Eighteen patients with schizophrenia and 18 healthy controls gave written informed consent according to the guidelines of the Human Subjects Review Committee of the University of Toronto and completed the study. Patients and healthy controls were all right handed, matched based on age, gender, education, and general cognitive estimates (see Table 1 and Appendix for details). The patients who participated in the study were medicated with atypical antipsychotics (see Table 2 for medication details). fMRI data from five patients and five controls could not be used: six scans were lost due to unsuccessful reconstruction; one participant's structural scan indicated brain abnormalities, and three participants moved more than the allowed 3 mm.

The US immediately followed the offset of the CS+ circle in 50% of the trials. The visual cues used as CS+ and CS− appeared for 5 s and the appetitive US and a neutral stimulus (a star) for 800 ms. A fixation cross was presented between trials for a period of 9 s. Trials were presented in a pseudorandom order to facilitate learning, but also to not require participants to be in the scanner for lengthy time periods. In total, there were 80 randomized trials in each session: (i) 20 CS+ paired with the US, (ii) 20 CS+ alone, (iii) 20 CS− paired with a neutral stimulus, and (iv) 20 CS− alone. The E-prime software (Psychology Software Tools, Inc., Pittsburgh, PA, USA) controlled the stimulus presentation. Immediately after participants left the scanner, they also completed a post-learning subjective-state questionnaire, in which they were asked to identify which colored circle had been associated with reward, and rate the amount of “pleasantness” they attributed to each visual cue.

Magnetic resonance imaging scans were acquired with a GE Signa 1.5 T scanner (General Electric, Waukesha, WI, USA) equipped with a standard head coil. Five hundred nine volumes of 28 contiguous axial 4.4 mm thick slices covering the whole brain were acquired using a T2 -sensitive spiral in-out sequence (TR = 2240 ms; TE = 40 ms; flip angle 85°; matrix 64 × 64; FOV 200 mm × 200 mm). This sequence was chosen because it has reduced susceptibility dropout in frontal and medial temporal regions including the VS (Glover and Thomason, 2004). The first three volumes were discarded to allow for T1 equilibrium effects, and the data from the remaining 506 volumes were used in the analysis.

Motion correction was performed using SPM99 (Wellcome Department of Cognitive Neurology, London, UK; Friston, 2007). Additional artifacts were removed using MELODIC (Multivariate Exploratory Linear Optimized Decomposition into Independent Components), an implementation of probabilistic independent component analysis (ICA; Beckmann and Smith, 2004). Artifacts were identified following the guidelines outlined by Beckmann and colleagues. An average of 40% of components per participant were interpreted as artifactual, and removed. There were no differences between the two groups with respect to the number of ICs removed. ICA correction increased the reliability of the voxel saliences representing the condition differences as assessed by bootstrap estimation of standard errors (see below).

After ICA correction, the images were spatially normalized with SPM99 to a locally created spiral template in Montreal Neurological Institute (MNI) space. The computed transformations were applied to all functional images, interpolated to isotropic voxels of 4 mm × 4 mm × 4 mm. The resulting images were smoothed using an 8-mm full width half-maximum isotropic Gaussian kernel. Analysis of Functional Neuroimaging software (http://afni.nimh.nih.gov/, Cox, 1996) was used to display statistical maps on a standard EPI template, and anatomical labels for local maxima were obtained using the SPM5 MNI Anatomy Atlas (Friston, 2007).

The fMRI data was analyzed using partial least squares (PLS; McIntosh and Lobaugh, 2004) to capture the brain regions that maximally represented group differences in BOLD activity in response to CS+ vs. CS− conditions. PLS is similar to principal components analysis (PCA) or to canonical correlation, with the exception of one important feature: PLS solutions are constrained to the part of the covariance structure that is attributable to the experimental manipulations or that relates to a given dependent measure. As such, PLS is ideal for data sets where the dependent measures within a block are highly inter-correlated or not of full rank, such as in the case of neuroimaging data, because items within a block are not adjusted for these correlations as they are in the canonical correlation approach.

Partial least squares computes an optimal squares fit across the entire dataset in time and space simultaneously, which requires the data to be in matrix format. Every row of the matrix contains data for one subject in one condition. The rows are arranged such that subjects are nested within condition blocks for control and schizophrenia groups, respectively. With 2 groups, 13 subjects per group, and 2 conditions (i.e., CS+ and CS−), there were 13 × 2 × 2 rows in the matrix. The columns of the matrix contained BOLD signal intensity for each voxel at each timepoint.

One of the advantages of using PLS is that it makes no assumptions about the shape of the hemodynamic response functions (HRFs). The HRF for any given condition normally lasts for several scans; therefore, a “lag-window” is a short signal segment within a trial that represents the response of each voxel. In the present study, a lag-window of eight lags from stimulus onset was used, representing eight TRs or an interval of approximately 18 s.

