|Home | About | Journals | Submit | Contact Us | Français|
The brain stores information in an associative manner so that contextually related entities are connected in memory. Such associative representations mediate the brain’s ability to generate predictions about other objects and events to expect in a given context. Likewise, the brain encodes and is able to rapidly retrieve the affective value of stimuli in our environment. That both contextual associations and affect serve as building blocks of numerous mental functions often makes interpretation of brain activation ambiguous. A critical brain region where such activation has often resulted in equivocal interpretation is the medial orbitofrontal cortex (mOFC), which has been implicated separately in both affective and associative processing. To characterize its role more unequivocally, we tested whether activity in mOFC was most directly attributable to affective processing, associative processing, or to a combination of both. Participants performed an object recognition task while undergoing fMRI scans. Objects varied independently in their affective valence and in their degree of association with other objects (associativity). Analyses revealed an overlapping sensitivity whereby left mOFC responded both to increasingly positive affective value and to stronger associativity. These two properties individually accounted for mOFC response, even after controlling for their interrelationship. The role of mOFC is either general enough to encompass both associations that link stimuli with reinforcing outcomes and with other stimuli, or abstract enough to use both valence and associativity in conjunction to inform downstream processes related to perception and action. These results may further point to a fundamental relationship between associativity and positive affect.
Affective and associative processes are both crucial for an individual’s ability to understand and act in the world. To best anticipate how to respond to a newly presented object, the brain immediately begins the process of object recognition while also extracting the motivational relevance (or affective value) of that object. In so doing, the brain quickly and efficiently activates relevant associations that give rise to focused predictions (i.e., “what other objects or contexts might go with this object?”; Bar, 2004, 2009; Chun & Jiang, 2003; Oliva & Torralba, 2007). The ease with which a stimulus brings to mind other related stimuli and contexts (what we’ll refer to as its associativity)1 is central to research into memory, prospection, imagination and scene construction (Bar, 2004; Bar, Aminoff, Mason, & Fenske, 2007; Barsalou, 2009; Bartlett, 1932; Bower, 2008; Eichenbaum & Fortin, 2009; James, 1890). Similarly, the brain quickly predicts an object’s affective value, in particular, its valence (i.e., “is this something pleasant/approachable or unpleasant/to-be-avoided?”; Barrett & Bliss-Moreau, 2009; Cabanac, 2002; Damasio, 1994; Rolls, 1986; Russell, 2003). These two domains of prediction are supported by vast yet largely non-overlapping psychological and neuroscientific literatures, despite indications that they might share some cognitive and neural mechanisms (Andrews-Hanna, Reidler, Sepulcre, Poulin, & Buckner, 2010; Barrett & Bar, 2009; D’Argembeau, et al., 2009). One region in particular that is consistently implicated in both fields of research is the medial orbitofrontal cortex (mOFC) (meta-analytic summary in Roy, Shohamy, & Wager, 2012).
Cognitive neuroscientific research has shown that regions of ventromedial prefrontal cortex (including mOFC) are involved in tasks that engage associative processing, including those that do so through recall of past autobiographical experiences (e.g., Burianova & Grady, 2007; Denkova, Botzung, Scheiber, & Manning, 2006); imagination of possible future events (e.g., Addis, Wong, & Schacter, 2007; Peters & Buchel, 2010; Szpunar, Watson, & McDermott, 2007); mind-wandering (e.g., Christoff, Gordon, Smallwood, Smith, & Schooler, 2009; Mason, et al., 2007); and through perception of contextually associative stimuli (objects/scenes); (e.g., Bar, 2004; Bar & Aminoff, 2003). These research areas have shown that mOFC is more active as tasks elicit greater associative processing (reviewed in Bar, et al., 2007; Buckner, Andrews-Hanna, & Schacter, 2008; for related meta-analytic summary, see Spreng, Mar, & Kim, 2009).
Other areas of human neuroscience have extensively explored the mOFC’s role in affective processing. This has been studied in the context of perception for stimuli both basic (e.g., tastes, smells; Grabenhorst & Rolls, 2008; Rolls, Kringelbach, & de Araujo, 2003) and complex (e.g., emotional faces, affect-laden scenes; Nielen, et al., 2009; O’Doherty, et al., 2003b); extinction of conditioned emotional responses (e.g., Kalisch, et al., 2006; Milad, et al., 2007); choices between different magnitudes and types of rewarding and/or aversive stimuli (e.g., Kim, Shimojo, & O’Doherty, 2006; O’Doherty, Kringelbach, Rolls, Hornak, & Andrews, 2001); and updating of associated outcome expectations over time (e.g., Daw, O’Doherty, Dayan, Seymour, & Dolan, 2006; Hampton, Bossaerts, & O’Doherty, 2006; Seymour, et al., 2005). Across these domains, mOFC activity typically correlates with whether and/or to what degree a stimulus is more pleasant or preferred (i.e., more positively valenced; meta-analytic summaries in Brown, Gao, Tisdelle, Eickhoff, & Liotti, 2011; Grabenhorst & Rolls, 2011; Kringelbach & Berridge, 2009; Liu, Hairston, Schrier, & Fan, 2011).
