|Home | About | Journals | Submit | Contact Us | Français|
VV, AR, and SAH designed the experiments. VV and AR led the programming of the experiments, data collection, and data analysis, in collaboration with SAH. AT assisted with general data analysis and led the meta-analysis. All authors contributed to writing and editing the final manuscript.
The dorsomedial prefrontal cortex (DMPFC) plays a central role in aspects of cognitive control and decision making. Here, we provide evidence for an anterior-to-posterior topography within the DMPFC using tasks that evoke three distinct forms of control demands – response, decision, and strategic – each of which could be mapped on to independent behavioral data. Specifically, we identify three spatially distinct regions within the DMPFC: a posterior region associated with control demands evoked by multiple incompatible responses, a middle region associated with control demands evoked by the relative desirability of decision options, and an anterior region that predicts control demands related to deviations from an individual's preferred decision-making strategy. These results provide new insight into the functional organization of DMPFC, and suggest how recent controversies about its role in complex decision making and response mapping can be reconciled.
The medial prefrontal cortex contributes to a variety of cognitive functions like conflict monitoring (Botvinick et al., 2001; Pochon et al., 2008), error detection (Carter et al., 1998), outcome evaluation (Bush et al., 2002; Gehring and Willoughby, 2002), reinforcement learning (Kennerley et al., 2006), decision making under uncertainty (Rushworth et al., 2005; Rushworth and Behrens, 2008), emotions (Etkin et al., 2006) and social interactions (Amodio and Frith, 2006; Behrens et al., 2008; Behrens et al., 2009). Meta-analyses indicate that these functions are supported by brain regions along the anterior cingulate sulcus and extending dorsally, hereafter collectively referred to as dorsomedial prefrontal cortex (DMPFC) (Ridderinkhof et al., 2004; Beckmann et al., 2009).
Several lines of evidence suggest that DMPFC may possess a topographic organization of function. Studies of the anterior cingulate cortex (ACC), typically defined to encompass medial prefrontal cortex more broadly and ventrally than DMPFC, have found distinctions based on emotional content (Bush et al., 2000) and social relevance (Behrens et al., 2008) of task stimuli. The DMPFC has also been postulated to shape activation in dorsolateral PFC (DLPFC; Kerns et al., 2004; Meriau et al., 2006), and its subregions differ in their anatomical connectivity accordingly (Beckmann et al., 2009). Given evidence that the DLPFC is organized along a posterior-to-anterior gradient of increasing abstraction (Koechlin et al., 2000; Christoff et al., 2003; Koechlin et al., 2003), we hypothesized that a similar organization would exist within DMPFC. Specifically, we predicted that posterior DMPFC, which shows greater connectivity to motor and pre-motor regions, would support response-related control while more anterior regions, which show greater connectivity to DLPFC, would support decision-related control. Moreover, based on evidence that DMPFC is also recruited for choices that run counter to behavioral strategies (Paulus et al., 2002; De Martino et al., 2006; Hampton and O'Doherty, 2007; Venkatraman et al., 2009), we also predicted that strategy-related control would evoke activation in the most anterior aspect of DMPFC.
We evoked these three distinct forms of control – response, decision, and strategy – within the same subjects and with distinct behavioral covariates. To evoke response-related control demands, we used a counting Stroop task (Fig. 1a), using the subject-by-subject response-incongruency effect as a measure of response-related control (Fig. 2a). For decision and strategic control demands, we used a decision-making task (Fig. 1b) where subjects had to decide between two stocks based on their attributes: the balanced stock had equal ratings on two attributes, while the extreme stock had a high rating on one attribute but a lower rating on the other. Subjects showed a strong bias for an attribute-balancing strategy, consistent with previous studies (Chernev, 2004, 2005). On each trial, the decision-related control demand was characterized by the difference in relative desirability of the two stocks (Fig. 2b). Strategic control demands were characterized according to the degree of bias toward the balancing strategy within each individual (Fig. 2c). Should these three measures evoke distinct foci of activation, as predicted, there would be strong evidence in support of a functional topography within DMPFC.
