Is the Whole the Sum of its Parts? Neural Computation of Consumer Bundle Valuation in Humans

Abstract

Humans are often tasked with making decisions about bundles of multiple items and very little is known about how the human brain aggregates, computes, and represents value in such cases. We investigated how the brain evaluates consumer items, both individually and in bundles, and how this activity relates to choice behavior. Human participants (N = 14; 7 female, 7 male) completed a deep-functional MRI protocol while we elicited behavioral valuations for single and bundled items. Behaviorally, we find that bundle values are sub-additively discounted compared to the sum of individual item values. Neurally, we find that the same distributed network in prefrontal cortex computes the value of a bundle and its constituent individual items, but the value representation undergoes a normalization that actively rescales across bundle and single item contexts. These findings suggest that generalized value regions contextually adapt within a valuation hierarchy, as opposed to utilizing an absolute value code.

Significance Statement

Humans routinely choose among bundles—meals, device-plus-plan, and vacation packages—yet how the brain evaluates multi-component options is unclear. Behaviorally, we show that bundle values are sub-additive relative to the sum of their parts. Using a deep-functional MRI protocol, we find that ventromedial prefrontal cortex and a distributed prefrontal network encode value for both single items and bundles, indicating a condition-independent, general value code. Critically, this code is scaled by context: activity is attenuated for bundles, and representational analyses favor models with relative, normalized value over absolute value. These findings reconcile a putative “common currency” with divisive normalization, demonstrating that value representations generalize across objects yet rescale to their distribution—an important principle for multi-component choice.

Introduction

In daily life, humans evaluate options with multiple components, such as meals, investment funds, or vacation packages. These decisions involve trade-offs between components, making bundles fundamental in economic analysis of consumer choice (Mas-Colell et al., 1995). Normatively, bundle evaluation involves valuing each component and aggregating them: the whole is the sum of its parts, formalized in attribute integration theories (Bettman et al., 1998).

The brain appears to encode partially distinct representations of individual attributes, integrating them to compute overall stimulus value in a “common currency” (Chib et al., 2009; Kahnt et al., 2011; Levy and Glimcher, 2012; Bartra et al., 2013; Lim et al., 2013; Clithero and Rangel, 2014; Suzuki et al., 2017; O’Doherty et al., 2021). However, research on value signals in bundling is inconclusive (Zyuzin et al., 2022), possibly due to the complexity of valuing and integrating distinct items. Monkey orbitofrontal cortex (OFC) neurons have been found to reflect bundle valuation, but responses to the bundle relative to its distinct components are unknown (Pastor-Bernier et al., 2019). This study examines if attribute integration theory generalizes to bundles, where components represent a different level of abstraction, as choice objects. The lateral OFC (lOFC) is implicated in encoding stimulus attributes, with inputs to ventromedial prefrontal cortex (vmPFC) proposed to implement value integration (Lim et al., 2013; Suzuki et al., 2017). Values for food and consumer items are reported in ventral aspects of vmPFC, with integrated value in a more dorsal area (McNamee et al., 2013; Clithero and Rangel, 2014). Here we test whether item and bundle values exhibit a hierarchical representation across prefrontal areas with separate regions encoding individual item and bundle values respectively. Alternatively, the same brain areas could encode the value of both single and bundled items, suggesting a generalized valuation process.

If there is a generalized valuation signal, a next question is how single item and bundle value codes are scaled. The primate brain appears to implement a relative value code adapting to the choice set context (Padoa-Schioppa, 2009; Louie et al., 2015; Khaw et al., 2017; Rustichini et al., 2017; Zimmermann et al., 2018). This compression can maintain resolution for accurate discrimination due to limited capacity in physiological systems (Polania et al., 2019). Cox and Kable (2014) found adaptation to distributions of items: wider ranges attenuate blood-oxygen-level-dependent (BOLD) responsiveness to stay within bounds. Behavioral evidence suggests similar re-scaling for attributes before aggregation (Dumbalska et al., 2020).

We hypothesize that value regions switch between single and bundle contexts through re-scaling the value code according to current context values, rather than using the same absolute scale across trials. In our experiment, item and bundle trials are randomly interleaved, preventing slow adaptation to one context. Context-specific normalization allows optimal encoding and discrimination within each context, supporting behavioral flexibility, while divisive normalization (Carandini and Heeger, 2012; Louie et al., 2015) and range normalization models offer potential mechanisms (Rustichini et al., 2017). Both models predict attenuated neural activity for bundles relative to single items, and we use neural data to compare these normalization codes to alternatives.

In our study, participants made choices about food items, non-comestible consumer goods, and bundles (Fig. 1), during functional MRI (fMRI). To maximize our ability to discriminate single and bundle item values, we used an elongated deep-fMRI design whereby each participant performed the task on three separate days, resulting in over 900 trials, considerably more than in typical fMRI designs.

Figure 1.

Experimental design. a, Willingness-to-pay (WTP) task (outside scanner each day). Participants reported their willingness-to-pay for either items or bundles of two items in a BDM auction. b, Choice task (inside the scanner, each day). During single item trials, a choice was made between an item and a reference monetary amount equal to the median bid of single item trials in the WTP task. Similarly during bundle trials, a choice was made between a bundle and a reference monetary amount equal to the median bid of WTP bundle trials. Participants had up to 5 s to make a choice indicated by a right or left button press. c, Depiction of the deep-fMRI scanning protocol. Single item and bundle trials were randomly interleaved within a run on each day. The experiment involved 3 d of scanning, 5 runs of the task per day, for a total of 15 runs.

We utilized both univariate and multivariate analyses. While the former allows investigation of the broad spatial topography of value codes, multivariate analyses can uncover more subtle distributed representations (McNamee et al., 2013; Suzuki et al., 2017), such as if single and bundle items are represented in a distributed fashion within the same region. We also used representational similarity analysis (RSA) to compare the possible implementation of different models of normalization on the value signal, for which RSA is well suited.

Materials and Methods

Participants

Participants (N = 14) were recruited from the general population through the Caltech Brain Research Participant System (7 females, 7 males, 24.9 ± 3.74 years, mean ± s.d.). Though we used a deep-fMRI design in which participants were scanned extensively over 3 d so as to enable robust estimates of values code in each individual participant, a sample size of 14 was also sufficient to support group level inference. Participants did not have any food allergies and were not dieting at the time of the experiment. They were given a participation fee of $40 ($20 an hour), in addition to receiving monetary, food, and non-comestible consumer goods as rewards depending on their choices in the experiment. Each subject gave their informed consent, and the study was approved by the Institutional Review Board of the California Institute of Technology.