Because these areas of research traditionally take place relatively independent of one another and with different types of stimuli/paradigms, the intersection between affective and associative processes, and the shared neural architecture that underlies them, remains poorly understood. This is important not only because it represents a crucial gap in our understanding of how these systems and processes interact, but also because it introduces some degree of ambiguity when interpreting the results of a given study. Specifically, the centrality of mOFC to a large body of research in these two separate fields, combined with the fact that both contextual associations and affective value appear to be processed automatically in perception and cognition (Bargh & Chartrand, 1999; Zajonc, 1980), makes it possible that studies intending only to manipulate affect have observed activation in mOFC related to the associative nature of their stimuli but attributed it to affect, and vice versa (cf. Peters, 2011). We were therefore particularly interested in elucidating whether the mOFC’s presumed stimulus processing role in one of these dimensions is accounted for by the other.
To address this key issue, in the current study we examined both affective and associative processing in the mOFC within a single experiment using stimuli that varied parametrically along each of these dimensions. Specifically, using one paradigm we tested whether the same region of mOFC shows increased activity in response to objects with greater positive affective value (higher valence) as well as to objects that are better able to elicit associations (higher associativity). One possibility is that mOFC performs a processing role corresponding exclusively to either affective or associative processing. If this is the case, our experiment should confirm that mOFC activity is entirely accounted for by either valence or associativity, suggesting that it is possible to reduce one psychological domain to another. A second possibility is that mOFC is a hetero-functional region that separately processes information along these different dimensions, or performs a more general function to which both affective and associative information contribute. In this case, mOFC might track each object dimension independently, resulting in a purely additive influence of the two in BOLD signal measured within mOFC. A third possibility is that mOFC responsiveness to valence and associativity might be dominated by an interaction between these two dimensions.
We first replicate previously separate sets of results showing BOLD activity in mOFC tracking positive value and increasing associativity simultaneously during object perception. Importantly, we extend these findings to show that, from the perspective of mOFC activity, neither of these dimensions is reducible to the other – valence and associativity are independently and additively related to activity in mOFC. Finally, we provide preliminary evidence suggesting that associativity-selectivity and valence-selectivity, at the voxel level, may not simply arise from separate regions within mOFC.
Twenty-three healthy right-handed subjects with no reported history of neurological or affective disorders and normal or corrected-to-normal visual acuity and color discrimination abilities were recruited for the fMRI experiment. Three subjects were excluded for excessive missed trials (>30% of total trials), and one for an incomplete session. Thus, these results will reflect analysis of nineteen (19) total subjects (12 female, age 19–36, mean age: 23.8). Written informed consent was obtained prior to the start of the scanning session, in accordance with a Human Studies Protocol (#2001P-001754) approved by Massachusetts General Hospital.
Images were acquired using a Siemens 3T Trio Tim MR magnet and a 32-channel RF head coil. We acquired functional image volumes as T2*-weighted echo-planar images (EPIs) with the following parameters: 36 interleaved slices, 2200ms TR, 28ms TE, 2.5mm thickness, 0.75mm gap, 64×64 matrix, 200mm FOV (resulting in an inplane voxel size of 3.125 × 3.125 × 2.5mm). Our fMRI sequence and slice prescription was optimized for reducing signal loss and distortion in the orbitofrontal cortex, including the use of a modified z-shim prepulse moment and 30° tilt of our slice prescription counterclockwise of the AC/PC line along the sagittal plane (based on recommendations of Deichmann, Gottfried, Hutton, & Turner, 2003). As a consequence of the limited slice prescription used in order to achieve optimal OFC signal, the most dorsal portions of posterior parietal cortex were not captured in the scan volume for a majority of subjects (13 out of 19). Each subject performed 3 functional runs, each consisting of 149 TRs. Each run included 11s of fixation at the beginning (to allow for the fMRI signal to reach steady-state), and the corresponding 5 EPI volumes were discarded from further analysis. Each session included the acquisition of two high-resolution T1-weighted Multi-Echo MPRAGE (MEMPRAGE) anatomical images (1mm isotropic voxels), which were later averaged together.
Our analyses explored three categories of valence (Negative, Neutral, Positive) and two categories of associativity (Weak and Strong). Our first stage of analysis focused on replicating previously separate findings in as broad a stimulus set as possible. For this stage, associativity analyses were therefore limited to neutrally valenced objects previously normed by Bar & Aminoff (2003). Our second stage of analysis was finer-grained and used object categories that fully crossed levels of valence and associativity, based on stimulus re-norming with a large group of raters. (See also Supplementary Discussion regarding motivation for and potential limitations of using independent ratings.)
