The structural images were then averaged together for displaying the functional data. Event-related BOLD responses were analyzed using a general linear model (GLM). Activity related to each trial was modeled with a stick function, representing the onset of the first image, convolved with the canonical hemodynamic response function. Serial correlations in the time series were accounted for using the autoregressive model [AR(1)]. A high-pass filter of 128 s was used. CHIR-99021 There was a covariate of interest for each confidence bin (i.e., 1–6) for each of the four trial types (i.e., scene/face different/same) for each of the eight

runs. Covariates of no interest were the six motion covariates for each run estimated during the realignment step of preprocessing. Contrast coefficients were weighted to account for different numbers of trial types in each run. Contrast images from first-level analyses were then entered into second-level analyses. 3DClustSim (Cox, 1996; http://afni.nimh.nih.gov/pub/dist/doc/program_help/3dClustSim.html) was used to determine the cluster correction for p < 0.05 across the whole brain (p < 0.001 and k = 86 voxels). We opted to define the hippocampal ROI functionally, rather than structurally, because it would give us more power to detect hippocampal involvement. Previous studies have consistently found

posterior, but not anterior, CH5424802 cost hippocampal involvement in scene perception (e.g., Lee and Rudebeck, 2010; Lee et al., 2008; Mundy et al., 2012), which is consistent with work in the rodent implicating the dorsal (septal) hippocampus in spatial context memory (Fanselow and Dong, 2010 and Moser and Moser, 1998). Accordingly, averaging over the entire anterior-posterior extent of the hippocampus could have reduced the power to detect hippocampal involvement. To save the functional clusters as ROIs, we used MarsBaR (M. Brett, J.-L. Anton, R. Valabregue, and J.-P. Poline, only International Conference on Functional Mapping of the Human Brain, 2002). Parameter estimates were extracted and averaged

for the voxels within the cluster for each confidence bin. Responses were collapsed across “same” and “different” trials because an insufficient number of misses prevented an examination of only “different” trials for confidence responses 1–3. For the analysis with “different” trials only, we restricted the analysis to response bins “4,” “5,” and “6” because there were adequate trial numbers in those bins for every participant (i.e., every participant met the criterion of at least 10 trials in each response bin; average number of trials were 20, 32, and 42 for “4,” “5,” and “6” response bins, respectively). Because only 1 participant met the criterion of more than 10 responses in each bin for the “1,” “2,” and “3” responses on “different” trials, we could not reliably extract parameter estimates for those responses (average number of trials were 6, 9, and 10 for “1,” “2,” and “3” responses, respectively). See Figure S1.