Twenty-three young adults (7 males, mean age = 21.9) participated in this study. All subjects acclimated to the scanning environment using a mock MRI scanner and participated in two short practice runs consisting of six trials each, one inside and one outside of the MRI scanner. Three subjects were excluded prior to data analysis – two due to excessive motion and one for poor behavioral performance – leaving a total of 20 subjects in the final analyses. We used an incentive-compatible design in which subjects received monetary compensation both for participation and for their performance in the experiment (see Supplementary Materials for details). All subjects gave written informed consent as part of protocols approved by the Institutional Review Boards of Duke University and Duke University Medical Center.
Response-related control demands were evoked using a counting Stroop task (Bush et al., 1998). We employed a block design for this task (Fig. 1A). In the Neutral condition, subjects were presented with multiple repetitions of animal words (e.g. dog, cat) and asked to count the number of times the word was presented. In the Incongruent condition, they were presented with multiple repetitions of number words (one, two, three or four) and asked to count the number of times the word was presented. However, the number of repetitions was always incongruent with the number word itself (for e.g., “two” presented four times). Each word was repeated between 1-4 times and subjects responded by pressing the corresponding button on a 4-button response box. We calculated, for each subject, the magnitude of the Stroop incongruency effect, measured as the differences in response times between the Incongruent and Neutral conditions, and used it as an index of response-control demands (Fig. 2A).
Decision-related and strategy-related control demands were evoked using an attribute-balancing task where all subjects were presented with a total of 90 sets of three real stocks, each rated on two attributes that provided real metrics of stock performance (Fig. 1B). To minimize the influence of subjects' prior knowledge or financial biases, the identity of the stocks and the nature of the attributes were concealed from the subjects until after the experiment. On each trial, the three stocks were labeled as “A”, “B”, and “C”, and the two attributes were labeled as “a1” and “a2” and shown as percentile rankings. The subjects' task was to predict the better-performing stock based only on these attributes.
All choice trials involved one stock that had balanced values for both attributes, one stock that had a good rating on attribute a1 but a poorer rating on attribute a2, and a third stock in which this pattern was reversed with slightly varying attribute ratings (see Supplementary Materials for more details). In each trial, subjects had to choose between a balanced stock and an extreme stock that was randomly selected from the other two alternatives for that trial. Choosing the balanced stock was consistent with the simple attribute-balancing heuristic identified in previous studies using similar paradigms (Chernev, 2004, 2005).
In addition, we randomly interspersed six catch trials, where the extreme alternative had higher ratings than the balanced option for both attributes and thus was the more desirable option. These trials thus served to ensure that subjects followed instructions and attended to the task. One subject who performed contrary to expectations in the catch trials was excluded from further analysis.
We manipulated the decision-related control demands across trials by varying the expected-value relationship (defined as the sum of two attributes for each stock) between the balanced and extreme choices (Fig. 2B). In 36 trials (Congruent condition), the balanced stock had a higher expected value compared to the alternative option. In another 36 trials (Incongruent condition), the balanced stock had a lower expected value compared to the alternative option. In the remaining 18 choice trials (Equal condition), the expected value for the balanced stock was equal to the expected value of the alternative option. Decision-related control demands were minimal (i.e., where decisions were easiest) in the congruent condition, followed by the incongruent condition, and were maximal in the equal condition (i.e., where both choices have similar expected value).
Finally, we defined strategy-related control demands (Fig. 2C) according to the degree to which a given choice (i.e., balanced or extreme, on each trial) deviates from the individual's empirical strategy preferences (i.e., the proportion of balanced choices across all trials). Subjects who were normally biased toward the balanced option would exhibit maximal strategy-related control when they choose the extreme option. However, subjects who were biased for the extreme option would exhibit maximal strategy-related control when they choose the balanced option. This analysis approach counterbalances for any main effects of either choice type, ensuring that our effects indeed reflect strategy-related control.
