Neurocomputational mechanisms underlying the subjective value of information

Humans have a striking desire to actively seek new information, even when it is devoid of any instrumental utility. However, the mechanisms that drive individuals’ subjective preference for information remain unclear. Here, we used fMRI to examine the processing of subjective information value, by having participants decide how much effort they were willing to trade-off for non-instrumental information. We showed that choices were best described by a model that accounted for: (1) the variability in individuals’ estimates of uncertainty, (2) their desire to reduce that uncertainty, and (3) their subjective preference for positively valenced information. Model-based analyses revealed the anterior cingulate as a key node that encodes the subjective value of information across multiple stages of decision-making – including when information was prospectively valued, and when the outcome was definitively delivered. These findings emphasise the multidimensionality of information value, and reveal the neurocomputational mechanisms underlying the variability in individuals’ desire to physically pursue informative outcomes.


Reporting for specific materials, systems and methods
We require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material, system or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response. Volunteers responded to an advertisement for the study. Sample size was comparable to other fMRI studies done in this area.
Participants completed an effort-based decision-making task outside the scanner, and an in-scanner information-seeking task. Order was counterbalanced across participants. A researcher was present to familiarise participants to the task. Stimuli were displayed on an MRI-compatible monitor positioned at the head of the scanner bore, and participants viewed the monitor through a mirror mounted on a 32-channel head coil. Functional data were acquired with a T2*weighted gradient echo-planar imaging (EPI) sequence using interleaved slice acquisition (TR 2,200 ms; TE 30 ms; flip angle 90°; 38 contiguous slices with a slice thickness of 3.0 mm without an interslice gap; voxel size 3.0 mm3 on a base matrix of 64 64 pixels, oriented along the AC-PC line).
Whole brain FMRIPREP standard pre-processing pipeline -see paper for details Spatial normalisation to the ICBM 152 Nonlinear Asymmetrical template version 2009c was performed through nonlinear registration with the antsRegistration tool of ANTs v2.1.0, using brain-extracted versions of both T1w volume and template. Brain tissue segmentation of cerebrospinal fluid (CSF), white-matter (WM) and gray-matter (GM) was performed on the brain-extracted T1w using fast (FSL v5.0.9).

ICBM 152 Nonlinear Asymmetrical template version 2009c
Physiological noise regressors were extracted applying CompCor. Principal components were estimated for the two CompCor variants: temporal (tCompCor) and anatomical (aCompCor). A mask to exclude signal with cortical origin was obtained by eroding the brain mask, ensuring it only contained subcortical structures. Six tCompCor components were then calculated including only the top 5% variable voxels within that subcortical mask. For aCompCor, six components were calculated within the intersection of the subcortical mask and the union of CSF and WM masks calculated in T1w space, after their projection to the native space of each functional run. Frame-wise displacement was calculated for each functional run using the implementation of Nipype.
Data were analysed using SPM12 (Wellcome Department of Imaging Neuroscience, Institute of Neurology, London, United Kingdom; http://www.fil.ion.ucl.ac.uk/spm), implemented in MATLAB. Each participant's data were modelled using fixedeffects analyses. The effects of the experimental paradigm were estimated for each participant on a voxel-by-voxel basis using the principles of the general linear model (GLM). Predictor functions were formed by modelling the onsets of the events of interest with a stick (delta) function convolved with the canonical haemodynamic response function. Lowfrequency noise was removed with a 128 s high-pass filter. The GLM included three regressors of interest: the Scenario event, the Reveal event, and the Outcome event, each of which was associated with a parametric modulator (see below). Other regressors which were included, but not analysed, included the motor events (i.e., the Choice and Effort events), and the onsets of the catch trials and their outcomes. The six head motion parameters derived during realignment (three translations and three rotations) were incorporated as additional nuisance regressors.
The main focus of this model-based fMRI study was to determine the neurocomputational mechanisms underlying: (1) the subjective valuation of information, and (2) the reduction of uncertainty across individual participants. To address the first goal, we computed the subjective value of information (i.e., k_i!I + k_w!W) for the chosen option on every trial for every participant using the parameters from our best-fitting model. In addition, we computed the subjective value of effort (i.e., k_e!E) for every trial using the same model. We then entered these two SVs as orthogonalised, parametric modulators for the Scenario event-related regressor.
To address the second goal, we computed the subjective value of information when it was finally delivered at the Reveal or Outcome screens. As for the first goal, the subjective value of information was defined through the winning model as k_i!I + k_w!W, which represents the amount by which uncertainty was reduced (as defined by the Rényi entropy function with a participant-specific alpha parameter), added to the valence of information. These subjective values were then entered as parametric modulators for the Reveal and Outcome events separately. Regression coefficients were estimated at the subject level using the standard restricted minimum-likelihood estimation implemented in SPM12. Variance inflation factors for all of our regressors were < 4, indicating that multicollinearity between regressors was not an issue in our design. A GLM design matrix for a representative participant is provided in Supplementary Figure 5.