Stimuli

Across participants, 70 food items and 40 non-comestible consumer goods were used as stimuli in the experiment. Food items included fruits, snacks, and mains (including microwaveable meals) that are available at local grocery stores. Consumer goods included a diverse array of items under $40 in price, including cell phone chargers, kitchen items, Caltech memorabilia, and books. Many of these items have been used in previous studies (Suzuki et al., 2017). The full list of items can be found in Tables S1 and S2.

Experimental tasks

Participants performed the experiment in three sessions on three separate days in order to maximize the amount of fMRI data within-subject. On each day, participants first performed a willingness-to-pay (WTP) task outside the scanner, then performed the choice task in the scanner where they were asked to choose between an item or bundle of two items versus a reference monetary amount. Participants were asked to refrain from eating 4 h before the experiment in order to ensure that food items were valuable to them. Compliance was confirmed by self-reports, that asked whether a participant complied, in addition to hunger ratings.

Each participant performed the WTP and choice tasks on a total of 40 unique items presented throughout the experiment. On a given day, the participant was presented 20 items, composed of 10 food items and 10 non-food items. Ten of these items were presented in all 3 d, and 10 new items were introduced every day. To construct bundles, every item was paired with each other, including pairs of the same items. Thus, 210 bundles were included each day (20 choose 2 = 190 + 20 pairs of the same item). On the final day of the experiment after the choice task, outside the scanner participants rated how familiar they were with each of the 40 items. For each item, the participant indicated their familiarity with the item on a continuous scale from “not at all’ to “very much” by moving a red pointer, with no time constraint as described previously (Suzuki et al., 2017).

WTP task (outside the MRI scanner)

Participants completed an un-timed Becker–DeGroot–Marschak (BDM) auction task to measure their WTP for items and bundles, with a procedure similar to previous studies (Chib et al., 2009; McNamee et al., 2013; Suzuki et al., 2017). The BDM auction is a reliable incentive-compatible method to elicit subjective values for items (Becker et al., 1964). Participants were endowed with a $20 budget in cash, and instructed that they can use this cash to purchase items from our laboratory store (and keep the money they do not use). In each trial, an item or bundle was shown and the participant was asked to type in how much they would be WTP from $0 to $20 for that item/bundle (Fig. 1). Participants first bid on the individual items (20 each day) and then bid on the bundles (210 each day).

Each trial was to be treated independently, as a random trial from the entire experiment was selected at the end of the experiment. If the selected trial was from the WTP task, the participants’ bid on that trial was then compared against a randomly generated price (uniform probability from $0 to $20), and if their bid was greater than or equal to that price, they received the item(s) and paid the corresponding price with their $20 budget. If their bid was less than the price, they did not receive the item(s), and they did not have to pay anything. Participants were explicitly instructed about this auction procedure, and about how the optimal strategy is to bid their true subjective value for a given item/bundle. With a questionnaire, we confirmed that participants understood the mechanism of the auction.

Choice task (inside the MRI scanner)

In the scanner, participants made choices involving the items and bundles previously bid on in the WTP task. Each trial involved a binary choice between an item or bundle and a reference monetary amount (Fig. 1b). The reference monetary amount was equal to the participant’s median WTP bid for that category (individual items were chosen against the median bid on individual items and bundles were chosen against the median bid on bundles). This ensured that the participants would choose the reference monetary amount about half the time and choose the item or bundle half the time. For a trial, the stimulus appeared in the middle of the screen, and the word “ITEM” appeared on the bottom left or right with equal probability while the reference monetary amount in “$X” appeared on the other side. Participants selected their choices with a button box, with the leftmost button indicating choosing the left option and vice versa for the rightmost button. Participants had 5 s to make a decision, after which their choice was presented on the screen for 0.5 s and followed by a jittered intertrial interval (intertrial interval phase, 2–7 s).

On each of the 3 d, participants were scanned for five runs. Each run included 62 trials, where each of the 20 individual items was presented once per run, and each of the 210 bundles in a day was presented once on that day. On day 1, anatomical scans were collected after the choice task.

fMRI data acquisition

The fMRI data was acquired on a Siemens Prisma 3T scanner at the Caltech Brain Imaging Center with a 32-channel radio frequency coil. At the end of the first day of scanning, T1 and T2 weighted anatomical high-resolution scans were collected with 0.9 mm isotropic resolution.

High-resolution data

A high-resolution partial volume slab was collected in three participants with a 1.5 mm isotropic voxel size and the following parameters: multiband acceleration = 4, 64 slices, TR = 1,100 ms, TE = 26 ms, flip angle = 63°, FOV = 192 mm × 192 mm, in-plane GRAPPA (R = 2), echo spacing = 0.68 ms. The protocol is optimized to view medial prefrontal cortex (mPFC) in high-resolution and therefore the partial volume slab cuts off portions of the motor cortex and parietal lobe. Echo-planar imaging (EPI)-based fieldmaps of positive and negative polarity were also collected before each run with similar parameters as the sequence. We collected high-resolution scans in three of the participants in order to maximize our sensitivity to topographical encoding of single versus/bundle items within vmPFC. Preliminary analysis did not find evidence in this region, so we instead collected data from the remaining cohort using a standard resolution sequence as described below. This allowed us greater coverage within the brain to search for differential encoding of single and bundle items (at the cost of having to do so at a lower resolution). We included the high-resolution participants in subsequent analyses where appropriate, and observed similar neural effects across both acquisition protocols.

Standard resolution data

A whole-brain multiband EPI protocol was collected in 11 participants with a 2.5 mm isotropic voxel size and the following parameters: multiband acceleration = 4, 72 slices, TR = 1,120 ms, TE = 30 ms, A-P phase encoding, −30 degrees slice orientation from AC-PC line, flip angle = 54°, FOV = 192 mm × 192 mm. EPI-based fieldmaps of positive and negative polarity were also collected before each run with similar parameters as the sequence.

fMRI preprocessing

Data was preprocessed using a standard pipeline for preprocessing of multiband data. Using FMRIB software library (FSL; Smith et al., 2004), images were brain extracted, realigned, high-pass filtered (100 s threshold), and unwarped. Images were denoised by independent component analysis component removal. Components were extracted using FSL’s melodic, classified into signal or noise with a classifier trained on separate datasets or manually classified for the high-resolution dataset. T2 images were aligned to T1 images with FSL FLIRT, and then both were normalized to standard space using Advanced Normalization Tools (ANTs) using CIT168 high-resolution T1 and T2 templates (Avants et al., 2009; Tyszka and Pauli, 2016). The functional data was first co-registered to anatomical images using FSL’s FLIRT, then registered to the normalized T2 using ANTs. For univariate analyses, data was spatially smoothed in FSL with a 5-mm FWHM Gaussian kernel. For multivariate analyses, data was spatially smoothed with a 2-mm FWHM Gaussian kernel.