Images of strongly and weakly associative objects were initially compiled from a set used in Bar & Aminoff (2003) (available at http://barlab.mgh.harvard.edu/ContextLocalizer.htm) as well as additional stimuli previously compiled using similar norming procedures. These neutrally valenced objects were supplemented with images of positive and negative affect-inducing objects compiled and modified primarily from the International Affective Picture System (IAPS; Lang, Bradley, & Cuthbert, 2008) as well as from online image searches. Each of the resulting 462 objects was viewed by 7 independent raters (14 raters total, each viewing half of the set) who provided valence ratings on a 7-point Likert scale ranging from Very Unpleasant to Very Pleasant (centered on Neutral). The averaged ratings and their variances were used to limit this set to objects consistently rated as positive (pleasant) and negative (unpleasant). These ratings were also used to limit objects previously categorized as strongly/weakly associative to those consistently rated as neutral. After further limiting the set to semantically non-overlapping stimuli, the resulting set of stimuli included 276 objects – 69 each of positive (valence M ± SD: 5.71 ± 0.43), negative (2.18 ± 0.39), weakly associative neutral (4.14 ± 0.45), and strongly associative neutral (4.18 ± 0.45). This set of 276 objects was used in the first stage of analysis. To allow for the second stage of analysis, each of these objects was re-normed for valence by an average of 39 new raters (range: 34–43), using the same procedure as above, and the stimulus set further limited to orthogonalize these ratings against ratings of associativity collected for the same objects (as described in the following section; see also Table 1).
The set of 276 objects described above was presented to an independent set of raters to determine the ease with which each image elicited associations with other objects. Each image was presented to an average of 39 raters (range: 35–47; 116 individuals each viewed approximately one-third of the total image set). Each image was presented separately and while it remained on the screen raters were asked to try their best to type the names of three separate objects that they associate with it. For instance, an image of a washing machine might elicit ‘detergent,’ ‘clothes,’ and ‘dryer’ as associates. Raters’ response time on this task was not limited, and they were required to produce at least one associate per image. This procedure allowed us to derive our key measures of associativity for a given image, including the number of times that object failed to produce all three associates (i.e., the proportion of associate slots left unfilled for a given object across participants) [ease/difficulty of association] and the response time (RT) to associate production (normalized for both word length and each rater’s mean typing rate) [speed of association]. These two measures – ease and speed of association – were significantly correlated across objects (r(276) = 0.53, p<0.001), and so were normalized and averaged together to form a single Associativity Index (with higher values representing objects that produced associations more easily and faster, thus Strongly Associative). If, in the example above, the washing machine consistently produced only one or two associates and/or those associates took a long time to generate, its associativity index would be low and we would describe it as weakly associative. If the reverse were true (i.e., all three associates were consistently given and relatively quickly), it would be considered strongly associative. While some raters participated in both valence and associativity norming, a given rater only viewed each object once, providing either a valence or an associativity rating for that object.
For both analysis stages these variables are discretized into categorical rather than continuous variables. The motivation for this was twofold. First, this is the approach used by a number of studies whose findings we were most closely attempting to replicate, both for associativity (e.g., Bar & Aminoff, 2003) and valence (e.g. Nielen et al., 2009; Ritchey et al., 2011; Sass et al., 2011). Second, because valence and associativity turned out to be intercorrelated (see Results) the second set of analyses tested for fully independent effects of the two object dimension on mOFC activity by forming stimulus categories matched along the potentially confounding dimension. This approach to orthogonalizing our variables of interest, which we determined to have less potential of being overly conservative than alternatives, partially guided the decision to use categorical variables for all of our main analyses (see also Supplementary Analysis).
The paradigm employed a rapid event-related design. Target images were color photographs (256×256 pixels) of objects in isolation presented on a white background. Each target image was presented briefly (150ms), and immediately followed by a colorful mask presented for 100ms (see Fig. 1). This was done in order to encourage the participant to bring to mind the object and its internal associations (affective and contextual) and minimize the degree to which the participant was modulating attention toward the image of the object in front of them, and the individual features thereof, as well as to encourage the participant to answer relatively automatically rather than to consciously deliberate over potential responses. A red fixation cross then appeared signaling the start of the response period and turned black after 1500ms, signaling the end of the response period. The black fixation cross remained on the screen for the duration of the inter-trial interval (ITI), which ranged from 200ms to 9250ms (to allow a jittered ITI in multiples of TR length and jittered stimulus presentation from the start of each TR). The fMRI session consisted of 276 unique trials, pseudo-randomly ordered across three functional runs, in addition to 28 practice trials (practice trials presented objects not included in the main task). Each target image was presented only once.