The present application of PLS employed the within-task mean-centering approach. Here, trials within each experimental condition were averaged and expressed as a voxel-by-voxel deviation from the grand mean across the entire dataset. Singular value decomposition (SVD) was then applied to the mean-centered deviation matrix. Mathematically, SVD simply re-expresses this matrix as a set of orthogonal singular vectors or latent variables (LVs), the number of which is equivalent to the total number of tasks. The LVs are analogous to eigenvectors in PCA and account for the covariance of the original mean-centered matrix in decreasing order of magnitude. For each LV, the two vectors are linked by a singular value (equivalent to the square root of the eigenvalues). The singular value indicates the proportion of crossblock covariance (i.e., covariance between the two blocks of data: brain activity and experimental design) that is accounted for by each LV.

The two vectors reflect a symmetrical relationship between the components of the experimental design most related to the differing signals in the voxels on one hand, and the optimal, in the least squares sense, spatiotemporal pattern of voxel signals related to the identified experimental design components on the other. The numerical weights at each voxel and timepoint are called voxel saliences. Those saliences identify the collection of voxels that are most related to the condition differences expressed in the given LV. The task saliences, on the other hand, indicate the degree to which each condition is related to the identified pattern of source waveform differences expressed in the given LV.

Structural equation modeling (SEM) analysis is a multivariate technique based on a structural model which represents the hypothesized causal relations between several variables (see McIntosh and Gonzalez-Lima, 1994; Buchel and Friston, 1997; Bullmore et al., 2000 for methodological details). In the context of fMRI research, these variables are the measured BOLD signal change of y₁ to y_n brain regions, and the hypothetical causal relations are based on anatomical connections between the regions. The strength of each connection y_i–y_j is specified by a “path coefficient,” which indicates how the variance of y_j depends on the variance of y_i if all other influences on y_j are held constant.

In the present study, the SEM analysis employed covariances of activity between regions of interest computed across participants and for each condition to determine path coefficients. The models were compared statistically to test for condition-specific differences in effective connectivity.

Five brain regions of interest in the right hemisphere that showed maximal task and group differences in the PLS analysis were selected based on prior knowledge of anatomical connectivity and involvement in reinforcement learning (Delgado et al., 2000; Knutson et al., 2000, 2001; Kirsch et al., 2003; Zald et al., 2004; Zink et al., 2004; Seymour et al., 2007). Furthermore, tracer studies in primates have also shown that the VS is anatomically connected to the dorsal striatum (Haber and Knutson, 2010), the hippocampus, and the medial prefrontal and orbitofrontal cortices (Haber et al., 1995; Ferry et al., 2000; Friedman et al., 2002). As the SEM analysis depends on an a priori model of anatomical connectivity, we specified only a small number of well validated connections in the right hemisphere and not across both hemispheres to exclude assumptions of commissural anatomical connectivity. The peak of the brain region activation (i.e., the BOLD signal at the third TR) was used in the analysis, averaging a neighborhood of 20 voxels around the location of the region of interest.

The selected regions were the VS [MNI (x, y, z) = 8, 4, −4], caudate (16, −4, 16), putamen (28, 8, −8), hippocampus (20, −36 −12), orbitofrontal cortex (20, 16, −24), and the superior medial frontal gyrus (16, 56, 8). Brain region amplitudes were obtained by extracting the BOLD signal from the task PLS analysis across the entire event window (i.e., eight lags) for both unpaired CS+ and CS− conditions. The peak of BOLD activity (i.e., BOLD signal from the third lag or the third TR) was used in the SEM analysis. We decided to select the time window that showed that largest differences between the CS+ and CS− conditions, because we anticipated that the fMRI data would exhibit significant temporal correlation due to low-frequency physiological fluctuation. Furthermore, we chose the third lag because, in comparison to the rest, it also showed robust and distinct functional connectivity patterns in the CS+ and CS− conditions across the two groups.

Structural equation modeling was performed using the program Amos 17.0 (SmallWaters Corp., USA) applying a maximum likelihood algorithm for estimating path coefficients. Statistical inferences about group differences were based on a hierarchical model approach. This approach compared an alternative model, in which all connections were allowed to vary between the two groups and the two conditions, to a null model in which all path coefficients were constrained to be the same across all groups and all conditions. In the null model, error variances were also set to be the same as those estimated in the alternative model.

The χ² goodness-of-fit statistic was used to assess the model's ability to reproduce the original correlation matrix. The difference ( ) was examined with the degrees of freedom equal to the difference between the degrees of freedom in the constrained and the free models. The test is a hierarchical test that determines whether a modification to the model leads to significant improvement in the goodness-of-fit of the model (McIntosh and Gonzalez-Lima, 1994; Protzner and McIntosh, 2006). Comparison of the models was done by subtracting the goodness-of-fit χ² value for the null model () from the χ² value for the alternative model (). If the χ² value for the null model was significantly larger than that of the alternative model, then the path coefficients that varied between conditions and groups were statistically distinct (Protzner and McIntosh, 2006).

