and dietary choice by changing attribute representations in domain-general and domain-specific brain circuits. eLife, 7, e31185. [DOI] [PMC free article] [PubMed] [Google Scholar]
  57. Yarkoni T., Poldrack R.A., Nichols T.E., Van Essen D.C., Wager T.D. (2011). Large-scale automated synthesis of human functional neuroimaging data. Nature Methods, 8(8), 665. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

scan-19-009-File012_nsz068
Click here for additional data file. (1.8MB, docx)

Articles from Social Cognitive and Affective Neuroscience are provided here courtesy of Oxford University Press

RESOURCES