Evidence accumulation, not “self-control,” explains why the dlPFC activates during normative choice

What role do cognitive control regions like the dorsolateral prefrontal cortex (dlPFC) play in normative behavior (e.g., generosity, healthy eating)? Some models suggest that dlPFC activation during normative choice reflects the use of control to overcome default hedonistic preferences. Here, we develop an alternative account, showing that an attribute-based neural drift diffusion model (anDDM) predicts trial-by-trial variation in dlPFC response across three fMRI studies and two self-control contexts (altruistic sacrifice and healthy eating). Using the anDDM to simulate a variety of self-control dilemmas generated a novel prediction: although dlPFC activity might typically increase for norm-consistent choices, deliberate self-regulation focused on normative goals should decrease or even reverse this pattern (i.e., greater dlPFC response for hedonic, self-interested choices). We confirmed these predictions in both altruistic and dietary choice contexts. Our results suggest that dlPFC’s response during normative choice may depend more on value-based evidence accumulation than inhibition of our baser instincts.


Introduction.
Self-control dilemmas typically involve trade offs between short-term, hedonic considerations and longer-term or more abstract standards and values. For example, social interactions often force an individual to weigh self-interest against norms favoring equity and other-regard.
Similarly, dietary decisions often require weighing the immediate pleasure of consumption against personal standards or societal norms favoring healthy eating. Understanding when, why, and how people choose normatively-preferred responses (e.g., generosity over selfishness, healthy over unhealthy eating, etc.) has represented a central goal of the decision sciences for decades. What neural and computational processes must be engaged to support more normative behavior? What makes such choices frequently feel so conflicted and effortful, and how can we make them easier? To what extent does following social or personal norms depend on activation in brain regions associated with cognitive control, such as the dorsolateral prefrontal cortex (dlPFC)?
Previous research has provided a wealth of evidence suggesting that the dlPFC may promote more normative choices in both the social and non-social domain. For instance, compared to unhealthy food choices, healthier choices in successful dieters were accompanied by greater activation in a posterior region of the dlPFC 1 . Greater dlPFC response in a similar region has also been observed when individuals make normatively-favored choices in both social decision making 2,3 and intertemporal choice 4,5 . Moreover, activation in the dlPFC increases when individuals explicitly focus on eating healthy 6 or on decreasing craving for food 7 . Electrical disruption of this area also decreases patience 8 and reduces normative behavior in social contexts like the Ultimatum game 9 . Collectively, these results support the notion that the dlPFC may be recruited to modulate values or bias choices in favor of normative responses, perhaps especially when those responses conflict with default preferences.
Yet a variety of results seem inconsistent with this view. For example, researchers often fail to observe increased dlPFC recruitment when individuals make pro-social or intertemporally normative choices [10][11][12] . Moreover, electrical disruption of the dlPFC has been observed both to decrease appetitive valuation of foods 13 , and increase generous behavior in the Dictator Game 9 .
Such findings conflict with the idea that this region consistently promotes normative concerns over immediate, hedonistic desires. Thus, how to predict whether and when one might observe a positive association between dlPFC response and choices typically associated with successful self-control remains unclear.
Here, we propose a computational account of fMRI BOLD response in the dlPFC that may resolve many of these apparent inconsistencies. This account draws on prior research in both perceptual and value-based decision making, which consistently finds that the posterior dlPFC region associated with normative "self-control success" also activates during choices that are more difficult to discriminate in simple perceptual and value-based choices lacking a self-control conflict, e.g., [14][15][16] . Our account is also inspired by findings that the dlPFC may be one hub in a larger neural circuit (encompassing additional regions like the dorsal anterior cingulate cortex that selects actions for execution using a process of evidence accumulation and lateral inhibition among competing action representations 17,18 . Based on this evidence, we developed a computational model of self-control dilemmas that successfully predicts not only when an individual will choose in normative rather than hedonistic fashion, but also when, why, and to what degree response in the dlPFC will be recruited during that process. We note also that, although we focus here on the dlPFC, our model also applies in theory when observing similar relationships to other brain areas frequently associated with conflict and cognitive control, including regions of the IFG/aIns and dACC. As with similar models of simple perceptual and value-based choices, our attribute-based neural drift diffusion model (anDDM) assumes that the brain makes decisions through a process of value-based attribute integration and competition ( Figure 1). More specifically, choices are resolved via competitive interactions between neuronal populations that output responses based on accumulated information about the value of choice attributes, weighted by their momentary goal relevance. Some of these attributes are associated with hedonism (e.g., self-regarding concerns in altruistic choice) and some are associated with social norms and standards for behavior (e.g. other-regarding concerns). For expository purposes, we refer to these respectively as hedonic and normative attributes. Intuitively, whether our computational algorithm makes a hedonistic or normative choice depends not only on the magnitude of hedonic and normative attributes, but also on their weight: higher weights on normative attributes lead to more normconsistent responses.
