Twenty-seven undergraduates from the National University of Singapore were recruited for this within-subject study through advertisements on a campus website. From this original pool, two were removed from analysis due to excessive head-motion in the scanner, one was excluded based on near-chance performance in both states, and another was excluded on the basis of image problems, giving a final sample of N
23 (12 male; mean age
21.3 years, SD
1.4 years). All subjects were right-handed, had no history of chronic physical or psychiatric disorders, or long-term medication use. They had regular sleep schedules and slept between 6.5–8 hours a night based on self-report, and were not extreme morning chronotypes as assessed by a modified Horne-Ostberg Chronotype Questionnaire 
Upon entering the study, subjects visited the lab for a briefing to practice the experimental task and to collect an Actiwatch (Actiwatch, Philips Respironics, USA) that they were instructed to wear at all times until the conclusion of the experiment. Subjects were also issued sleep diaries on which they were to record the onset and offset of all sleep bouts. Sleep history was checked prior to each of the fMRI scanning sessions, and participants who did not comply with a regular sleep schedule (>6.5 hours of sleep/night; sleep time no later than 1:00 AM; wake time no later than 9:00 AM) were excluded.
At least five days after the briefing, subjects returned to the laboratory for the first of two experimental sessions. In the rested wakefulness (RW) condition, subjects reported to the lab at approximately 7:30 AM. After filling in a questionnaire to assess their subjective level of sleepiness (the Karolinska Sleepiness Scale), they underwent an fMRI scan during which they performed a task involving selective attention to two different classes of stimuli: faces and houses (see fMRI procedures below for detailed description). Anatomical scans were also acquired during this time. fMRI scanning in the RW state typically began at about 8:00 AM. In the sleep deprivation (SD) condition, subjects reported to the lab on the evening prior to their fMRI scan. Subjects' actigraphy records were used to confirm they had awakened at their regular time on that day, and had not taken any daytime naps. Subjects remained awake overnight in the laboratory under the constant supervision of a research assistant. They were permitted to engage in light recreational activities, but were not allowed to smoke or consume caffeine. Every hour, participants performed the Psychomotor Vigilance Test and rated their subjective sleepiness using the Karolinska Sleepiness Scale. In the SD condition, subjects underwent an fMRI scan as in the RW condition, but at 6:00 AM. The order of scanning sessions was counterbalanced across subjects (RW session first; N
12) to minimize potential order confounds. Sessions were separated by at least one week, so that subjects undergoing the SD session first had sufficient time to fully recover from the effects of sleep loss.
Permission to conduct this study was granted by the Singapore General Hospital IRB, and all subjects provided written informed consent prior to participation. Subjects were financially compensated for their time. The individual providing the example face in provided written informed consent for the publication of this image.
Schematic of the object-selective attention task.
Subjects were shown blocks consisting of 6 novel targets (grayscale images of three faces and three houses) and 30 scrambled images that were of approximately equivalent luminance as the target pictures (). Equal numbers of male and female faces bearing neutral expressions were presented. Target stimuli were randomly interleaved with the scrambled images such that the interval between two targets ranged between 10 s and 14 s (mean
12 s). The interstimulus interval for presentation varied randomly between 0.5 s and 3.5 s (mean
1.75 s), except after the appearance of a target, when it was held constant at 2 s. This was to allow subjects adequate time to respond before the next stimulus onset.
At the start of each block, an instruction screen lasting 2 s was presented to the subject, informing them to either attend to faces, attend to houses, or passively observe the stimuli. This was followed by a further 2 s delay before the first stimulus appeared. In each of the ‘attend’ conditions, subjects were instructed to respond to the target by pressing a button with the right hand. In the ‘observe’ condition, subjects simply viewed the stimuli without making any response (). Thus, in the “attend to face” blocks, attend face (AF) and ignore house (IH) events were generated, and in “attend to house” blocks, attend house (AH) and ignore face (IF) events were generated. Observe face and observe house (OF and OH) events were generated in the blocks where stimuli were passively observed. fMRI runs consisted of 4 blocks of fixation (20 s) interleaved with 3 task blocks (77 s). Subjects performed 6 runs in total (all possible permutations of the task blocks) during each scanning session.
Finally, at the end of the RW session, subjects were scanned while they viewed blocks of faces and houses; data from these scans served as functional localizers that allowed us to identify the fusiform face area (FFA) and parahippocampal place area (PPA) for each individual subject 
. Functional localizers consisted of eight stimulus blocks interleaved with nine fixation blocks, and lasted 6 minutes and 16 seconds each. Each stimulus block comprised either 18 faces or 18 houses, presented at the rate of 1 per second.