Subjects first participated in four runs of the attribute-balancing decision task, each containing 24 gambles and lasting approximately 6 minutes. Each trial began with the display of all three stocks for 4 or 6s. Subjects were instructed previously to examine the stocks as presented (Fig. 1B). Subsequently, two of the three stocks were highlighted in red (one involving balanced attributes and other extreme attributes). Subjects were instructed to choose between one of these two stocks and incentivized with a monetary bonus for picking the better performing stock (see Supplementary materials). They had 6s to make their choice, whereupon two arrows appeared specifying which button corresponded to which choice. The association of the buttons to choice was random, with equal probabilities of each option being associated with left or right buttons. The mapping of buttons for each trial was not revealed until the response phase, further reducing any possibilities of response-related incongruency effects during the decision phase. Subjects were instructed to arrive at their decision during the 6s interval and to press the button corresponding to their choice as soon as the arrows appeared. Response times were defined as the interval between the appearance of arrows and the button-press response. The decision and response phases were explicitly separated to prevent the contamination of decision effects with response-preparation effects. During the inter-trial interval of 4-8 seconds, a fixation cross was displayed on the screen. Notice that no feedback was provided at the end of each trial and hence there was no explicit learning during the decision phase of the task.
After four runs of the decision task, all subjects completed one run of the Counting Stroop task. To maximize power for the contrast between Neutral and Incongruent trials, our design alternated between blocks of 6 trials of each type (3s between trial onsets, 18s blocks, 17 blocks/run), without any fixation or non-task blocks. A final 6-minute run, whose data are not discussed in this manuscript, resolved a subset of the trials from the decision phase to actual stock labels and attributes. The winning stock was then identified based on data we collected on the real performance of these stocks (see Supplementary materials), and subjects received a monetary bonus based on the performance of their selected stock on two randomly chosen trials.
Before the fMRI data collection, subjects had the opportunity to practice the experimental task (without reward) in two six-trial blocks, one presented outside the MRI scanner and the other presented within the MRI scanner but prior to collection of the fMRI data. All stimuli were created using the Psychophysics Toolbox for MATLAB (Brainard, 1997; Pelli, 1997) and were presented to the subjects via MR-compatible LCD goggles. Subjects responded with the index fingers of each hand via a MR-compatible response box.
We acquired fMRI data on a 4T GE scanner using an inverse-spiral pulse sequence with parameters: TR = 2000ms; TE = 30ms; 34 axial slices parallel to the AC-PC plane, with voxel size of 3.75*3.75*3.8mm. High-resolution 3D full-brain SPGR anatomical images were acquired and used for normalizing and coregistering individual subjects' data.
Analysis was carried out using FEAT (FMRI Expert Analysis Tool) Version 5.63, part of FSL (FMRIB's Software Library, www.fmrib.ox.ac.uk/fsl) package (Smith et al., 2004). We removed artifacts in the fMRI timeseries using MELODIC and in-house scripts, with three a priori criteria for identifying noise-related components: power at high frequencies, rapid changes within short time windows, and correlation with estimated head motion. The following pre-statistics processing steps were applied: motion correction using MCFLIRT, slice-timing correction, removal of non-brain voxels using BET, spatial smoothing with a Gaussian kernel of FWHM 8mm, and high-pass temporal filtering. Registration to high resolution and standard images was carried out using FLIRT.
For the response-related control analysis, we used one regressor to model activation in the incongruent blocks of the Stroop task (i.e., treating the neutral blocks as baseline). We introduced the mean response-time difference between Incongruent and Neutral trials for each subject as a covariate in the across-subjects regression analysis to determine the focus of activation for response-related control. The decision-related control analysis used six total regressors to model activation in the attribute-balancing task: three task regressors for the three conditions (congruent, incongruent and equal), one regressor for the initial presentation of the stocks, one regressor for catch trials, and one regressor for the response period (scaled by response time). We defined active voxels using a conjunction analysis of [equal > incongruent] and [incongruent > congruent]. The model used for the strategy-related control analysis contained five total regressors: two regressors modeling the balanced and extreme choices of subjects in conditions that cause the greatest decision conflict (trials where the absolute magnitude of expected value difference between the two options was less than 8), one regressor modeling the responses in the remaining conditions, one regressor for the initial presentation of the stocks, and one regressor to model the subject responses. We introduced each subject's relative bias towards the balancing strategy (z-transformed proportion of balancing choices) as a covariate in the contrast between the balanced and extreme choice regressors.