What role does the dlPFC play in the anDDM? The observation of increased posterior dlPFC response when people choose consistently with normatively favored goals (e.g., healthy over unhealthy choices) has been taken to suggest that this region acts either to modulate the processing of attribute values or their weights in favor of normatively-favored goals 1,6 , or to inhibit hedonistic reward-related responding 19,20 . In contrast, we propose that activity in this region reflects processes related to the response selection stage of decisions. This suggests that dlPFC response during normative choice represents a downstream consequence of valuation processes, rather than a direct causal influence on them. To support this argument, we use the anDDM to simulate when and why we might observe greater activity in the dlPFC (and regions with similar response profiles) when resolving a choice. As we describe below, these simulations suggest that normative choices should be associated with greater neural activation in the dlPFC only when two things are true: hedonic attribute values directly oppose normative attribute values, and hedonic attributes receive more weight as inputs to the anDDM. In contrast, when normative attributes receive more weight, hedonistic choices should produce greater activity in the dlPFC and other areas associated with response selection.
We then used these observations to make two predictions. First, if people by default favor hedonic over normative attributes, then most studies will observe greater dlPFC response when people choose the normatively-favored option. This prediction does not strongly distinguish our account from alternatives. However, our model makes a second, more novel prediction: if a normally hedonistic decision maker focuses on normative goals, this should reduce activation in the dlPFC when choosing the normatively-favored option. A straightforward reading of an attribute-weighting account predicts the opposite: a normally hedonistic individual who deliberately attempts to focus on normative responding should show increased activation in the dlPFC in order to alter attribute weighting in favor of normative goals 19,21 . We test these two alternative predictions across three studies and two canonical self-control contexts in which people frequently struggle to align their actual behaviors with normative goals: altruistic and dietary choice. In all cases, results strongly supported the predictions of the anDDM. These findings raise new and important questions regarding the role of the dlPFC-and effortful selfcontrol more generally -in promoting normative choice.

Simulating the dilemma of self-control
Although self-control dilemmas can take a variety of forms, for expository purposes we here take a single, typical self-control dilemma: a decision maker deciding whether to indulge in a decadent snack or opt for something healthier. This example allows us to capture two critical features: first, self-control dilemmas typically involve making decisions about options that vary in the magnitude or value of hedonic and normative attributes (e.g. tastiness and healthiness).
Second, the decision-maker must weigh these attributes based on goals that can vary in their relative strength at different times. At a nice restaurant, tastiness may be prioritized. When trying to lose weight, healthiness is prioritized. We used simulations to explicitly capture these two features.
Simulations were realized using a neural network instantiation of our anDDM 18 where choices result from dynamic interactions between two separate but intermingled pools of neurons representing the different options under consideration ( Figure 1). Activation in each pool accumulates noisily based on a combination of external inputs from hedonic and normative attributes weighted by their current subjective importance, inhibitory inputs from the other pool, and recurrent self-stimulation (see Methods for details). This model generated predictions for how magnitudes and weights for hedonic and normative attributes influence the likelihood of a virtuous (i.e., healthy) choice, response time [RT], and neural response. These simulations yielded three key observations about behavior and neural response, which we describe in the context of food choice but apply in theory across any self-control dilemma that requires weighing hedonic rewards against normative values and goals.

Figure 1.