Predictions of normalization

Let the vector $\mathbf{v}=[v_i,v_j,\dots ]\in {\mathbb{R}}_{+}^N$ denote the values of individual items i, j, … that make up a bundle. We derive two formal results about the normalization function $Z(\mathbf{v}):{\mathbb{R}}_{+}^N\to {\mathbb{R}}_{+}^N$:

$$Z(\mathbf{v})=\frac{\mathbf{v}}{\sigma +\sum _{n\in N}{v}_n},(1)$$

with contrast parameter σ > 0, which implements a linear constraint on neural activity (Webb et al., 2020). For simplicity we will derive these results for two-item bundles, but the argument extends easily to N-item bundles.

Let v_i,j denote the value of the bundle of two items i and j. Theories of linear value integration would imply that v_i,j = v_i + v_j. Our first results demonstrates that normalization implies that bundle value is concave in item valuations, as detailed in the following proposition. A graphical demonstration can be found in Figure 2.

Figure 2.

Normalization. A normalization model predicts that the bundle value should be a concave function of item values if the reported bid is the sum of the normalized values.

Proposition 1

Concavity of Normalized Bundle Valuations

If value is constructed via a normalization function Equation 1, and the valuation of a bundle is the sum of the normalized valuations, $v_{i,j}=z_i+z_j=\frac{v_i+v_j}{\sigma +v_i+v_j}$, this sum is concave in the sum of the item values for all σ > 0.

Proof.

Let $\overline{v}\equiv v_i+v_j$. Therefore the bundle valuations can be written as a function of this sum, $v_{i,j}=\frac{\overline{v}}{\sigma +\overline{v}}$. The second derivative of this function with respect to that sum, is $\frac{{\mathrm{\partial }}^2{v}_{i,j}}{\mathrm{\partial }{\overline{v}}^2}=\frac{-2\sigma }{(\sigma +\overline{v}{)}^3}<0$ when σ > 0. Therefore it is concave.

Our second result concerns the responsiveness of the value normalization function, with respect to one or all of its arguments. This responsiveness determines the scale of the value function and will therefore be reflected in the magnitude of its first derivative. If the measurement of neural activity is linear in value, then this derivative will determine how neural activity responds to a change in value. We show here that for a normalization given by Equation 1, and a linear measure of neural activity, the neural measure will attenuate for bundles. Specifically, the first derivative of the value function will be largest for single items and will attenuate for bundles. Intuitively, the responsiveness of the normalization function attenuates because there is an additional item that must be normalized against (both the numerator and the denominator increase by the same amount).

Proposition 2

Attenuation of Normalized Value for Bundles

The first derivative of the normalization function Equation 1 for bundles is positive and strictly decreasing in v₂.

Proof.

For simplicity, we consider the derivative of Z([v₁, v₂]) with respect to an arbitrary argument v₂:

$$\frac{\mathrm{\partial }Z([v_1,v_2])}{\mathrm{\partial }{v}_2}=\frac{\sigma }{(\sigma +v_1+v_2{)}^2}.(2)$$

When v₂ = 0, and the normalization function is applied to only a single item, the first derivative of the function (i.e., its responsiveness) will be largest. It will therefore be attenuated for bundles (v₂ > 0). If the measurement of neural activity of a bundle is linear in these normalized values, $z_1+z_2=\frac{v_1+v_2}{\sigma +v_1+v_2}$, then its derivative is the sum of Equation 2 for both items and the responsiveness of this measurement to value would be similarly attenuated on bundle trials.

Behavioral analyses

Linear and non-linear regression analyses were performed to model how bundle value is computed as a function of the values of the constituent items in the bundle, v_ij = f(v_i, v_j). The linear model represents bundle value as a linear combination of the individual item values:

$$v_{i,j}={\beta }_0+{\beta }_1{v}_i+{\beta }_2{v}_j.$$

Item i and item j simply correspond to the item shown on the left and right, respectively, during stimulus presentation. Since the effect of screen location should not matter, this model can be further reduced to v_i,j = β₀ + β_sum(v_i + v_j) to improve efficiency. The mixed-effects model was estimated across all subjects, with subject-specific random effects terms for intercept and slope. Additional non-linear mixed-effects models were also constructed.

Power:

$$v_{i,j}={\beta }_0+{\beta }_1*v_i^{{\beta }_2}+{\beta }_3*v_j^{{\beta }_4}.(3)$$

Logarithmic:

$$v_{i,j}={\beta }_0+{\beta }_1*\log ({\beta }_3+v_i)+{\beta }_4*\log ({\beta }_5+v_j).(4)$$

Normalization:

$$v_{i,j}={\beta }_0+{\beta }_1\frac{v_i+v_j}{\sigma +v_i+v_j}.(5)$$

All models and model statistics were estimated with Matlab, with the fitlme and nlmefit functions, and model fits were evaluated with BIC and R². Reported BIC scores reflect the sum of BIC across participants. To plot the fitted line/curve for each model in Figure 3b, bundle value was computed from the fitted function parameters with v_i set equal to v_j for all values from 0 to 10 (with a 0.01 step size). The sum of these values (double of the value set both for v₁ and v₂) is represented by the x-axis. A mixed-effects linear model also used to analyze bundle type (food, trinket, mixed, and duplicates of the same item). Due to limited data, a random effect (across subjects) was only used for the intercept and with fixed effects controlling for type.

Figure 3.

Willingness-to-pay (WTP) behavior. a, Histograms of WTP bids for individual items and bundles across all subjects. b, Density plot of the WTP (value) of a bundle versus the sum of the values of the constituent items in a bundle (across all subjects). The diagonal represents where bundle values equal the sum of the constituent item values. The model fits from four models are plotted over the density plot. All models conclude that bundle value is a sub-additive function of the individual item values, as they extend below the diagonal. c, Linear model fits stratified by bundle type.