Subjects were instructed to rate how common each object was on a five-point scale ranging from not at all common to extremely common. This task was chosen in order to require some high level of object recognition without focusing attention explicitly on either affective or associative qualities of the object, an assumption we later verify by comparing ratings of commonality across image categories. Responses were provided on a five-button MR-compatible response box. The order of the five-point scale was counterbalanced across subjects to prevent confounds between rating and motor mapping, and subjects practiced the appropriate mapping to proficiency before the practice trials began. Stimulus presentation and response collection was performed using Psychtoolbox (http://www.psychtoolbox.org; Brainard, 1997) running on Matlab (http://www.mathworks.com), controlled by a MacBook Pro laptop with a monitor resolution of 1024×768 and refresh rate of 60 Hz.
Structural and functional imaging analyses were performed using the Freesurfer and FS-FAST analysis tools and processing stream developed at the Martinos Center for Biomedical Imaging (http://surfer.nmr.mgh.harvard.edu). Data from individual fMRI runs were first motion corrected using the AFNI motion correction algorithm (http://afni.nimh.nih.gov/afni/) in which all images were aligned to the first image of the first functional run. The data were then spatially smoothed using a Gaussian full-width at half-max (FWHM) of 6 mm. The first five volumes were removed from each fMRI run to allow for signal stabilization. The intensities for all runs were globally rescaled such that the in-brain mean intensity was 1000. Signal intensity for each condition was then computed and averaged throughout all the runs. Each subject’s fMRI volumes were also co-registered to their own high-resolution structural volume, and a semi-automated procedure was used to segment this structural data into gray/white matter components and to extract the outer cortical surface for each hemisphere as a topologically preserved spherical representation (software and documentation is available at http://www.nmr.mgh.harvard.edu/freesurfer).
For the first analysis stage, all events were modeled according to the four object categories, and for the second analysis stage an additional condition modeled any stimuli excluded from the six more conservatively defined categories. Analyses reported include any trials where participants failed to respond in time (which were infrequent and not influenced by object category; see Behavioral Results), but all results remain unchanged if these are modeled as a separate condition (see also Supplementary Analysis). The estimated hemodynamic response was defined by a gamma function of 2.25s hemodynamic delay and 1.25s dispersion. Data were then tested for statistical significance for each individual (first-level) and contrast maps were constructed comparing the BOLD estimates for each condition. While the a priori focus of our analyses was on a specific set of ROI’s (described below), to allow comparison with previous findings and test for overlap between key contrasts whole-brain random effects analyses were performed at the group (second) level with an omnibus F-test and individual t-tests over contrast maps generated at the first level. In order to reduce the influence of noise at the first-level, these random effects analyses weighted first-level contrast maps by the inverse of their variance (Thirion, et al., 2007). Whole-brain statistical maps were corrected for multiple comparisons using a cluster significance threshold of p<0.05 (corrected), with a cluster-defining (voxelwise) threshold of p<0.01. Since primary analyses are restricted to a priori ROI’s, we chose a standard but relatively liberal cluster-defining threshold in order to describe the full extent of our network of activations. However, because cluster-correction was performed in volumetric space, activations shown on the cortical surface (Fig 3, left) are not cluster-corrected but instead set at a more conservative voxelwise threshold of p<0.001. All whole-brain analyses were performed for confirmatory rather than exploratory reasons (i.e., ROI selection was based strictly on anatomy rather than functional activation). In order to visualize overlap between individual contrasts, a conjunction was performed over these corrected maps using the minimum statistic (Nichols, Brett, Andersson, Wager, & Poline, 2005). While ROI analyses were performed on cortical surfaces in each participant’s native space, group-level analyses were performed after transforming each participant’s data into normalized (Talairach) space.
Given a priori predictions about the involvement of the two regions in our task, we used Freesurfer’s automated segmentation and parcellation algorithms (Desikan, et al., 2006; Fischl, et al., 2002; Fischl, et al., 2004) to define anatomical regions of interest (ROI’s) within left and right mOFC based on each subject’s individual anatomy (see Fig. 3b for an example). The parcellation algorithms used gyral and sulcal landmarks from each individual’s surface anatomy to define an ROI on each hemisphere which extended rostrally/caudally according to the boundaries of the medial orbital gyrus, was bounded on the orbital surface by the midpoint of the olfactory sulcus, and was bounded on the medial surface by the inferior boundaries of cingulate and superior frontal gyri (Desikan, et al., 2006). Averaged beta weights were extracted from this ROI for each of our conditions, and then converted to percent signal change values. In order to determine whether BOLD activity in this region tracked one or both of our parameters of interest (or their interaction), these values were then entered into a mixed-effects Valence x Associativity ANOVA model, with subject as a random effect. We used a two-tailed p-value less than 0.025 (p<0.05, corrected for the left and right hemisphere a priori ROI’s) as our significance threshold in these ROI analyses, and we note the cases in which this correction was applied when reporting the results of the ROI analyses.