Attribute-based neural drift diffusion model (anDDM) of normative choice. Each option's hedonic and normative attributes (e.g., tastiness = +5 and healthiness = 0 for the sundae) are weighted by their current importance (e.g., w Taste [w T ] and w Health [w H ]) and summed to construct relative option values [V O1 -V O2 ]. These values, corrupted by momentary noise at time t [ε 1 (t)], serve as the external inputs to two mutually inhibitory neuronal pools representing the two options. Neural activation in these two pools (red and green lines in upper right plot) accumulates over time until one hits a predefined threshold, determining both the simulated response time (RT) and the simulated choice. Choices are classified as normative if the option with higher normative attribute value (in this case, higher healthiness, i.e. the apple in option 2) is selected. The sum of neural activation across the two pools can be used to simulate expected neural signals at the time of choice, and can be convolved with the canonical hemodynamic response function to construct a predicted BOLD signal for each choice (lower right inset).
Observation 1: The likelihood of a normative choice depends on the value of hedonic and normative attributes. To capture the idea that some choices (e.g. ice cream vs. Brussels sprouts) represent more of a self-control conflict than others (e.g. strawberries vs. lard), we simulated a single decision maker facing choices between hypothetical options that independently varied the relative value of normative and hedonic attributes (e.g. the foods' relative healthiness and tastiness). In the context of food choice, we classified a simulated choice as normative (healthy) when the simulation selected the option with higher healthiness. Choices were classified as hedonistic (unhealthy) otherwise. To determine the effect of current behavioral goals, we simulated the decision maker's choices for a variety of different weights on healthiness (w Health ) and tastiness (w Taste ). averaging over different instances of w Taste and w Health ). As can be seen, the magnitude and sign of the two attributes matters: she tends to choose more healthily when one option dominates on both healthiness and tastiness (no-conflict trials). She chooses less healthily when one option is tastier while the other is healthier (conflict trials). She is least likely to choose normatively when the difference in tastiness is large and the difference in healthiness is small. Thus, our simulations make the commonsense prediction that attribute values matter in determining the overall likelihood that an individual makes a healthy/normative choice.
Observation 2: The likelihood of a normative choice depends on weights given to normative and hedonic attributes. We next attempted to capture the idea that an individual might vary from context to context in the goals that they prioritize, and that the essence of self-control is to prioritize (i.e., assign a higher weight to) normative attributes like healthiness, or to deprioritize (i.e., assign a lower weight to) hedonic attributes like tastiness. We thus simulated the decision maker in different goal states by assuming different weights on hedonic and normative attributes (i.e. tastiness and healthiness). We show two example simulations in Figure 2b-d. Unsurprisingly, the decision maker chooses healthily less frequently when weight on tastiness is higher than weight on healthiness. However, these differences are starkest in conflict trials, and essentially vanish for no-conflict trials (Figure 1d). another, and is less likely when tastiness and healthiness conflict. Warmer colors indicate a higher likelihood of choosing the healthier option. Specific goals (b) prioritizing tastiness or (c) prioritizing healthiness alter the overall frequency of healthy choice, although in both contexts unhealthy choices are more likely for large differences in tastiness and small differences in healthiness. (d) The overall likelihood of a healthy choice (averaged for all combinations of conflict or no conflict choices). Goals prioritizing tastiness (black bars) produce fewer healthy choices than goals prioritizing healthiness (gray bars), but only when tastiness and healthiness conflict. Bottom: e-g) The computational model can also simulate expected neural activity (i.e. aggregate activity in the two neuronal pools, summed over decision time: 1 + !"#$ 2) when choosing healthy [H] or unhealthy [UH] options, as a function of relative option values and different goals. Warmer colors indicate more activity when a healthy choice was made (i.e., Activity H > Activity UH ). h) Overall difference in neural activity for H compared to UH choices for goals prioritizing tastiness (black bars) and healthiness (gray bars), divided as a function of attribute conflict. In no conflict trials, healthy choices elicit less activity regardless of goal (i.e. Activity H < Activity UH ). In conflict trials, however, healthy choices elicit more activity (i.e. Activity H > Activity UH ), but only when goals prioritize tastiness. Identical results are obtained when substituting RT for neural response (see Supplementary Figure 1). Comparing differences in simulated neural response for healthy and unhealthy choices yields two important conclusions. First, when options do not conflict on healthiness or tastiness (i.e. one option is better on both), healthy choices generally elicit less activity than unhealthy ones ( Figure   2e). Notably, for no-conflict trials this holds true irrespective of whether a decision maker is currently