We also calculated a generalized r2-statistics for each model, which is a standardized measure of model fit analogous to accounted variance ( Nagelkerke, 1991). It is computed as r2=1−L/Lrandomr2=1−L/Lrandom. Stimuli were presented on a gray background using Cogent 2000 (http://www.vislab.ucl.ac.uk/cogent.php) running in MATLAB. Stimuli were presented using an LCD projector running at a refresh rate of 60 Hz, viewed by subjects via an adjustable mirror. Data were acquired with a 3T scanner (Trio, Siemens, Erlangen, Germany)

using a 12-channel phased array head coil. Functional images were taken with a gradient echo T2∗-weighted echo-planar sequence (TR = 3.128 s, flip angle = 90°, TE = 30 ms, 64 × 64 matrix). Whole brain coverage was achieved LY2835219 cost by taking 46 slices in ascending order (2 mm thickness, 1 mm gap, in-plane resolution 3 × 3 mm), tilted in an oblique

orientation at −30° to minimize BMS-354825 signal dropout in ventrolateral and medial frontal cortex (Weiskopf et al., 2006). Subjects’ head was restrained with foam pads to limit head movement during acquisition. Functional imaging data were acquired in three separate 415-volume runs, each lasting about 21 min. The first five volumes of each run were discarded to allow for T1 equilibration. A B0-fieldmap (double-echo FLASH, TE1 = 10 ms, TE2 = 12.46 ms, 3 × 3 × 2 mm resolution) and a high-resolution T1-weighted anatomical scan Oxalosuccinic acid of the whole brain (MDEFT sequence, 1 × 1 × 1 mm resolution) were also acquired for each subject. Image analysis was performed using SPM8 (rev. 3911; http://www.fil.ion.ucl.ac.uk/spm). EPI images were realigned and unwarped using field maps (Andersson et al., 2001). Each subject’s T1 image was segmented into gray matter, white matter, and cerebrospinal fluid, and the segmentation parameters were used

to warp the T1 image to the SPM Montreal Neurological Institute (MNI) template. These normalization parameters were then applied to the functional data. Finally, the normalized images were spatially smoothed using an isotropic 8-mm full-width half-maximum Gaussian kernel. FMRI time series were regressed onto a composite general linear model (GLM) containing regressors representing the time of the choice, the time of the outcome screen, and any button presses during the choice period. The outcome regressor was modulated by a number of model derived decision variables. Modulators for outcome were: prediction errors for the individual resource outcomes and the portfolio outcome (δ1, δ2, δp), the absolute deviation of the portfolio outcome from the target (|δp|), resource risk (h1, h2), risk prediction errors (ε1, ε2), the correlation strength of the resources (ρ), and the correlation prediction error (ζ). A further modulator captured the anticipated magnitude of actual weight updating in the next trial (|wt − wt+1|).