In order to perform a more conservative test against the possibility that Valence- and Associativity-sensitivity revealed in mOFC voxels was actually occurring in segregated sets of voxels within our anatomically defined ROI, we extracted averaged percent BOLD signal change estimates for each valence condition from the twenty most associativity-selective voxels in left mOFC. Selectivity was assessed on a within-subject basis by rank-ordering voxels (without consideration for spatial contiguity) based on p-values from the omnibus F-test for a voxel-wise GLM which included only category labels for Strong and Weak associativity. We then tested for a significant effect of valence category in these associativity-selective voxels using a mixed-effects ANOVA. We also performed the reverse analysis, testing for associativity-selectivity in the most valence-selective voxels. To test for robustness, this analysis was repeated for samples of between one and 100 peak voxels.
Participants performed a simple task requiring the recognition of visually presented objects while undergoing fMRI scans (Fig. 1). Images of isolated objects were presented briefly (150ms) and participants were asked to judge how common the object was. All objects were independently rated for hedonic valence and associativity, and planned ROI analyses tested for changes in mOFC BOLD signal across levels of valence and associativity. Our first stage of analysis focuses on four object categories (Negative, Neutral-Weak, Neutral-Strong, Positive; for examples, see Fig. 2a), defined across the full set of stimuli viewed in the scanner. Our second stage narrows this stimulus set to allow for fully crossed and orthogonalized analyses of all six possible categories (Negative-Weak, Negative-Strong, Neutral-Weak, Neutral-Strong, Positive-Weak, Positive-Strong; Fig. 4a).
Participant ratings and reaction time (RT) were entered into separate mixed-effect ANOVAs, with participant as a random effect. Based on the four initial categories, we found a significant overall main-effect of condition on ratings of commonality (F(3,54) = 101.2, p<0.001), such that negative objects were consistently rated less common on average than objects in the other three conditions (see Supplementary Materials for secondary analyses controlling for this). There was no significant difference between commonality ratings for the remaining three conditions based on post-hoc contrasts (F(3,100) = 0.92, p>0.40). There was no significant main effect of condition on response time (RT) (F(3,54) = 2.0, p>0.10), suggesting no clear difference in task difficulty across the four conditions. The lack of commonality differences (with the exception of Negative objects) and lack of RT differences remained true after trials were later re-sorted for the second analysis stage (see Table 1 legend). Participants missed relatively few trials (median = 2.2%, mean ± SD = 5.7 ± 6.9%), with no effect of condition (F(3,54) < 0.10).
An additional finding emerged related to the independent ratings of our two object dimensions of interest. We found that an object’s valence and degree of associativity were moderately correlated (r(276) = 0.31, p<0.0001; see Supplementary Fig. 1), such that objects that were rated as more positive tended to be more strongly associative. While this relationship is interesting in its own right, and further motivates the need for disambiguation of previous mOFC findings, we also control for this as a potential confound in our second set of analyses.
Our analyses focused on left and right mOFC a priori ROI’s using surface anatomic landmarks in each participant’s native space (see Methods and example in Fig. 2b). Since we had no clear a priori hypothesis about one hemisphere or another, our ROI analyses were Bonferroni-corrected for the dual a priori hypotheses. We extracted estimates of BOLD signal from these mOFC ROI’s for each of the four broader stimulus categories and found a significant main effect of condition (Fig. 2c; left: F(3,54) = 19.3; right: F(3,54) = 8.52, p’s<0.001, corrected). A series of planned post-hoc contrasts confirmed that both associativity and affective value engaged mOFC. Directly replicating previous studies of associativity (Aminoff, Schacter, & Bar, 2008; Bar & Aminoff, 2003), we found that neutral objects with strong associations significantly increased BOLD activity in mOFC, when compared to weakly associative objects (left: F(1,54) = 33.3; right: F(1,54) = 15.4; p’s<0.001, corrected). Mirroring findings for a number of different kinds of affectively valenced stimuli (Brown, et al., 2011; Chib, Rangel, Shimojo, & O’Doherty, 2009; Lebreton, Jorge, Michel, Thirion, & Pessiglione, 2009; Nielen, et al., 2009; O’Doherty, Critchley, Deichmann, & Dolan, 2003a; Sass, et al., 2011), BOLD activity in mOFC also increased significantly for Positive objects compared with the neutrally valenced objects (i.e., Neutral-Strong and Neutral-Weak combined) and compared with the Negative objects (left: F(3,72) = 8.2, p<0.001, corrected; right: F(3,72) = 3.4, p<0.05, corrected; Fig. 2c).
