|
nif:isString
|
-
All participants gave written informed consent and were compensated for their participation. The study was approved by the Institutional Review Board of the National Institutes of Health, Bethesda, Maryland, USA.
30 Caucasian participants (11 male), living in the Washington D.C. area. One participant’s data was excluded from analyses of test phase eye-movements due to partial data corruption.
We used an EyeLink II headmounted eye-tracker (SR Research, Mississauga, ON, Canada), and sampled pupil centroid at 500 Hz. The default nine point calibration and validation sequences were repeated throughout the experiment. Both eyes were calibrated and validated, but only the eye with the lowest maximum error was recorded for the trials following a particular calibration. Calibration was repeated when maximum error at validation was more than 1° of visual angle. Before each trial, a drift correction was performed. Default criteria for fixations, blinks, and saccades as implemented in the Eyelink system were used.
We collected 32 Caucasian-American, 32 African-American, and 32 Chinese face images (16 male and 16 female for each race), for a total of 96 grayscale neutral expression frontal-view face images (see Fig 1A for examples). All Caucasian faces were taken from the neutral expression 18 to 29 age group of the Productive Aging Lab Face Database established by the University of Texas at Dallas (http://vitallongevity.utdallas.edu/stimuli/facedb/categories/neutral-faces.html) [35]. African-American faces were taken from the neutral expression 18 to 29 age group of the Productive Aging Lab Face Database, from the MacBrain (“NimStim”) Face Stimulus Set made by the MacArthur Foundation Research Network on Early Experience and Brain Development (http://www.macbrain.org/resources.htm), and from the Color FERET Database (http://www.nist.gov/itl/iad/ig/colorferet.cfm) [36,37] established by the United States Department of Defense (DOD) Counterdrug Technology Program. All Chinese faces were taken from the CAS-PEAL Face Database (http://www.jdl.ac.cn/peal/index.html) [38] established by the ICT-ISVISION Joint Research and Development Laboratory (JDL) for Face Recognition. The use of different face databases for different races of faces meant that not all aspects of the images were uniform, and this could spuriously impact results. Therefore in order to address this limitation, the images were modified to make several image properties alike across all face stimuli. Each face was scaled to have a forehead width subtending 10 degrees of visual angle at presentation and was rotated to correct for any tilt of the head. Images were cropped to remove most of the background, but not the hair or other external features, and all images were equated for overall luminance. At presentation, images were centered on a black background. To eliminate any possible stimulus bias as the source of any laterality effects, half of the faces were randomly left-right flipped across the vertical midline of the image for each participant.
Figure data removed from full text. Figure identifier and caption: 10.1371/journal.pone.0148253.g001 Study design. (a) Four example face stimuli. (b) AOIs for one face showing the calculation of start positions, which were determined separately for each face and defined relative to that face. Green dots schematically illustrate the potential start positions relative to the upcoming face. Left and right start positions were equidistant from centers of the nearest eye, nose and mouth AOIs. Upper and lower start positions were equidistant from the centers of the two eye or two mouth AOIs, respectively. Dotted blue lines schematically illustrate that the start positions were equidistant from the centers of the indicated AOIs. (c) Trial sequences in study and test phases. A face was only presented if the participant successfully maintained fixation for a total of 1.5 seconds. After face onset in the study phase, participants were free to study the face for up to 10 seconds and pressed a button to begin the next trial. In the test phase, faces were presented for one second only and participants responded with button presses to indicate whether the face was ‘old’ or ‘new’.
The website of the Productive Aging Lab Face Database states: “This [database] contains a range of face of all ages which are suitable for use as stimuli in face processing studies. Releases have been signed by the participants we photographed and the faces may be included in publications or in media events.” Development of the MacBrain Face Stimulus Set was overseen by Nim Tottenham and supported by the John D. and Catherine T. MacArthur Foundation Research Network on Early Experience and Brain Development. Please contact Nim Tottenham at tott0006@tc.umn.edu for more information concerning the stimulus set. Portions of the research in this paper use the FERET database of facial images collected under the FERET program, sponsored by the DOD Counterdrug Technology Development Program Office. The research in this paper use the CAS-PEAL-R1 face database collected under the sponsor of the Chinese National Hi-Tech Program and ISVISION Tech. Co. Ltd. For the purposes of analysis and for aspects of our experimental design, rectangular areas-of-interest (AOIs) were manually drawn prior to the experiment for each face around the right and left eyes, bridge of nose (i.e. middle of eye region), right and left half of nose, and right and left half of mouth (Fig 1B, for example) using EyeLink Data Viewer software. These AOIs were never visible to participants during the experiment. The edges of the AOIs were defined with designated criteria that were applied to all face stimuli as consistently as possible with respect to the spacing relative to relevant facial feature landmarks. Specifically, eye AOI edges were defined in the x-dimension as just beyond each canthus of the eye, with slightly more space between the lateral canthi and the respective edges of the AOIs than that for the medial canthi due to the nearness of the bridge of the nose to the medial canthi. In the y-dimension, the eye AOI edges were defined as from the lower section of the malar fold to just beyond the superior eyelid. The bridge of nose AOIs were defined in the x-dimension as the whole space between the two eye AOIs. In the y-dimension, the edges were identical to those of the eye AOIs. The nose AOI was defined in the x-dimension as just beyond the edges of the alar lobules. In the y-dimension, the lower edge of the nose AOI bisected the philthrum and the top edge was identical to the lower edges of the eye and bridge AOIs. The mouth AOI was defined in the x-dimension as just beyond the oral commissures. In the y-dimension, the lower edge was at a similar position relative to bottom edge of the lip that tended to land about the labiomedial crease, when it was visible. The upper edge of the mouth AOI was identical to the lower edge of the nose AOI. When nose and mouth AOIs were split into two halves (for the purpose of placing the pre-stimulus start positions), simple bisection in the x-dimension was performed. We subsequently conducted analyses of AOI areas, widths, and heights to provide rough indices of race of face physiognomic differences and variability (Figures F-H in S1 File).