Regions of interest

Regions of interest (ROIs) were defined using the AAL database (Tzourio-Mazoyer et al., 2002). The labels used in the paper are mapped to the original ROI names as follows: ACC $\mathrm{\_}$pre: rACC, ACC $\mathrm{\_}$sup: dACC, Frontal $\mathrm{\_}$Inf $\mathrm{\_}$Orb $\mathrm{\_}$2: vlPFC, Frontal $\mathrm{\_}$Med $\mathrm{\_}$Orb: vmPFC,OFCant: OFCant, OFClat: OFClat, OFCmed: OFCmed, OFCpost: OFCpost, Frontal $\mathrm{\_}$Sup $\mathrm{\_}$Medial: dmPFC, Frontal $\mathrm{\_}$Sup $\mathrm{\_}$2: dlPFC, Frontal $\mathrm{\_}$Mid $\mathrm{\_}$2: MFG, Frontal $\mathrm{\_}$Inf $\mathrm{\_}$Tri: IFG.

Univariate analyses

Univariate analyses were conducted in SPM12. General linear models (GLMs) were constructed to examine how subjective value and choice are encoded in the brain during the choice task. This GLM included a regressor for value, time-locked to the onset of the trial as a parametric modulator, which was modified in separate GLMs to test hypotheses about value representation. The stimulus onset regressor additionally had another parametric modulator for trial type (−1 for single item trials and 1 for bundle trials). Thus, a contrast corresponding to the representation of bundle value or single item value could be computed as the interaction between value and trial type. Reaction times were additionally included as a third parametric modulator. Regressor for the choice made on a trial was time-locked to the time of choice, with separate regressors for choosing an item/bundle and choosing the reference monetary amount. Regressors of no interest were included: left and right button presses (duration = 0), motion regressors, and run. Missed trials were not modeled. Data from all three experimental days for each subject were consolidated into a single model, with separate sessions and distinct nuisance regressors for each day. The three participants that underwent the high-resolution scanning protocol are absent from the univariate analysis because that protocol was focused on ventral and frontal areas (with a limited field of view) with no coverage in dorsal frontal areas.

Different representations of value were included in the value regressor. The value contrast in Figure 5b, c used a normalized version of value, where value was z-scored by trial type (item or bundle). The absolute value model uses the raw WTP $ amount that the subject rated an item/bundle outside the scanner. WTP minus reference (value difference) used WTP minus the reference monetary amount on that trial.

Figure 5.

Normalization of the value code. a, The distributions of value by condition from example absolute and relative value codes. An absolute value code (left) represents value according to the participants’ WTP bids, an incentive-compatible measure of subjective value. A relative value code (right) normalizes value by condition, plotted here after z-scoring individual item values and bundle values separately for an example subject. A relative code adapts to the current context and places values from different distributions on the same scale. b, Contrast of normalized value > absolute value. A cluster in vmPFC is better fit by a relative value model, indicating that the value representation in this region is normalized. Clusters defined by a p < 0.001 uncorrected cluster forming threshold followed by p < 0.001 FDR cluster-level correction (N = 14). c, Contrast of normalized value > value difference (WTP bid $ amount—the reference $ amount), with a cluster in vmPFC defined similarly (N = 14). d, Model comparison (change in BIC) of different representations of value compared to the null model (with no value component but including nuisance controls). Models are fit via RSA. More negative values imply a better model fit, implicating the relative value model as the best account for the neural code in each region (N = 14). e, RSA Value coefficients from the Relative model, by subject. Value coefficients are attenuated for bundle trials relative to item trials, as predicted by normalization models. Horizontal lines reflect the average coefficient ± the standard error across all subjects and ROIs.

Multivariate analyses (MVPA)

To examine the nature of the bundle value code as the fine-grained distributed level, we implemented MVPA and RSA analyses. Analyses used the PyMVPA toolbox (Hanke et al., 2009), Scikit-learn functions, and custom Python code.

MVPA samples

To prepare for MVPA and RSA, we extracted trial-by-trial voxel-wise fMRI responses. A GLM was designed for each participant that modeled the onset and duration of each trial separately to extract the voxel responses that were unique to each trial. Other regressors of no interest modeled the other events and were not separated by trial (one regressor across an entire day): outcome phase (onset and duration), left and right button presses (duration = 0). As in the univariate analyses, data across all 3 d in a subject were included in the same model, with different days and separate regressors per day entered as different sessions. After estimating the models, the parameter estimate maps (beta maps) for each trial were concatenated into a 4D dataset, with length equal to the number of trials a subject completed across the experiment.

Cross-decoding analysis

To test for distinct bundle value codes, single item value codes, and general value regions, we implemented a cross-decoding analysis similar to previously used methods (McNamee et al., 2013) at the ROI level. The 4D dataset of beta maps for every trial were loaded with PyMVPA functions. Value (the WTP bid) was z-scored by trial type (item or bundle) and included as the targets to predict in the PyMVPA dataset. Then the voxels for each ROI were entered as features for the cross-decoding procedure. Ridge regression was used as the decoder in all analyses (Scikit-learn’s, linear $\mathrm{\_}$model.Ridge), with alpha =10³ (this parameter was optimized with sweeping). Each run was used as a separate cross-validation fold. Decoders were trained on the training samples from one trial type and tested on samples from the held out run (leave-one-run-out cross-validation). Decoders were tested on the samples from both trial types in the test set, even though they were trained on one type. However, decoder predictions were quantified separately for individual item samples in the test set and bundle samples in the test set. This ensured that performance could be compared across trial type separately from within trial type. Prediction accuracy was quantified by the Pearson correlation between the predictions and true value labels and averaged across cross-validation folds. This resulted in four decoder prediction accuracy metrics for each ROI and subject according to the four train/test splits: “train and test on single items”, “train and test on bundles”, “train on single items/test on bundles”, “train on bundles/test on single items.” The average prediction accuracy across participants for each of these train/test splits is plotted in Figure 4c. Significance was assessed for each split versus chance level (r = 0) with two-sided one-sample Wilcoxon signed rank tests at p < 0.05 and FDR-corrected for multiple comparisons (ROIs) at q = 0.05. To assess if a decoder’s prediction accuracy was significantly different between conditions in the test set, we used two-sided two-sample Wilcoxon signed rank tests at p < 0.05 and FDR-corrected for multiple comparisons (ROIs) at q = 0.05.

Figure 4.