prioritizing tastiness or healthiness (Figure 2f-g). Second, and more importantly, when attributes conflict, network activity during healthy vs. unhealthy choices shows a striking dependence on an individual's goals (i.e. the relative balance of w Health and w Taste ). In conflict trials, hedonism-favoring goals (i.e., w Taste > w Health ) result in higher activity on average when choosing healthily (Figure 2h). This difference becomes exaggerated as the magnitudes of tastiness and healthiness increase (Figure 2f). In contrast, when goals prioritize normative attributes like healthiness (i.e., w Health > w Taste ), simulated neural responses are lower on average for healthy compared to unhealthy choices (Figure 2g,h). Thus, neural response is positively associated with normative choice (i.e., greater neural activity to choose normatively instead of hedonistically) only when the decision maker places a higher weight on hedonistic than normative attributes. The same is true of simulated RTs, which are often used as a proxy for both choice difficulty and the presence of control (Supplementary Figure 1). Thus, in the anDDM the observation that normative choices activate brain areas associated with cognitive control might simply indicate that hedonic attributes are currently weighted more highly.

Testing computational predictions using fMRI data
The anDDM accurately predicts dlPFC activity across a variety of contexts.
It is currently unknown whether activity in the dlPFC region frequently associated with selfcontrol might reflect activation patterns in the anDDM in the same manner as simple choice 18 .
We thus began by verifying that trial-by-trial simulated neural activity in the anDDM correlated with activity in this region for complex, multi-attribute choices typical of different real-world self-control dilemmas. Note that, while this correlation could occur because the dlPFC performs the precise computations carried out by the anDDM, such a correlation could also occur if the dlPFC performs separate computational functions that activate proportionally to anDDM activity.
In either case, we would expect trial-by-trial activity of the dlPFC to correlate with predictions of the anDDM.
Our analysis focused on three previously-collected fMRI datasets 22   instructions asked participants to focus on different pro-social motivations (ethical considerations, partner's feelings) as they made their choice. (c) In Study 3, we examined the generalizability of the model-based predictions in another choice domain. Here, we manipulated weights on food's healthiness (normative attribute) and tastiness (hedonic attribute) using a Food Choice Task. In all studies, participants had 4 seconds to decide, and gave their response on a 4-point scale from "Strong No" to "Strong Yes".
We predicted that dlPFC activity should correlate parametrically with simulated activity of the anDDM during self-control dilemmas. To test this notion we first fit computational parameters of the anDDM to each participant's behavior (see Supplemental Figure 2 for model fits). We then asked whether parametric variation in the measured BOLD signals within the dlPFC ROI correlated with simulated response across all three fMRI studies (see Methods for detail).
As predicted, a three-way conjunction analysis identified the left dlPFC at our a priori threshold The preceding analysis confirmed that activity in the left dlPFC covaries with predicted activity simulated in the anDDM in three independent fMRI studies. We next confirmed the central prediction of our simulations concerning the relationship between normative choices and activity in the dlPFC. In particular, models suggesting that the dlPFC promotes normative choices 1,6,20 imply that norm-consistent choices should be accompanied by greater activation in the dlPFC (as has been observed previously). Moreover, this should be especially true when people focus on normative goals 6,7 , since those goals support norm-sensitive behavior and might require the override of default hedonistic preferences 19,24 . The anDDM makes the opposite prediction. While neural activity in the model (and by extension the dlPFC) can be higher for normative compared to hedonistic choices, this should be true only when goals lead to stronger weighting of hedonic attributes and attribute values conflict (c.f. Figure 2h) (Figure 4d, paired-t 43 = 2.98, P = .005). A whole-brain analysis confirmed that this pattern was specific to the dlPFC, as well as the dACC and insula/IFG regions also associated with the anDDM, rather than a general property of neural activity (see Supplementary Table 3 for details).  Methods for details of model fitting). As expected, weights for self-interest (a hedonic attribute, Having confirmed that the regulatory focus manipulation altered weights on hedonic and normative attributes, we next asked if this manipulation affected BOLD response during generous vs. selfish choice in the dlPFC, consistent with predictions of the anDDM. In particular, given that all trials involved conflict between normative and hedonic attributes, we predicted that in the Natural condition, where participants generally placed higher weight on self-interest (a hedonic attribute), generous choices should elicit higher activation. In contrast, in the Partner condition, which elicited higher weight on normative attributes (i.e., other's outcomes and fairness), selfish choices should elicit the greatest activity in the dlPFC. The Ethics condition, which elicited similar weights across the attributes, should lie in between.