Based on the subjective-state questionnaire following the experiment, controls rated the CS+ as more pleasant than the CS− [1.92 vs. 0.69; z = 2.86; p < 0.004; for the CS+ vs. CS−, respectively]. Although patients reported awareness of stimulus associations following the experiment, they did not significantly distinguish between the CS+ and the CS− in terms of pleasantness [1.25 vs. 0.58; z = 1.72; p = 0.08; for the CS+ vs. CS−, respectively].

Galvanic skin conductance was analyzed for unpaired CS+ and CS− trial types only in which there was no subsequent outcome. The percent of trials in which GSR values significantly exceeded baseline values following the onset of either the CS+ or CS− stimuli was analyzed using repeated measures analysis of variance (ANOVA) with group assignment as the between-subject factor and the CS+ and CS− unpaired conditions as the within-subject factors. A main effect of stimulus type was detected with a larger percent of trials showing increased GSR values to the unpaired CS+ condition [F(1,21) = 16.82, p < 0.001]. As it can be seen from Figure 1, an interaction between condition and group assignment was also detected [F(1, 21) = 7.1, p < 0.015] with patients exhibiting less differentiation between the CS+ and CS− conditions, decreased GSR signal to the CS+ as compared to controls. The group by condition interaction was primarily driven by the larger percentage of trials exhibiting increased GSR values in the unpaired CS+ condition in controls compared to patients (i.e., an average of 23% in controls compared to 13.5% in patients).

To further examine the GSR patterns driving the group x condition interactions, we also compared CS+ to CS− conditions within each group; this analysis was also performed on unpaired trial types. Significant differences between stimulus types were observed in the control group [t(11) = 4.25, p < 0.003] with larger GSR values to the CS+ compared to the CS− condition. No significant differences between the unpaired CS+ and CS− conditions were observed in the schizophrenia group [t(9) = 1.65, p = 0.13].

The first LV, calculated using the mean-centered PLS analysis, captured group differences in the CS+ vs. CS− contrast (LV1 = 57.41% crossblock covariance, p < 0.012). Controls showed increased BOLD activity to the CS+ compared to the CS−, and patients reflected the opposite pattern, increased BOLD activity to the CS− compared to the CS+ condition. The task saliences along with the voxel saliences for LV1 are included in Figure 2. The brain scores in Figure 2 indicate how strongly individual subjects express the condition differences identified in LV1. The collection of voxels that reliably expressed the condition differences with bootstrap ratios ≥ 3 are listed in Table 3.

Group differences in BOLD activity between CS+ and CS− conditions were observed in several brain regions including striatum, the right VS, right caudate, bilateral putamen (Figure 3), and right hippocampus, right ACC (BA 25), bilateral medial PFC (BA 10), and right orbitofrontal cortex (BA 11; Figure 4). In contrast to the control group, patients exhibited increased BOLD activity to the CS− compared to CS+ conditions in the aforementioned regions (see Figures 3 and 4).

The omnibus (p < 0.0001) between the alternative and the null models indicated significant overall group by condition differences. Refer to Table 4 for the individual path coefficients across the two groups and two experimental conditions.

In comparison to controls, patients exhibited significantly larger connectivity strengths from the (i) VS to the caudate, (ii) VS to the putamen, (iii) VS to the hippocampus, and (iv) VS to the OFC in the CS− condition, and from the VS to the superior medial frontal gyrus in response to both CS+ and CS− conditions. While controls showed increased connectivity from the OFC to the putamen, patients exhibited the opposite connectivity pattern in the CS− condition, from the putamen to the OFC. Increased effective connectivity from the OFC to the hippocampus was observed in the CS+ condition across both control and patient groups.

Pairwise comparisons between the two groups within each condition indicated a significant difference of model fit between the two groups in the CS− condition with (p < 0.0001) and a difference of model fit in the CS+ condition with (p = 0.07) that did not exceed conventional statistical thresholds.

In the CS+ condition, path coefficients from the VS to the OFC (BA 11) were significantly lower in patients compared to healthy controls (, p < 0.02; Table 5; Figures 5A,B).

In the CS− condition, path coefficients from the VS to the OFC and from the VS to the hippocampus were significantly larger in patients compared to healthy controls (, p < 0.0001;, p < 0.0001). Larger connectivity strengths from the OFC to the putamen were observed in healthy controls and opposite connectivity strengths, from the putamen to the OFC, were detected in patients (, p < 0.01; Table 5; Figures 5C,D).