Neural representation of subjective value and choice. a, Areas more active when the item or bundle was chosen over the reference monetary amount. Significant clusters in dmPFC/SFG, ACC, vmPFC, and angular gyrus, defined by a p < 0.001 uncorrected cluster forming threshold followed by p < 0.001 FDR cluster-level correction (N = 11). b, Neural correlates of item/bundle value. Significant cluster in anterior vmPFC (N = 14), defined similarly to previous pane. c, MVPA cross-decoding analysis results (N = 14). Ridge regression decoders were trained on samples from a condition and tested on held out samples from either condition (left: trained on single items, right: trained on bundle trials, top row: group result, and bottom rows: the first and the last subject). Asterisk * denotes significant prediction accuracy on a test partition corrected for multiple comparisons. At the group level, there were no significant paired differences in prediction accuracy between test conditions for any ROI. Error bars reflect SE across participants.

Representational similarity analysis

To examine whether the representational geometry of the regions of interest correlated to an absolute value or relative value code, we conducted RSA. The 4D dataset of beta maps for every trial were loaded with PyMVPA functions as in the cross-decoding analysis. Trial-by-trial neural dissimilarity matrices (DSMs) were constructed for every ROI by computing pairwise comparisons of the beta map across trials with PyMVPA’s PDist function. Euclidean distance was used as the distance metric. For all DSMs, within-day comparisons were removed to avoid potential confounds due to similarity being driven by patterns being in the same run or day. These neural DSMs were then compared against several theoretical model DSMs representing different hypotheses about value encoding:

1. Null model: Contains only nuisance regressors that account for potential confounds, including left versus right button presses, and item or bundle trial, to account for the large visual differences between the trial types. This model has no parameters.

2. Absolute value model: Value is represented according to raw WTP bids, regardless of trial type (individual item or bundle). This model has one parameter.

3. Subtractive normalization model: Value representation is normalized by subtracting the mean value within each stimulus category, adjusting for category-specific baselines. This model has one parameter.

4. Z-score normalization model: Value is normalized within each trial type using z-scoring (subtracting the mean and dividing by the standard deviation), allowing for direct comparison between individual items and bundles despite their different value distributions. This model has one parameter.

5. Relative model: Value is represented according to WTP bids, but the value coefficient can freely vary between trial types. A smaller coefficient in bundles is consistent with both range and divisive normalization. This model has two parameters.

The model implementation computed pairwise distances between these normalized values to create a model DSM. This was then combined with nuisance regressors through the formula:

$$\begin{array}{rl}\text{Predicted DSM}& =a_0+a_1\cdot {\text{ButtonPress}}_{\text{DSM}}\\ & +a_2\cdot {\text{TrialType}}_{\text{DSM}}+{\text{Value}}_{\text{DSM}}.\end{array}(6)$$

Where a₀ through a₂ are fitted parameters that control the contribution of each regressor to the predicted neural representation.

All models were fit to the neural DSMs using non-linear least squares optimization implemented in the LMfit Python package. To avoid potential confounds from within-day temporal autocorrelations, comparisons between trials from different scanning days were prioritized in the analysis. The goodness-of-fit for each model was assessed using the Bayesian Information Criterion to assess the accuracy between the model predictions and the empirical neural DSMs while correcting for the number of parameters in the model.

Results

Bundle value is a sub-additive function of the individual item values

We first examined how the bidding behavior for bundle valuations related to the valuations of the constituent items. Each participant first reported how much they would be WTP for various foods and non-comestible consumer goods (items) and pairs of these items (bundles; Fig. 1). These WTPs are collected using an incentive-compatible procedure validated in previous studies, and they therefore provide a measure of an individual’s subjective valuation of each item or bundle (Becker et al., 1964; Chib et al., 2009; McNamee et al., 2013; Suzuki et al., 2017). Participants were allocated a budget of $0–$20 and submitted a wide distribution of WTP bids across categories, albeit with most items and bundles valued at small dollar amounts (Fig. 3a; mean WTP value for each category: individual item food = $3.34 ± 3.30 s.d., individual item trinket = $3.30 ± 3.50 s.d., food bundle = $5.96 ± 4.44 s.d., trinket bundle = $4.92 ± 4.81 s.d., and mixed bundle $5.394 ± 4.65 s.d.).

In Figure 3, we plot the density of bundle valuations relative to the sum of the individual item valuations. While the bundle values tend to increase monotonically with the sum of the individual item values, as might be expected, there is substantial density below the diagonal suggesting that bundle value is a sub-additive function of the individual item values. To test this hypothesis, we modeled the value for a bundle of items, v_i,j, as a function of the constituent item values, v_i and v_j. The theory of attribute integration specifies the value of a bundle as a linear combination of values of the individual items (v_i,j = β₀ + β₁ v_i + β₂ v_j). Using a linear mixed-effects regression, we found that both parameter estimates for the individual items were significantly <1 ( ${\hat{\beta }}_1=0.73$, p < 10⁻¹⁰, t = 6.24; ${\hat{\beta }}_2=0.73$, p < 10⁻¹¹, t = 6.72; ${\hat{\beta }}_0=0.83$; R²= 0.78), consistent with the hypothesis that bundle value is a sub-additive function of the individual item values (Fig. 3b). We additionally tested three non-linear mixed-effects models which allow for concave functions of the combined items (a power model, a logarithmic model, and a divisive normalization model, see Methods). The best-performing model was a divisive normalization model which normalizes the bundle components relative to each other (ΔBIC = 182), see Table 1. The power model and logarithmic model, which allow for parametric non-linear functions on the valuations, also fit the data better than the linear model (ΔBIC = 129, ΔBIC = 65). The parameters of all of these non-linear models are estimated to yield concave value representations (Fig. 3b), thus participants discount their subjective value of a bundle more as the individual item values increase.

Table 1.

Behavioral models of bundle value

Model	BIC
Normalization	38,996
Power	39,049
Logarithmic	39,115
Linear	39,178

BIC scores reflect fit across participants. Each model fits random effects parameters per participant.

Although our experiment was not explicitly designed to measure substitutability in bundle evaluation, the mixture of categories in our experiment does enable a coarse analysis. If substitutability (together with diminishing marginal utility) were the predominant factor in the observed sub-additive bundle valuations, then we would expect substitution to be largest for bundles of the same item, smaller for bundles that are within-category, and smaller yet for bundles that are mixed across category. Therefore the magnitude of the trinket-mixed and food-mixed category comparisons is critical. Because both food and trinkets are more substitutable than mixed (the former are subsets of the latter), the bundle valuations should be more sub-additive in these categories.