To test these predictions, we performed one-way repeated measures ANOVAs with condition  Table 3 for details).

Regulatory effects on healthy vs. unhealthy choice (Study 3).
In Study 3, we sought to replicate the finding that a regulatory focus on normative attributes reduces activation in the dlPFC, but in a new, non-social domain: healthy eating. During the Food Choice Task in Study 3, we manipulated attribute weights by instructing participants either to "Respond Naturally", "Focus on Health", or "Focus on Taste" while making their choice. Normative (i.e., healthy) choices were defined as selecting the food with higher subjectively perceived healthiness (see Methods for details). Note that the "Focus on Health" instruction aimed to increase weight on healthiness (w Health ), a normative attribute. Extending results of Study 2, the "Focus on Taste" condition was designed to enhance the weight on tastiness (w Taste ), the hedonic attribute, which should preserve or even enhance the difficulty of normative choices that we observed in natural choice settings in study 1 and 2. This allowed us to verify that our findings are specifically driven by changes in weights, not simply because we asked participants to perform a cognitive task.
To confirm that the regulatory manipulation influenced attribute weights, we performed one-way  Table 3 for details).

Regulatory effects in the absence of conflict (Study 3).
Our analyses so far focused on conflict trials, since simulations suggest that these trials show the biggest differences as a function of attribute weights (Figure 2). The design of Study 3, which included a subset of trials with no attribute conflict, also allowed us to test one further prediction of the anDDM. In Observation #3, we found that normative choices should only be associated with increased neural activity when hedonic and normative attributes conflict (Figure 2h). When attributes do not conflict, the anDDM predicts that normative choices should on average result in lower neural response. Moreover, the anDDM suggests smaller differences in response across goal contexts favoring hedonism or health norms. This suggests that, in contrast to conflicted choices, there should be less effect of regulatory focus on dlPFC response during no-conflict choices.
To test this prediction, we first performed a one-way repeated measures ANOVA with condition when normative and hedonic attributes did not conflict is generally low, and shows little to no effect of regulatory focus or the relative weight on tastiness and healthiness.

Discussion
When and why do normative choices (i.e., those choices that conform to abstract standards and social rules) recruit regions associated with cognitive control like the dorsolateral prefrontal cortex (dlPFC)? Simulated activity from an attribute-based neural drift diffusion model A large literature, generally consistent with models that assume normative behavior requires controlled processing, suggests that the dlPFC activates when prepotent responses conflict with desired normative outcomes 27,28 . The neural activity of the anDDM, which arises from mutually inhibitory pools of option neurons receiving weighted inputs from hedonic and virtuous attributes, is in some ways consistent with such an interpretation. However, it calls into question assumptions that prepotency equates to hedonism, or even to automaticity 29 more generally. Instead, our model suggests that the "prepotent response" may correspond, at least in the realm of value-based decision making, to choices consistent with the choice attribute that is currently receiving higher weight, regardless of the source of that weight. In other words, even when higher weights on normative attributes derive primarily from a deliberative, regulatory focus, as in our final two studies 23  This observation may help to explain why some researchers have found evidence consistent with greater response in the dlPFC promoting normative choice 1,3,6,9,30,31 , while others have not [10][11][12] .