Event-related regressors were generated by convolving impulses at the onsets of events of interest with a double-gamma hemodynamic response function. Within-subject analyses were carried out using a fixed-effects model across runs for each subject. Across-subjects analyses were carried out using FLAME Stage 1, which is FSL's implementation of mixed-effect analyses. All statistical images were thresholded using clusters determined by z > 2.3 and a whole-brain-corrected cluster-significance threshold of p < 0.05. The same analysis procedures and thresholding were used for analyses involving a between-subjects covariate. The MRIcron package was used for visualizing brain images (Rorden et al., 2007).
For the Stroop task, subjects were significantly more accurate in the Neutral compared to the Incongruent condition [t(19)=2.46, p<0.05]. Similarly, subjects also took significantly longer in the Incongruent condition compared to the Neutral condition [t(19)=3.88, p<0.001]. Taken together, these findings are consistent with increased demands for response-related control in the Incongruent condition.
For the attribute-balancing task, subjects showed a significant bias toward the balanced alternative, preferring it in 58% of the trials. Within the equal condition, they preferred the balanced alternative in 65% of the trials [t(19) = 3.9, p<0.005]. Despite this overall bias, subjects' choices were still sensitive to changes in expected value. Consistent with our predictions for decision-related control demands, we found that subjects' response times were slowest for the equal trials, intermediate for incongruent trials, and fastest for congruent trials (Supplementary Table S1). Finally, subjects also showed substantial individual variability in their choice preferences, providing a foundation for our analyses of strategic control (Supplementary Fig. S1).
We found increased activation in the dorsomedial PFC, dorsolateral PFC and parietal cortex for the Incongruent condition in the Stroop task (Supplementary Table S2, Supplementary Fig. S2). When we used the difference in response time between Incongruent and Neutral conditions as a covariate, we found that activation in left dorsolateral prefrontal cortex and a posterior region of the DMPFC (pDMPFC) increased linearly with increasing response conflict (Fig. 3). The pDMPFC activation cluster was localized to the posterior part of the cingulate sulcus (BA 32). Activation within this cluster also correlated with an independent trait measure of motor impulsiveness (Patton et al., 1995) across subjects (r=0.49, p<0.05, Supplementary Fig. S3).
We next sought to identify brain regions whose activation was driven by the decision phase (when subjects deliberated between the two stock options) in the attribute-balancing task. We performed a conjunction analysis to identify regions showing significant effects for each of the two individual decision-related contrasts: equal > incongruent and incongruent > congruent. We found increased activation in right dorsolateral prefrontal cortex, right inferior parietal lobule, and a region of DMPFC anterior to that observed for response-related control (Supplementary Fig. S4A, Supplementary Table S3). The DMPFC activation cluster for this contrast was also along the cingulate sulcus (BA32), anterior to the cluster for response-related control, and we refer to it as mDMPFC. Within this region, activation increased in a stepwise manner for our three levels of decision-related control (Fig. 4). Notably, these effects cannot be attributed to response conflict, given that they are time-locked to the decision phase that occurs before response mappings had been indicated to the subject.
Next, we sought to identify regions that indexed strategic control demands across subjects, using individuals' preference for the balancing strategy as a covariate in our between-subject analysis. (We report main-effect analyses unrelated to strategy-related control in the supplementary material; see Supplementary Fig. S4B). We found that individual differences in strategic preferences across subjects predicted the relative activation evoked by balanced and extreme choices in only one region: anterior dorsomedial prefrontal cortex (aDMPFC; Fig. 5), along the paracingulate region (BA 9). People who expressed a greater preference for the balanced option exhibited a greater increase in DMPFC activation for the extreme option compared to the balanced option, and vice versa. The difference in activation also correlated with an independent trait measure of need for cognition (Epstein et al., 1996), which measures the relative cognitive effort associated with decision making (r=0.60, p<0.005, Supplementary Fig. S5).