However our behavioral results do not exhibit this pattern. Figure 3c depicts the results of a mixed-effects regression (with a random effect over subjects and fixed effects for each category) that uses the “mixed” category as a baseline. The slope is indeed smallest for bundles that are duplicates of the same item (β_sum = 0.59 ± 0.01, β₀ = 0.98 ± 0.30), which is consistent with the economic principle of diminishing marginal utility. The substitutability hypothesis would imply larger coefficients for food and trinket categories (i.e., closer to “1”) relative to mixed. However, we do not find that the food or trinket fixed effects are significantly larger than mixed (Food: β_sum = 0.65 ± 0.01, β₀ = 1.68 ± 0.29; Trinket: β_sum = 0.61 ± 0.09, β₀ = 1.02 ± 0.29; Mixed: β_sum = 0.66 ± 0.01, β₀ = 1.00 ± 0.28). In fact, the trinket fixed effect is significantly smaller (p < 3 × 10⁻⁷) compared to mixed, implying more subadditivity in this category, not less.

We also examined reaction times during the choice task to assess whether bundle and single item evaluations differ in processing speed (Supplementary Fig. 3). Bundle trials showed significantly faster reaction times compared to single item trials (bundle: M = 1.989 s, SD = 0.704; single item: M = 2.119 s, SD = 0.694). We also plot RT by choice difficulty, operationalized as WTP—the reference amount. Both trial types show similar patterns with longer reaction times for difficult choices near the reference value, consistent with an inverse-U relationship where choices closest to indifference require more deliberation.

Spatial representation of subjective value and choice in the brain

We next investigated the spatial representation of subjective value and choice in the brain during the choice task to determine whether different regions, or a hierarchy of regions, encoded single versus bundle trials. After the WTP task, participants were scanned using fMRI while they performed a choice task with the same items and bundles. Three participants were scanned with a high-resolution fMRI protocol (voxel size = 1.5 mm isotropic) designed to record from mPFC regions with high fidelity. The remaining 11 participants were scanned with a standard whole-brain protocol (2.5 mm isotropic). On each trial in the choice task, a participant made a choice between an item (or a bundle) versus a reference monetary amount equal to their median WTP bid on that category (Fig. 1b). This ensured that participants would value the item/bundle in isolation and choose the item/bundle about half the time. We observed a slight bias toward choosing the money on individual item trials (54.0% vs 46.0).

Univariate analysis

To examine whole-brain correlates of bundle and single item valuation, we first constructed a univariate GLM that included regressors for a stimulus WTP rating time-locked to the onset of the trial, and a regressor for the choice made on that trial time-locked to the choice (in addition to a regressor for trial type and other covariates of no interest; see Methods). Several regions were more active when participants chose an item or bundle versus when they chose the reference monetary amount, including clusters in dmPFC areas such as superior frontal gyrus (SFG), anterior cingulate gyrus (ACC), vmPFC, and the angular gyrus (p < 0.001 FDR-corrected cluster-level, Fig. 4a). Small subregions of ventral striatum were also significant p < 0.001, but these striatal activations did not survive the cluster-level FDR correction (p < 0.001). The three participants that underwent the high-resolution scanning protocol are absent from this specific whole-brain analysis because that protocol was focused on ventral and frontal areas (with a limited field of view) with no coverage in dorsal frontal areas. By including the regressor for the WTP of the item/bundle shown in a trial in the same GLM, we were also able to isolate subjective value signals independent of choice. After cluster-level false discovery rate correction (p < 0.001, FDR), one large cluster in the anterior portion of vmPFC and frontal pole showed a positive correlation to subjective value (Fig. 4b). These results are consistent with previous findings of a general value signal (Levy and Glimcher, 2012; Bartra et al., 2013; Clithero and Rangel, 2014) and set the stage for further examination of how the representations in these regions are modulated over single item and bundle trials.

We next investigated whether anatomically distinct brain regions represent the value of individual items and the value of bundles. One possibility is that additional regions are recruited when computing the value of bundles, since this involves a hierarchical process of valuing the individual items and then integrating them to evaluate the bundle. To examine whether the brain recruits additional circuitry when evaluating a bundle versus valuing a single item, we tested for the interaction of value and trial type in the previously described GLM (contrasts: bundle value > single item value and single item value > bundle value). At the group level, no clusters survived correction for either comparison. Therefore we do not find evidence for a topography of separable single item value and bundle value codes at the univariate level (Supplementary Fig. 1).

Multivariate analysis

We next aimed to determine whether the value of bundled items might be represented differently from single items at a more distributed level. We utilized multivariate pattern analysis (MVPA) in order to detect such distributed codes. Using all 14 subjects, we implemented decoding analyses across conditions to test for the existence of distinct bundle value codes, single item value codes, or general value regions. A ridge regression decoder was trained on distributed voxel patterns in several ROIs in PFC. The decoder was trained on samples of the valuations from one trial type (e.g., single item trials) and then tested separately on a held out run from either trial type (i.e., a leave-one-run-out cross-validation). Decoder performance was assessed with a Pearson correlation between the predicted values and the true values in the test set. To test the hypothesis of a condition-independent general value region, we analyzed if the decoder could predict the valuations of samples drawn from the other condition. This would identify general value regions that utilize the same distributed code when computing the value of both single items and bundles. To test for condition-dependent value regions, we analyzed if the decoder could only predict samples within the condition it was trained on, while failing to predict samples from the other condition. This would identify distinct single item value regions or bundle value regions.

We found that value could be predicted at the group level across all four train/test splits (“train and test on single items”, “train and test on bundles”, “train on single items/test on bundles”, and “train on bundles/test on single items”) in vlPFC, dmPFC, dlPFC, MFG, and IFG (Fig. 4c; two-sided one-sample Wilcoxon signed rank test, p < 0.05, FDR-corrected across ROIs at q = 0.05). Decoding across both item and bundle trials in these regions indicates the presence of distributed, condition-independent value codes. We also observed above-chance decoding for the “train on single items/test on bundles” split in rACC, dACC, vmPFC, anterior OFC, lOFC, and medial OFC, providing additional—though asymmetric—evidence for general value coding in these regions (as prediction accuracies did not reach significance for the reverse split).

In order to depict how value decoding varies across individuals, individual subject results are also shown for two representative subjects in Figure 4c, and the remaining subjects in Supplementary Figure 2. Six other PFC ROIs were significant for at least one split at the group level and individual participants exhibit interesting heterogeneity in which ROIs are representing value. Thus as depicted in Figure 4c, while some ROIs passed FDR correction and others did not at the group level, decoding strengths were broadly similar across regions, consistent with an individual-specific distributed value coding across PFC rather than sharp localized differences.