While our analyses focus explicitly on individually-defined ROI’s, results of a confirmatory group whole-brain analysis (Fig. 3, Supplementary Table 1) show that activations in mOFC appear prominently in each of the three whole-brain corrected contrasts of interest. These analyses also reveal expected patterns of coactivation for associativity in retrosplenial and parahippocampal cortices, which have been observed across all previous studies of associativity (Bar et al., 2007) and many studies of episodic and autobiographical memory more generally (Buckner, Andrews-Hanna, & Schacter, 2008; Roy, Shohamy & Wager, 2012). And regions coactivated with mOFC in valence contrasts, including dorsal mPFC and amygdala, are also commonly coactivated with this region across studies of affective processing, particularly in the case of the amygdala (Lindquist & Barrett; Roy, Shohamy & Wager, 2012). A conjunction of these contrasts confirmed that the greatest degree of spatial overlap fell the within mOFC and, in particular, left mOFC (Fig. 3e). One immediate implication of these results is that the differential activation in mOFC to strongly versus weakly associative objects reported in previous research is robust to more careful control for valence. However, this still leaves unanswered the question of whether the effect of valence on mOFC is mediated by associativity. We turn to this question next in our second analysis stage. Because our investigation was primarily concerned not with the function of all of mOFC but specifically with those regions related to both affect and associations, remaining analyses focus on left mOFC, where selectivity was substantially stronger for both dimensions (as discussed above and further confirmed by a significant ROI x condition interaction: F(3,54) = 10.13, p<0.0001). Bonferroni correction continued to be applied to adjust for our two a priori hypotheses.
Subsequent analyses revealed that the effects of valence and associativity on mOFC activity are independent of and additive with one another, indicating that the mOFC’s role in affective processing is not attributable to associativity, or vice versa. To confirm that mOFC activity was independently influenced by both object properties, this more conservative second stage of analysis focused on a subset of objects that were classified into one of six possible combinations of valence and associativity level (Negative-Weak, etc.). These objects were also selected so that categories at the same level of associativity were matched for valence, and vice versa, to address any potential confounds between the two (see Table 1 and examples in Fig. 4a). We were then able to run a 3 (valence: Positive, Negative, Neutral) × 2 (associativity: Strong vs. Weak) ANOVA for BOLD signal in the left mOFC. We found that left mOFC BOLD activity revealed main effects of valence (F(2,36) = 4.88, p<0.05, corrected) and associativity (F(1,18) = 25.7, p<0.0005, corrected; valence x associativity: F(2,36) = 2.70, p>0.15, corrected), indicating that both properties contributed to independent increases in mOFC activity (Fig. 4b).
These results thus far suggest that activity in mOFC independently increases with positive affective value and increased associativity. One possibility that remains is that these results arise from two segregated but partially overlapping regions within left mOFC, one selective only for valence and the other selective only for associativity. This hypothesis can not be disconfirmed by comparing peak coordinates or exploring activation overlap at differing thresholds because it is always possible that voxels most selective for one parameter are still significantly selective for the other parameter, but that their significance fails to meet the given threshold. In order to provide evidence to militate against the possibility that our independent effects of affect and associativity originate from segregated regions within mOFC, we instead identified the twenty voxels within mOFC that were most selective for associativity (on a within-subject basis) and extracted from each participant the average percent BOLD signal change for each of the valence conditions. A mixed-effects ANOVA (treating participant as a random effect) revealed a significant effect of valence within these voxels (F(2,36) = 4.98, p<0.02). We then did the same for BOLD estimates of strongly versus weakly associative objects within voxels most selective for valence and again found a marginally significant effect for associativity (F(1,18) = 4.16, p<0.06) (Supplementary Fig. 2a).2 To ensure that neither set of estimates benefited from error variance accounted for by the other parameter, voxel selection and beta estimate extraction were derived from statistically independent general linear models over the same data set (one GLM accounting only for levels of associativity and the other only levels of valence). These findings provide preliminary support for the hypothesis that overlapping populations of voxels within the mOFC were independently responsive to both an object’s affective value and the strength of its associativity with other objects/contexts.
The mOFC has been implicated in research on both affective and associative processes, raising the important question of whether one set of proposed functions better accounts for the other. In the current study, we measured mOFC BOLD activity while participants viewed objects that were varied orthogonally in affective valence and associativity. Replicating findings from separate literatures, we demonstrated that activity in the same region of left mOFC was proportional to an object’s valence (specifically, whether it was positively valenced or not) (Brown, et al., 2011; Grabenhorst & Rolls, 2011; Kringelbach & Berridge, 2009) and to its associative strength (Aminoff, et al., 2008; Bar & Aminoff, 2003). Crucially, we found that activation in this mOFC region was not reducible to either valence or associativity, even after controlling for interdependencies between the two dimensions. Both properties independently and additively modulated left mOFC activity, despite the fact that neither was integral to the task at hand. Moreover, sensitivity to an object’s affective value and its associative strength seemed to arise from overlapping rather than separate populations of voxels within the anatomically defined mOFC.3
In terms of both structure and connectivity, it is not surprising that the mOFC seems important for both contextually associative and affective predictions. The mOFC’s proposed role in object recognition is supported by its strong connectivity with magnocellular visual processing regions along the dorsal stream (Barbas, 2007a, 2007b; Carmichael & Price, 1995b); the mOFC also sends projections back to the ventral visual stream, potentially influencing early stages of object processing in inferotemporal cortex (Kveraga, Ghuman, & Bar, 2007; Kveraga, et al., 2011; Summerfield, et al., 2006). The mOFC has also been described as a center of integration for information regarding affective significance or value, and it sends and receives projections from regions involved in processing and regulating the autonomic physiology associated with motivated states, including the amygdala, hypothalamus, cingulate cortex, and the brainstem, as well as indirect projections from the ventral striatum (Carmichael & Price, 1995a; Öngür & Price, 2000). Furthermore, because the mOFC receives afferent projections from all sensory modalities (largely via its dense connections with lateral aspects of OFC), it is optimally situated for linking representations to inform judgments (Grabenhorst & Rolls, 2011; Murray & Wise, 2010; Wallis, 2007).