To verify that the clusters implicated in strategy- and decision-related control were functionally distinct, we conducted a post hoc ROI analysis within the aDMPFC and mDMPFC regions. We found that aDMPFC did not show a linear relationship with increasing decision conflict, nor did mDMPFC show a significant correlation with strategic variability across subjects (Supplementary Fig. S6). Thus, we conclude that these regions made distinct functional contributions to performance of our experimental tasks.
The results presented above provide strong evidence for a functional topography within the DMPFC: the most anterior cluster tracked strategy-related control, a middle cluster predicted decision-related control, and a more posterior cluster tracked response-related control (Fig. 6). To ascertain the consistency of these results with the prior literature, we conducted a meta-analysis of 53 studies that reported DMPFC activation in tasks involving decision making and/or cognitive control. We classified each study as involving either decision-related or response-related control, based on task properties, and included every reported coordinate of medial frontal cortex activation in an Activation Likelihood Estimation (ALE) analysis (see Supplementary Material). We exclude strategy-related control in this meta-analysis due to lack of prior work in this area (but see De Martino et al., 2006; Venkatraman et al., 2009).
Our meta-analysis revealed a strong spatial dissociation within DMPFC between activations related to decisions and to responses, with more anterior regions being involved in decision-making and more posterior regions with response-related cognitive control (Supplementary Fig. S7). The centroids of maximum likelihood overlapped with the observed activations for decision-related and response-related control demands in our study (Fig. 6, solid squares), and both are posterior to the focus for strategy-related control in aDMPFC.
We used tasks that evoked three distinct forms of control demands – response, decision, and strategy – and found strong evidence for a posterior to anterior topography within the DMPFC. Three distinct foci of activation were identified based on separate behavioral measures using data from the same individuals. These results both provide new insight into the functional organization of DMPFC and suggest a reconciliation of recent controversies about that region's role in complex decision-making and cognitive control.
Control demands play an important role in adaptive decision-making, particularly since decision preferences are strongly influenced by context (Simonson and Tversky, 1992; Tversky and Simonson, 1993). For example, people avoid decision options that seem extreme, whether compared to their alternatives or whether having attributes with highly disparate values (Chernev, 2004, 2005). Options that are balanced (e.g., their scores on various attributes are more equal) thus frequently serve as a desirable compromise relative to options containing attributes with high dispersion. We introduced a similar biasing context in the current study, which used an abstract market environment with anonymized options to control for subjects' differential prior exposure to stocks. This task allowed us to manipulate the degree of decision-related and strategy-related control independently, the former by varying the relative desirability of the choices and the latter by measuring individual variability in strategic bias.
Several aspects of our experimental design allow us to conclude that these distinct forms of control were represented in distinct regions within DMPFC. First, we explicitly separated the decision phase from the response phase so that activations related to decision control cannot be attributed to confounding effects from response selection or motor preparation (Pochon et al., 2008). Second, subjects received no immediate feedback about their decisions when choosing between different stocks; instead, feedback was provided in a separate, later run. This allowed us to rule out effects of outcome history when characterizing our subjects' decision preferences, while also precluding alternative explanations for the observed DMPFC activation like error detection, reinforcement learning, and signaling reward prediction errors (Kiehl et al., 2000; Kerns et al., 2004; Kennerley et al., 2006; Kim et al., 2006). We do not claim that DMPFC plays no role in these functions, but note that they did not contribute to the activations observed in this experiment. Third, we used an independent counting Stroop task to elucidate neural mechanisms of response-related control. Fourth, all analyses were conducted within the same set of subjects, as critical for making clear spatial comparisons. And, finally, our distinct forms of control demands were each associated with a unique and independent behavioral covariate that was well-controlled within each task, eliminating potential effects of task-specific materials.