To identify a condition-dependent value region, we assessed if a region’s decoder only had significant prediction accuracy when trained and tested on a single condition (i.e., only significant for “train and test on single items” but not for the other three partitions). No region met this criteria at the group level, however the individual subject results show interesting heterogeneity. Additionally, there were no significant differences in a decoder’s prediction accuracy between conditions for any ROI (two-sided two-sample Wilcoxon signed rank test p < 0.05 and FDR-corrected for multiple comparisons q = 0.05). For example, although there is a 0.06 difference in Pearson correlation between “train on bundles/test on single items” and “train and test on bundles” for dACC, this difference is not significant after correcting for multiple comparisons (p < 0.426). Thus, the cross-decoding analyses do not yield any evidence for the existence of a distributed item and/or bundle-specific value code in any ROI. Instead, these findings are consistent with a distributed general value code for both bundles and single items.

Normalization of the value code

If the brain’s value regions represent value in a manner that generalizes across individual items and bundles of those items, a subsequent question emerges: how does the value code adaptively normalize to trial type? The common currency hypothesis suggests that options are encoded with the same value scale so they can be compared (Padoa-Schioppa and Assad, 2008; Chib et al., 2009; Levy and Glimcher, 2011; Lin et al., 2012), thereby value regions may code for items and bundles on the same range that scales with their WTP ratings. However, given the biological constraints of neurons, in order to evaluate decisions on drastically different value scales with the same valuation system (e.g., decisions about which house to buy or which entree to buy), the brain must also adaptively normalize the value code to the distribution of values available in a decision-making context (Louie et al., 2015). Since the distribution of bundle values is skewed toward larger values compared to the distribution of single item values (Fig. 3a), the neural code may normalize to the context of these different trials. A value region could therefore encode a $4 rated item very differently than a $4 rated bundle if options are appraised based on their utility relative to the other options within a condition (Fig. 5a). Previous work has established relative value coding empirically in experiments where a different distribution of values is presented in a block design (Padoa-Schioppa, 2009; Louie et al., 2011; Khaw et al., 2017; Zimmermann et al., 2018), presumably because the value code adapts gradually over time to match the distribution in each block. But crucially in our experiment, bundle trials and single item trials are randomly intermixed within a block. Therefore any adaptation, if present, must be instantaneous rather than slowly learned.

To test whether the neural representation codes for value in absolute or relative terms, we constructed GLMs with different value regressors in all subjects. One GLM included a regressor for value according to the WTP bid in dollars for the item or bundle displayed in that trial, while in a separate GLM value was normalized (with a z-score) by trial type. Value contrasts were then compared at the second level (normalized value > absolute value and absolute value > normalized value). A significant cluster in vmPFC emerged for the normalized value > absolute value comparison (peak voxel: x = −2, y = 62, z = −7; t₁₃ = 5.75; p < 0.001 uncorrected cluster forming threshold followed by p < 0.001 FDR cluster-level correction), suggesting that the neural representation of value in vmPFC is normalized (Fig. 5b). No significant clusters emerged after correction for the opposite analysis (absolute value > normalized value). An alternative explanation for this result is that vmPFC is computing the difference in value between the item/bundle presented in a trial and the reference monetary amount it is choosing against. This subtractive normalization computation would also produce a relative value code, but is a simpler form of adaptation that does not place the single item value and bundle value distributions on the same scale. We therefore constructed another GLM with a modified value regressor equal to the WTP bid minus the reference monetary amount, and the resulting value contrast was tested against the fully normalized value contrast. Again, a significant cluster in vmPFC resulted from the normalized value > value difference contrast (peak voxel: x = 0, y = 52, z = −14; t₁₃ = 4.83; p < 0.001 uncorrected cluster forming threshold followed by p < 0.005 FDR cluster-level correction; Fig. 5c), and the opposite contrast yielded no significant clusters. These results demonstrate that—at least at the univariate level—vmPFC normalizes the value code by condition even though that code is spatially consistent across conditions.

To examine a relative value representations at the multivariate level, we performed RSA in all subjects. RSA provides a data-driven method to examine the structure of neural representations by constructing DSMs according to how the multidimensional voxel space changes from trial-to-trial or from condition-to-condition (Kriegeskorte et al., 2008). These data-driven DSMs can then be compared to model DSMs that encode how the features of the task, such as value, evolve from trial-to-trial. We built trial-by-trial DSMs for every ROI and correlated them to model DSMs for four potential value representations: absolute value, subtractive normalization (subtracting the mean value of a stimulus in a category), z-score normalization, and a relative normalization model which allows the value coefficient to differ over trial type (item and bundle). In the latter model, this value coefficient will reflect the responsiveness of the neural code to value. Both divisive normalization and range normalization predict that this responsiveness should attenuate on bundle trials relative to single item trials. For divisive normalization, the responsiveness will attenuate because an item is normalized relative to an additional item (see Methods). For range normalization, it is because the range of valuations increases. Therefore an attenuation of the value coefficient would be consistent with a relative normalization model, but inconsistent with an absolute value model.

We compared all of the above models to a null model which only contains nuisance regressors. As shown in Figure 5d, the relative normalization model of value provides a much better fit in three value ROIs at the group level, vmPFC, OFCmed, and dmPFC. Additionally, over all subjects, the average value coefficient is significantly lower in bundles than for single items in all three of these areas (t = −3.359, p < 0.001; Fig. 5e). This suggests that the neural geometry in these value regions is most reflective of a relative value code in which the responsiveness of neural activity is attenuated in bundle trials.

Discussion

In real-world decision-making, consumers often make choices between options that are each made up of multiple goods. However, it is unknown how the brain constructs bundle values of multiple items. To investigate this, we used a BDM auction procedure to elicit subjective values for food items, non-comestible consumer goods, and bundles of these items and scanned participants while they made choices with these items and bundles.

We found that bundle value is computed as a sub-additive function of individual item values, therefore bundles are systematically discounted in relationship to the sum of individual item values. Concave non-linear models capture this behavior, suggesting that participants discount subjective values of a bundle more as the values of constituent items increase. The best-performing behavioral model normalizes item values relative to their sum; however, other non-linear models also capture this behavior.