Taken together, previous research suggests that the mOFC engages in predictive processing across both cognitive and affective domains by allowing for associations between a given stimulus and both its higher-order sensory relations (i.e., other stimuli/contexts) and affective outcomes (i.e., valence) to be rapidly computed and used for the proactive generation of the corresponding predictions. One possible interpretation of the current results is therefore that mOFC plays different roles for a number of independent psychological functions, or that its role is fundamentally associative but that these associations take on a number of different forms. This model gains support from the fact that the diverse afferents to mOFC described above make it an ideal information-processing hub for independently collecting and processing information related to stimulus valence (e.g., from lateral OFC and the amygdala) and to associated stimuli and contexts in memory (e.g., from hippocampus and parahippocampal cortex). Under this interpretation, our results would be a natural consequence of this region’s putative role in separately maintaining and/or transferring information related to stimulus associations, defined broadly to include associations between stimuli and other stimuli/contexts, and associations between stimuli and outcome values.
In contrast, another equally plausible interpretation is that the convergence of pleasantness and associativity within the mOFC suggests that this brain area subserves a more unified purpose to which both of those psychological properties relate. The affective and reinforcement literature has suggested that such a unitary function is generating positive (and/or negative; Damasio, 1994, 1996) affective states and/or encoding positive affective value in a stimulus (Kringelbach & Berridge, 2009; Lebreton, et al., 2009; Rolls, 1986). Building in part on psychological mechanisms previously theorized to underlie emotion (Barrett, 2006; Cabanac, 2002; Damasio, 1994; Rolls, 1986; Russell, 2003; Weiskrantz, 1968), recent research into decision-making and reinforcement learning has suggested that the kind of affective value encoded by mOFC represents the expected value of options under consideration, and can be more appropriately considered as an abstract “currency” for guiding decisions (Kable & Glimcher, 2009; Rangel & Hare, 2010; Wallis, 2007). From this perspective, our findings would raise the important question of whether determining an object’s overall expected value may include/require consideration of both its affective value as well as its associative value (i.e., the ease or multiplicity of association between it and other objects/contexts in memory). This estimate of associativity can aid the expected value computation in a number of possible ways, including as a proxy measure of possible states and physical actions that must be evaluated in relation to this stimulus (Rangel, Camerer, & Montague, 2008). From the perspective of the processes of recall and prospection, a stimulus may also gain some currency for downstream information processing if it is better able to link up stimuli in memory or allow fluid production of future plans through its promiscuous binding to other stimuli (i.e., a type of informational processing fluency; cf. Kurth-Nelson, Bickel, & Redish, 2012; Murray & Wise, 2010). In other words, a stimulus might be encoded as more valuable merely because it elicits a large number of associations (Bar, Shenhav, & Devaney, submitted). For the same reason, mOFC activity may increase with the meaningfulness (and identifiability) of an object (Chaumon, Kveraga, Barrett, & Bar, under review) because of the associative activation that a meaningful object elicits. While the exact kind of value evinced by associativity is still a matter of speculation, if it turns out to be the case that object associativity is another kind of value being encoded in mOFC then one would indeed expect to see the independent encoding for valence and associativity found in our data.
Our data and the interpretations offered above are also broadly consistent with a recent review by Roy, Shohamy, & Wager (2012). In integrating across research on memory and affect, the authors proposed that the overarching function of this region (specifically, ventromedial PFC) may be described as determining a stimulus’ “affective meaning,” combining both conceptual and affective inputs towards this goal (e.g., Kumaran, Summerfield, Hassabis, & Maguire, 2009). Under this view, the mOFC may integrate value representations with the actual stimuli/contexts associated with an object, consistent with the mOFC’s proposed role in representing specific states and stimulus contingencies relevant to learned reward values (i.e., “model-based” learning; Bornstein, Nylen, & Steele, 2011; Schoenbaum, Takahashi, Liu, & Mcdannald, 2011). The result would be increased mOFC activity for increasing numbers of readily available stimuli/contexts (i.e., stronger associativity). In this respect, future studies should consider whether functional connectivity between mOFC and regions coactivated by increasing associativity (e.g., parahippocampal and retrosplenial cortices) predicts the degree to which choices made in the scanner reflect model-based versus model-free learning and inference (Daw et al., 2011). It would also be interesting to see if the reverse is true for mOFC connectivity with regions like the amygdala.