Variations that influence the weight on normative attributes-whether across individuals, goal contexts, or paradigms-will tend to reduce statistical significance and increase heterogeneity in the link between neural activation in the dlPFC and normative choice. Fortunately, our model provides a way to predict both when and why dlPFC activity will be observed. For example, in the domain of intertemporal choice, our model predicts that making future outcomes more salient should amplify their weight in the choice process, promoting patience while decreasing dlPFC activation. This is exactly what is observed empirically 12  to recompute values based on contextual information 32 . Yet we find that the relationship between self-control "success" and "failure" in this region reverses when participants actively focus on health: dlPFC now responds more strongly to unhealthy choices. These results seem incongruent with the notion that this area down-regulates weight on norm-inconsistent considerations and upregulates norm-consistent ones, since we observed decreased responses in this area in the context of increased normative choice and increased weight on normative attributes (Figure 4, c.f. 23 ).
Instead, this region appeared to correlate with the evidence accumulation stage of decisions, rather than with the evidence construction stage, responding during decision conflict generally, regardless of whether that conflict derived from greater weighting of hedonic or normative attributes.
We emphasize, however, that our results and conclusions apply narrowly to the area of dlPFC identified. The anDDM-related dlPFC region in this study lies posterior and dorsal to another dlPFC area that we have observed, in these same datasets, to track hedonistic and normative attributes in a goal-consistent manner and to serve as a candidate for mediating regulationinduced changes 23 . Furthermore, gray matter volume in this more anterior dlPFC area, but not in the posterior dlPFC region identified here, correlates with regulatory success 33 . Thus, while some areas of the dlPFC may indeed play an important role in promoting self-regulation and virtue by altering attribute weights in decision value, we suspect that they are anatomically and computationally distinct from the region of the posterior dlPFC sometimes assumed to serve this role. Future work will be needed to better delineate subregions of the dlPFC, and to determine the unique role each one plays in promoting normative choices.
The close correspondence between predictions of the anDDM and activation patterns in the dlPFC makes it tempting to conclude that this region performs this computational function.
While this hypothesis is consistent with results from single-cell recordings 17,34 , we also acknowledge that the dlPFC has been associated with many computational functions and roles, not all of which are mutually incompatible. Thus, it is possible that the dlPFC region observed here performs some sort of process that is correlated with, but not identical to, the neuronal computations of the anDDM. Future work, including computational modifications or additions to the anDDM, as well as recordings from other modalities 34 , may help not only to elucidate the precise computational functions served by this area, but also the ways in which it promotes adaptive choice and virtuous behavior. Work extending these findings to other domains of virtuous choice, such as moral decision making or intertemporal choice, may also help to identify the commonalities and differences across different self-control dilemmas.

Computational Model Simulations
Our attribute-based neural drift diffusion model (anDDM: Figure 1) assumes that brain areas involved in decision making (particularly those that convert preferences into action) contain two spatially intermingled populations of neurons representing the options under consideration (here denoted as Option 1 and Option 2), with instantaneous firing rates (FR) at time t of FR 1 (t) and Firing rates in the two pools evolve noisily over time according to the following two equations: where the noise terms ! ( ) are normally distributed ~N(0,.1), ≥ 1 represents recurrent autostimulation from the pool onto itself, ≥ 0 represents inhibitory input from the other pool, and ! and ! represent external inputs proportional to the overall values of Options 1 and 2, determined by the weighted sum of their choice-relevant attribute values: Thus, each pool's activity receives an external input proportional to its value relative to the other option. In our simulations, we assumed two independent attributes: one related to hedonism (e.g., tastiness of a food) and one related to norms and standards (e.g. healthiness), although in principle any number and type of attribute could occur. Using these equations allowed us to simulate the dynamically evolving balance of excitation and inhibition across the two neuronal populations, and to derive distributions of both response times (RTs) and neural response. We label the final output (i.e., choice) of the system as "normative" if it results in selecting the option with the higher unweighted value for the normative attribute (e.g., the option with higher healthiness).