MR imaging was conducted using a 3T Siemens Tim Trio scanner (Siemens, Erlangen, Germany) fitted with a 12-channel head coil. Participants viewed stimuli through a set of MR-compatible LCD goggles (Resonance Technology, Los Angeles, USA) and responded using their right index finger via a MR-compatible button box. Performance was continually monitored by a research assistant who noted all lapses and eye closures (through use of an eye tracking device). Subjects were prompted to attend to the task through an intercom system when they failed to respond to two consecutive trials, or when epochs of eye closure exceeded 3 seconds. Functional images were collected using a gradient echo-planar imaging sequence (TR: 2000 ms; TE: 30 ms; flip angle: 90°; field-of-view: 192×192 mm; matrix size: 64×64). Twenty-eight 3-mm axial slices aligned to the intercommisural plane and covering the whole brain were acquired. Directly following the functional data collection, a high-resolution T1 coplanar image was acquired. Finally, a high-resolution 3D MPRAGE sequence was obtained so that anatomical images could be normalized into common stereotactic space.
Image Preprocessing and Analysis
MRI data were analyzed using Brain Voyager QX version 1.10.1 (Brain Innovation) and Matlab R13 (Mathworks). Functional images were aligned across scanning runs to the first image of the final run. Intrasession image alignment to correct for motion was performed using the first acquisition of the final functional run as the reference scan. Interslice timing differences within each functional acquisition were corrected using cubic spline interpolation. We performed Gaussian filtering in the spatial domain by applying an 8 mm FWHM smoothing kernel. Linear signal drift, and signals of lower than 3 cycles/functional run were removed. Finally, all images were registered to their respective individual 3D high-resolution T1 anatomical image, and normalized to Talairach space 
Functional imaging data were analyzed using a general linear model with 13 predictors in an event-related analysis. Twelve of these predictors were created with a 2×2×3 model using all combinations of state (RW/SD), stimulus type (house/face) and trial type (attend/observe/ignore). We modeled events by convolving a stick function with a double-gamma, canonical hemodynamic response. Only correct ‘attend’ responses were analyzed. A thirteenth predictor was created to model all lapses (non-responses within 2 s) in each state; these events were not subsequently analyzed any further. As we did not want to include periods of data that included frequent microsleeps, runs in which there were >50% of undetected targets were not entered into the model. We excluded 14 out of 288 runs (4.9%) from the analysis for this reason.
In order to identify cognitive control regions activated above threshold by selective attention to houses as well as faces, we computed the conjunction of two contrasts: attend house (AH) vs. baseline and attend face (AF) vs. baseline in the RW state. To control for Type I error, voxels were processed using an iterative cluster size thresholding procedure 
that considered the spatial smoothness of functional imaging data when generating activation maps based on a corrected cluster threshold (p
<.05). Subsequent to this, a voxel-level threshold of at least p
<.001 (uncorrected) for t
maps was applied.
To characterize state-related differences in control region activation during task performance, we compared activation within a 10×10×10 mm cube of voxels surrounding the peak voxels obtained from the conjunction analysis described above in addition to running an ANOVA-based analysis. The frontal and parietal regions selected from the conjunction analysis have previously been identified as important areas involved in selective attention 
. These ROIs were then interrogated to evaluate the relative magnitude of activation for attend, ignore and observe conditions across the two states. All secondary statistical tests were conducted using SPSS version 17.0 (SPSS Inc., Chicago, IL).
Analysis of object-selective attention within the ventral visual cortex was ROI-based. The PPA and FFA were defined by a separately conducted localizer scan performed for each individual as described previously. PPA ROIs comprised a 10×10×10 mm cube of voxels that surrounded the one voxel showing maximum difference in activation between house and face blocks. We focused our analysis on the PPA as it has been shown to yield more discriminating and spatially more consistent, selectivity data 
. Furthermore, because there was no hemispheric asymmetry of PPA activation, activation magnitude data for all conditions—AH (attend house), IH (ignore house) and OH (observe house)—were obtained from both the left and right PPA and averaged. Activation magnitude across trial type and state was evaluated using paired t-tests. We opted not to use analysis of variance (ANOVA) as we had specific a priori hypotheses, and because some of the comparisons in the 2-way ANOVA would not have been meaningful (e.g. AHRW
Psychophysiological interaction (PPI) analysis 
was performed by extracting the time series of activation from a 10 mm cubic region around the peak voxels identified by the conjunction of AH vs. baseline and AF vs. baseline contrasts within the left intraparietal sulcus (IPS; Talairach co-ordinates: −27, −58, 37) as well as the left inferior frontal gyrus/insula (Talairach co-ordinates: −36, 11, 4). We selected these regions due to their known involvement in biasing object-based attention, and for consistency with a companion study 
To carry out PPI analysis, we used a linear model with three predictors: the time course of activity in the seed ROI, a task predictor coding for activity within task blocks (AH vs. IH or AH vs. OH) and a PPI term. To construct the PPI term, the deconvolved time-course of the relevant seed region was multiplied with a vector containing the psychological variables of interest. This product was then re-convolved with a canonical hemodynamic response function 
. The coefficient of this third, interaction term, is the one of interest in PPI analyses. Statistical maps of functional connectivity for each state were computed by conducting two-tailed, one sample t-tests on parameter estimates of the PPI (RW and SD) thresholded at p
To evaluate the robustness of the findings, we compared PPI in the AH vs. IH as well as AH vs. OH contexts as both comparisons evaluate object-selective attention.