At the neural level, we tested for absolute or relative value signals, across single item and bundle choices, at univariate and multivariate levels. Previous research has identified a spatial topography in OFC where food value was represented in posterior mOFC and value codes for consumer goods were represented in anterior mOFC (McNamee et al., 2013). In the present study, we investigated whether a spatial topography would also exist in relation to the valuation of single items versus bundled items. One possibility is that bundle item value, being more complex in the sense of being composed of multiple items as features, would be represented in a distinct region of vmPFC compared to single items. However, we found no evidence for separate single item and bundle value regions, or for a topography of value complexity. An anterior portion of vmPFC exhibited subjective value across both individual item trials and bundle trials. This result is consistent with a large literature finding category-independent value signals in vmPFC (Bartra et al., 2013), yet evidence for this signal in bundles has been elusive (Zyuzin et al., 2022). Given our finding that BOLD signal responsiveness to value is attenuated in bundle trials, this may have required higher-powered analyses than used hitherto to observe. In a cross-decoding analysis, distributed value codes were revealed throughout PFC, including in dmPFC, vlPFC, and MFG for both bundles and single items, however we found that none of these regions differentially coded for one or the other. This suggests that prefrontal cortex represents a generalized value signal across a valuation hierarchy of items and bundles.

While we did not find evidence for spatial topography with regard to the value of single and bundled items, we did find that value is encoded in a relative as opposed to absolute manner. Specifically, a form of normalization of value is occurring in these areas, such that the value representation is attenuated on bundle relative to single item trials. Using RSA, we compared neural activity against different forms of normalized and non-normalized value signals. We found that a scale normalized value code provides the best account of the neural data compared to other forms of non-normalized value signals. This is consistent with both divisive and range normalization models that predict an attenuation of value for bundles. While we highlight divisive or range normalization as possible mechanisms for explaining this attenuation in bundle trials, our experimental design cannot distinguish between them at the behavioural or neural level. Future studies which alter the number and composition of bundles are needed to do so (Daviet and Webb, 2023). The form of relative coding may also depend on the experimental paradigm used. Studies of learning experiments in monkeys also find that bundle evaluation is not linear in its component values, while lesions of medial frontal cortex cause these evaluations to reduce violations of monotonicity (Papageorgiou et al., 2017). This may be further evidence that the form of relative coding in learning paradigms may differ from simple choice paradigms (Bavard and Palminteri, 2023).

Our findings have implications for the common currency hypothesis (Levy and Glimcher, 2012; McNamee et al., 2013; Clithero and Rangel, 2014). Partly consistent with a common currency, we found that value is encoded in a general way across both single items and bundles, within and across categories of goods. While this shows that value is represented by the same distributed voxel pattern, this code does normalize across contexts. Thus, our findings do not support a context-invariant common currency representation. Instead, the brain appears to be utilizing a value representation scaled by context.

Our behavioral results are compatible with the expectation most consumers have that bundles of multiple goods are usually discounted in comparison to the total price of the constituent products purchased separately (Nagle and Holden, 1987; Thorne, 2004). Value meals at fast food restaurants, snack variety packs, and season tickets for sports teams all represent discounted bundles in the real-world. However, the incentives of buyers and sellers in the marketplace are usually different than in our experiment. Sellers typically offer bundles at a discount to ensure that high-WTP customers pay a higher price, as in discounting an item when purchased in bulk (i.e., second degree price discrimination). Alternatively, this increases the probability of a buyer purchasing an additional good they would not have purchased otherwise. In our experiment, each item and bundle is bid on separately and a BDM auction trial is selected at random, so when evaluating a bundle in the experiment, the option to purchase them separately is not available. Therefore, our findings suggest that bundle discounting may be a more general phenomenon of multi-item valuation and not just of market equilibrium pricing strategies. In particular, we found that bundles of the same item were discounted slightly more than other bundles, consistent with the law of diminishing marginal utility (an additional unit of a good has a lower value). More broadly, our finding that valuations are normalized may account for the observation that people tend to “underbid” in auctions like the BDM design used here (Noussair et al., 2004; Berry et al., 2020; Rosato and Tymula, 2024). A model in which valuation is relative (i.e., multiplicative) can explain non-linear patterns of underbidding that we observe. For single items (or pairs of the same item), such underbidding is indistinguishable from diminishing marginal utility. For bundles, the relative coding of valuations could explain why bundle valuations are non-linearly scaled as a function of their value. Additionally, our finding that bundle values are systematically discounted relative to the sum of their components complements other behavioral accounts of subadditivity (Yadav, 1994; Bettman et al., 1998; Papageorgiou et al., 2017).

Microeconomic theory also accounts for the possibility that bundles of goods are substitutes or complements (Mas-Colell et al., 1995; Varian, 2014). For example, tea and coffee could be viewed as substitutes because both offer a warm caffeine boost. When a consumer does not buy both at a cafe, this can by represented as the bundled utility being less than the sum of independent utilities. Though our experiment was not designed to study substitution effects, one feature of our design suggests that substitutability is not the primary mechanism explaining the pattern of bundle valuations. If participants in evaluated a large majority of the stimuli as substitutes with diminishing marginal utility, then the degree of subadditivity should be larger for bundles of items from categories that are subsets of a larger set (since they are more likely to be substitutes). We did not find this pattern within the food and trinket categories relative to “mixed” bundles from both categories. Still, the question of the neurocomputational basis for substitutes and complements is an exciting one. Single neuron recordings in rhesus macaque OFC corresponds to indifference curves over bundles derived from behaviour—when a subject’s choices exhibit more substitution, this is reflected in OFC activity (Pastor-Bernier et al., 2019). Moreover, this activity exhibits a reduced increase in activity to successively larger reward amounts, analogous to diminishing marginal utility (Pastor-Bernier et al., 2021). However how the brain constructs the value of bundles containing substitutes or complements, and how this is constructed from their component attributes, is an interesting question for future brain research. In particular, a study design which directly manipulates the substitutability of items, rather than just randomizing them into bundles or loosely within/across categories, could examine how these neural representations might differ.

To conclude, we show that the human brain represents the value of bundles in the same brain areas that encode the value of single items. We found no evidence for a topographical arrangement of value representations within the vmPFC or elsewhere, when comparing single item and bundled item value signals, suggesting that these value signals are not hierarchically structured at least with respect to neuroanatomical organization. However, the responsiveness of brain activity within those brain regions is attenuated for bundles relative to single items. Our findings thus indicate that value representations for complex combinations of stimuli utilize similar neural representations, but these are scaled differently depending on their context, as opposed to utilizing an absolute value code.

Data and Materials Availability

Code is available at https://github.com/locross93/Bundle_Value.git. fMRI Data is available at https://osf.io/hxa9p/.