To simulate everyday self-control dilemmas using this framework, we simulated choices between two options representing different combinations of hedonistic and normative attributes, allowing the relative value difference between an option and its alternative on a given attribute to We also sought to capture in our simulations the notion that a decision maker can vary from moment to moment in their commitment to and desire for hedonistic vs. normative goals. For example, a dieter may begin to relax the importance they place on norm-consistent attributes like healthiness once they reach their target weight, resulting in more unhealthy choices. In the main text (and Figure 2), we focus on simulations for two different goal contexts: one with a higher weight on tastiness, a hedonic attribute (i.e., wT = .05, wH = .02) and one with a higher weight on healthiness, a normative attribute (wT = .02, wH = .05). For simplicity, we assumed that all choices Using these values and weights, we simulated choice frequencies, total neural activation (summed across the two neuronal pools), and RTs for each of the different hypothetical option pairs/attribute combinations, probing the effects of attribute weights, attribute magnitudes, and attribute conflict (i.e. match or mismatch between the signs of nornative and hedonic attribute).
Results of these simulations are displayed in Figure 2. Code is available at [link released after publication].

Experimental Studies
Details about portions of Studies 1, 2 and 3, as well as neuroimaging parameters, have been reported previously 22,23 . Here, we highlight in brief the most important details for the current work.
All participants received a show-up fee of $30 as well as an additional amount ranging from $0-$100, depending on the outcome of the task (see below

Manipulating Normative Goals (Studies 2 & 3)
Our computational model simulations suggested that the extent to which normative choices are associated with greater neural response depends to a large extent on the priority or weight given to normative vs. hedonic attributes. We thus capitalized on the design of Studies 2 and 3, which manipulated attention to different attributes (and corresponding weights), allowing us to test specific predictions of the anDDM. focus on the specific instruction given for all trials within a block. However, participants knew that they would receive the outcome of one of their choices, and were told that they should choose according to their preferences regardless of the instruction, thus encouraging participants to choose in a way that reflected their current decision value for the item.

Defining Normative Choice
Behavioral definition of generosity. All choices involved a tradeoff between maximizing outcomes for the self or for the other. We therefore label specific decisions as normative (i.e., generous) if the participant accepted a proposal when $Self < $Other, or rejected one when $Self > $Other. Choices were labeled as hedonistic (i.e., selfish) otherwise.
Behavioral definition of healthy choice. In the Dietary Choice Task, we separately examined trials requiring a tradeoff between taste and health (i.e. conflict trials where a food was rated either as healthy but not tasty, or as unhealthy but tasty) as well as trials with no tradeoff (i.e., no-conflict trials where a food was both tasty and healthy, or both unhealthy and not tasty). In both cases, we label specific decisions as normative (i.e., healthy) if the participant either accepted a healthy food, or rejected an unhealthy food. All other choices were labeled as hedonistic (i.e., unhealthy).

Computational Model Fitting
We used a Bayesian model-fitting approach to identify best-fitting model parameters  Supplementary Table S1) were used to simulate trial-by-trial activation across the two neuronal pools for use in the GLMs described below. Importantly, parameter values identified by this fitting procedure suggested that the model provided a good fit to behavior across all three studies (Supplementary Figure 2).

Neuroimaging Analyses
GLM 1a: Correlates of the anDDM (Study 1). We used GLM 1a to identify brain regions where activation varied parametrically according to the predictions of the anDDM in Study 1 (Altruistic Choice Task). To this end, we determined that the best BOLD approximation of the anDDM was a parametric modulator with a value consisting of the sum total of the simulated response across both pools of neurons, averaged over all simulations terminating in the observed choice on that trial within ±250ms of the observed RT, and modulating a boxcar function with onset at the beginning of the choice period and having a duration of the RT on that trial (see Supplemental Methods for further detail on selecting the best regressor). To simulate expected anDDM activation on each trial, we generated 5000 simulations using the best-fitting parameters for each participant and the estimated value of the proposal and default on each trial (i.e., w Self *$Self + w Other *$Other + w Fairness *|$Self -$Other|).
Then, for each subject we estimated a GLM with AR(1) and the following regressors of interest: We then computed subject-level contrasts of the anDDM parametric modulator (R3) against an implicit baseline. Finally, to test the hypothesis that anDDM responses might correlate with activation in the dlPFC, we subjected this contrast to a one-sample t-test against zero, thresholded at a voxel-wise P < .001, and a cluster-defining threshold of P < .05, small-volume corrected within a 10-mm spherical region of interest (ROI) centered on the peak coordinates of activity for the contrast of normative (healthy) vs. hedonistic (unhealthy) choice in a previous study of self-control in dieters 1 . Regions lying outside of this ROI were also identified at a voxellevel P < .001 uncorrected and a whole-brain cluster-corrected level of P < .05.