Several studies have found increased activation in the DMPFC associated with complex decision making (Paulus et al., 2002; Walton et al., 2003; Rushworth et al., 2004; Zysset et al., 2006; Botvinick and Rosen, 2008; Pochon et al., 2008), though these findings could be confounded by activation related to response preparation (Pochon et al., 2008). A more recent study using a perceptual decision-making task sought to explore the role of DMPFC in decision conflict, defined as the difficulty arising from choosing between two equally likely choices, while accounting for response-related activation (Pochon et al., 2008). They demonstrated greater DMPFC activation for decisions involving alternatives that were equally attractive, even in the absence of an explicit response (e.g., precluding an explanation in terms of response-related control). Activation within this region also varied with subjective ratings of decision difficulty. While Pochon and colleagues did not include additional conditions to allow testing of topographic organization, the manipulation of decision conflict in that study led to activation in regions of the DMPFC similar to those associated with decision-related control in our current study.
Moreover, there are initial suggestions that DMPFC is also recruited when individuals make choices that run counter to general behavioural tendencies or strategies (Paulus et al., 2002; De Martino et al., 2006; Hampton and O'Doherty, 2007), although these have heretofore been discussed in terms of decision preferences. Studies involving complex decision making have often focused on identifying brain systems that shape behaviour towards or against particular choices. Yet, it is becoming increasingly apparent that people employ a variety of strategies to simplify the representations of decision problems and reduce computational demands (Tversky and Kahneman, 1974; Payne et al., 1988; Payne et al., 1992; Gigerenzer and Goldstein, 1996). In this study, we show that activation in a similar region of the DMPFC predicts variability in strategy-related control across subjects. Specifically, individuals with greater bias for the balancing strategy exhibited a greater increase in activation for the extreme choices compared to the balanced choices, and vice versa. Strikingly, the cluster within DMPFC associated with strategy-related control was anterior to and functionally distinct from the clusters associated with decision-related control and response-related control. Therefore, a parsimonious explanation for the role of DMPFC in complex decision making is that distinct functional clusters might be associated with distinct aspects of decision making, including strategy preferences and response preparation.
One potential conjecture for functional specialization within the DMPFC is related to differences in connectivity of these clusters to other regions in the brain. A commonly held framework, one advanced in different guises by different theorists, suggests that lateral prefrontal cortex contains a topographic organization along its posterior to anterior axis (Koechlin et al., 2000; Koechlin et al., 2003). More posterior regions, those immediately adjacent to premotor cortex, are associated with setting up general rules for behavior. Conversely, more anterior regions support the instantiation of rules for behavior based on the current context. Findings from functional neuroimaging studies argue for further divisions within anterior prefrontal cortex, such that regions around the frontal pole support relational integration, or the combination of disparate information into a single judgment (Christoff et al., 2001). We speculate that the different regions of DMPFC differ in their lateral prefrontal targets.
Initial evidence for such a functional organization stems from a recent study that demonstrates choice-specific changes in the functional connectivity of the DMPFC with other regions involved in decision-making (Venkatraman et al., 2009). Similarly, Beckmann and colleagues used magnetic resonance diffusion tractography to delineate probabilistic anatomical connectivity of the cingulate cortex to other brain regions (Beckmann et al., 2009). The authors found that the lateral prefrontal cortex exhibited the highest probability of anatomical connectivity with a cluster that corresponds spatially to the region that predicts decision-related control in our study, while the premotor and precentral cortices showed highest probability of connection with a cluster that predicted response-related control in the present study. Under this perspective, the current results point to a generalized role for the DMPFC in cognitive control, but specific computational roles for its subregions depending upon the task demands and current context. Elucidating the functional connectivity of the different clusters in the DMPFC with other brain regions – ideally, using within-task measures derived from the same subjects – may hold the key to fully understanding its role in decision-making.
We thank James Bettman, John Payne, and Mary Frances Luce for discussions about the attribute-balancing paradigm. This research was supported by the US National Institute of Mental Health (NIMH-70685) and by the US National Institute of Neurological Disease and Stroke (NINDS-41328). SAH was supported by an Incubator Award from the Duke Institute for Brain Sciences.
Supplementary Information accompanies the online version of the paper.