It further remains to be determined how value-related encoding in mOFC influences cognitive processing in general and the process of object association in particular. Our finding that stimulus pleasantness and associativity are correlated suggests that one possible outcome of mOFC valuation might be to modulate the extent of associative processing for a given stimulus (Bar, 2009). In this view, positive affective states are linked with disinhibited association (i.e., a greater readiness to form or activate associations between stimuli) and negative affective states are directly related to the inhibition of stimulus-stimulus associations. While the associations attached to a stimulus are more likely to be stored in the medial temporal cortex and/or sensory cortices, according to this model mOFC could exert inhibitory control over how much associative activation is afforded downstream (Bar, 2009). In this kind of model, mOFC would be involved in overseeing the scope of associative processing such that positively valenced signals would trigger a broadening of this scope, and signals indicating negative valence in an object would result in narrowing this scope through inhibition of associative processing by mOFC (cf. Sass, et al., 2011), similarly to what is seen in the narrowing of attentional focus for negatively valenced stimuli (Baddeley, 1972; Gasper & Clore, 2002). While still largely hypothetical, this model gains support from recent findings in rhesus macaques showing that nearby perigenual ACC (BA 32) projects to regions of parahippocampal cortex in ways that would allow it to exert excitatory and inhibitory control over the local information processing circuitry (Bunce & Barbas, 2011). Moreover, previous behavioral studies have identified a link between positive mood states and the generation of broader associations (Brunye, et al., under review; Clore & Huntsinger, 2007; Fredrickson, 2004; Isen, Johnson, Mertz, & Robinson, 1985; Mason & Bar, 2011).
Our study carries broad implications for research into affective as well as associative processing. In particular, as it concerns research into the functions of mOFC, both types of stimulus property should be taken into consideration at the design and interpretation stages. While our results suggest that the processing of associativity and valence can vary independently of one another in mOFC, we also find that valence and associativity are correlated and activate overlapping regions in mOFC. Therefore, mOFC activity observed while varying stimuli along only one of these dimensions may be partially or even largely, though not necessarily entirely, attributable to the other. More generally, the nature of the relationship between ratings along these two dimensions (at the psychological level) remains a topic of great importance. It may be the case either that positively valenced objects facilitate associative activation and the generation of predictions, or that the greater availability of associates for an object makes it perceived more positively or less negatively, in line with the proposal that broader associative activation is linked with better mood (Bar, 2009). Gaining a deeper understanding of how we encode and react to these two stimulus dimensions, and their interaction, will lend importantly to our broader understanding of a more fundamental relationship between factors that have been traditionally dichotomized (and provided a perhaps false appearance of independence) into the domains of cognition and affect.
We are grateful to J. Boshyan, M. Rosen, and K. Shepherd for assistance in data collection, T. Benner for technical guidance, E. Aminoff for assistance in stimulus collection, and S. Gagnon for assistance with analysis and helpful comments on the manuscript. This work was supported by an NSF Graduate Research Fellowship awarded to A.S., National Institutes of Health Director’s Pioneer Award (DP1OD003312) to L.F.B., and NIH R01EY019477 and NSF 0842947 to M.B.
1We note that while affective predictions are also based on associations – namely associations between stimuli and reinforcing outcomes – we will use the term associative to refer to stimulus-stimulus and/or stimulus-context associations exclusively (irrespective of object valence).
2These same general results also hold when varying the number of voxels sampled between a single peak voxel and 100 peak voxels (Supplementary Fig. 2b).
3As with any neuroimaging study, our results cannot speak to whether such functional segregation exists at the neuronal level. Previous electrophysiological findings suggest that intermixed neurons in OFC and mPFC can code for different stimulus properties (e.g., identity, value; Padoa-Schioppa & Assad, 2006) and other decision variables (e.g., effort required, probability of outcome; Kennerley, Dahmubed, Lara, & Wallis, 2009), distinctions that can be lost at the level of fMRI. Therefore, while we think it interesting to note that we didn’t find clear evidence for functional segregation in mOFC based on the level of resolution at which neuroimaging studies are performed, we are cautious not to interpret the results as evidence of shared function of the individual neurons in this region. With this in mind, our discussion of these findings is still consistent with previous accounts of overarching regional function, including in research mentioned above where different sensitivities of individual neuronal populations were found (Padoa-Schioppa & Cai, 2011; Wallis & Kennerley, 2011).