GLM 1b: Correlates of the anDDM (Study 2). GLM1b was similar to GLM1a, with the exception that we estimated regressors for each condition separately. R1, R4, and R7 were boxcar functions representing the choice period for the Natural, Ethics, and Partner conditions, respectively. R2, R5, and R9 modulated R1, R4 and R7 with the decision value on that trial. R3, R6, and R9 modulated R1, R4, and R7 using the estimated activation of the anDDM on that trial.
A single contrast representing neural correlates of the anDDM was constructed by combining R3, R6 and R9 at the subject-level and performing a one-sample t-test against zero, thresholded at a voxel-wise P < .001 and a small-volume cluster-corrected level of P < .05 within the dlPFC ROI described above.
GLM1c: Correlates of the anDDM (Study 3). GLM1c was similar to GLM1b, but applied to the Food Choice Task. R1, R4, and R7 were boxcar functions representing Natural, Taste, and Health focus conditions. R2, R5, and R8 were parametric modulators representing the decision value on that trial, and R3, R6, and R9 were modulators consisting of anDDM activity simulated using healthiness and tastiness ratings as attributes. Similar to Studies 1 and 2, correlates of the anDDM were identified in this study thresholded at a voxel-wise P < .001 and a small-volume cluster-corrected level of P < .05 within the dlPFC ROI described above.
ROI definition. Based on GLMs 1a, b and c, we identified a region of the left dlPFC consistently associated with the anDDM across all three studies through a three-way conjunction analysis using the imcalc function in SPM12, with each individual study map thresholded at P < .05, small-volume corrected, and a minimum overlap of > 5 contiguous voxels. Outside of this ROI, we also identified regions significant across all three studies at P < .05, whole-brain corrected.
This identified just three regions, located in the left dlPFC, left IFG, and dACC ( Figure 4 and Supplemental Figures S3 and S4). We then interrogated activation within these regions specifically for the contrast of normative vs. hedonistic choice, using GLMs 2a, b and c, as specified below. Second, we computed the subject-level contrast image [R3 -R1], which identified regions with differential response for generous compared to selfish choices. Seven subjects were excluded from this analysis for having fewer than 4 generous choices over the 180 trials. We computed the average value of this contrast within the three anDDM ROIs specified above. As a supplementary analysis, we also asked whether any voxels beyond these regions demonstrated a significant effect, using a whole-brain analysis thresholded at P < .001, uncorrected (see Supplementary Table S3). which identified regions with differential response for generous compared to selfish choices in each condition. We computed the average value of each of these contrasts within the three anDDM ROIs specified above. As a supplementary analysis, we also asked whether any voxels beyond these regions demonstrated a significant effect in any condition, using a whole-brain analysis thresholded at P < .001, uncorrected (see Supplementary Table S3).
GLM 2c: Healthy vs. Unhealthy decisions in the Food Choice Task (Study 3). GLM 2c was analogous to GLM 2b, but examined healthy vs. unhealthy choices in the Dietary Choice Task, separately for conflicted trials (i.e. healthy but not tasty foods and tasty but unhealthy foods) and for unconflicted trials (i.e. healthy and tasty foods or unhealthy and not tasty foods). It included the following regressors of interest: R1) A boxcar function for the choice period on conflicted trials when the subject made a healthy choice (i.e., accepted a healthy-but-not-tasty or rejected a tasty-but-unhealthy food) in Natural Focus trials. unhealthy choices, in each condition separately and separately for conflicted vs. unconflicted trials, were computed in a manner identical to GLM2b. We analyzed activation for these contrasts specifically within the three ROIs identified as anDDM regions. As a supplementary analysis, we also report results at the whole-brain level at P < .001, uncorrected, in Table S3.
Unreported details are as in GLM 2a.
Data Availability. Behavioral data and all analysis code are available on the Open Science Framework at [link released after acceptance for publication]. Neuroimaging data are available upon request to the authors.