We varied race of face stimulus (Caucasian, African, Chinese) and pre-stimulus fixation location (“start position”) across the trials of the experiment comprised of two phases: study and test. We systematically varied start position because fixation patterns are affected by visuo-motor factors (e.g. start position) in addition to stimulus factors (face) [33,34]. During the study phase, participants observed 48 faces (16 of each race, 8 male for each race) in a self-paced manner (up to 10 seconds, self-terminating trials with a button press). At test, participants observed 96 faces (the 48 study phase faces plus 48 new faces) for a limited duration (one second only) and indicated whether or not they recognized each face as one observed during study (old/new task) with a button press. Participants were instructed to respond within two seconds following stimulus onset, as soon as they thought they knew the answer (Fig 1C). The experiment was programmed in Python and interfaced with the eye-tracker. Start positions were either above, below, right of, or left of the internal features of the upcoming face stimulus (see Fig 1B for examples). Coordinates for a given start position were calculated uniquely for each face stimulus to be equidistant from all of the nearest internal facial features. Specifically, for right and left start positions, the unique coordinate that was equidistant from the centers of the nearest eye, nearest half-nose, and nearest half-mouth AOI was calculated numerically for each face. Upper start positions were equidistant from the center of the two eye AOIs, and the lower start positions were equidistant from the two half-mouth AOIs. Distances from the upper and lower start positions to their respective AOI centers were constrained to be the mean of the right and left start position distances from their respective AOI centers. Before stimulus onset, participants fixated at the start position, indicated by a standard Eyelink II calibration target (0.17° diameter black circle overlaid on a 0.75° diameter white circle) on the black screen. Participants initiated the trial by pressing a button while looking at the fixation target. In this period, a drift correction was performed. A colored dot (0.5° diameter) remained after drift correction, and the stimulus appeared only after a participant had fixated at the dot for an accumulated total of 1500 ms. This ensured that drift correction and fixation were stable prior to stimulus onset. If more than 1500 ms of fixation away from the start position accumulated before the trial could be initiated, drift correction was repeated. A fixation was considered off the start position if it landed more than 0.5° from the center of the dot. Dot color changed successively from red to yellow to green in order to signal to the participant that a maintained fixation was successfully detected at the start position. In both the study and test phases, there were equal proportions of trials for each combination of levels of the factors of face race, face gender, and start position. When a given face was presented in both the study and test phases, the face images were identical across study and test phases. This practice had the advantage of making analysis more straightforward and easily interpreted since changes in viewpoint, emotional expression, lighting, etc. would not serve as confounds for eye-movement differences; however, this practice also has the limitation of potentially allowing simple image matching mechanisms, in addition to the more abstract facial identification mechanisms of interest. The particular subset of faces used in the study phase was randomized across participants. Of the faces presented in both study and test phase, half of the faces were presented with the same start position at study and test and for the other half, the start position on the other side of the face was used (e.g. left to right start position between study and test; upper to lower between study and test).
Fixation and AOI data were obtained through EyeLink Data Viewer software by SR Research. Subsequent analyses on these data and behavioral data from the test phase were performed with custom Matlab (The MathWorks, Inc., Natick, MA, USA) code. ANOVAs were performed in SPSS (IBM, Somers, NY).
We assessed participants’ discrimination performance, response bias, and reaction time on the old/new recognition task in the test phase. d' (d’ = z(hit rate)—z(false alarm rate)) and criterion c (c = -[z(hit rate) + z(false alarm rate)]/2) were computed for discrimination performance for each participant, broken down by race of face and start position. Reaction times were analyzed for correct trials only. Reaction time values more extreme than 2.5 standard deviations from the median within each condition and participant were excluded from analysis. Reaction time analyses were broken down by Race of Face and Start Position conditions with analysis being performed on the medians calculated for each subject. Greenhouse-Geisser correction was applied if any of the factors or interactions of a given ANOVA violated sphericity.
We assessed the relative frequencies of fixations across the AOIs as a function of our experimental manipulations. The AOIs used were left eye, bridge, right eye, nose (left and right sides combined to be comparable with prior studies), mouth (left and right sides combined), and other (outside the defined AOI regions). Given the variable numbers of fixations across trials and across participants, only the first five fixations of each trial were included in the analyses of the study phase. Participants rarely made fewer than five fixations in study phase trials and, further, the first few fixations are likely to be the most essential for the task, as indicated in prior research [39]. For the test analysis phase, all fixations within the entire stimulus viewing time (limit of one second per trial) were included. Relative frequency was calculated for each AOI as the number of actual fixations divided by the total number of possible fixations across all trials of the given condition for each subject (e.g. 16 study phase trials with Chinese faces multiplied by 4 fixations per trial = 64 total possible fixations across all Chinese face study phase trials). ANOVAs on relative frequencies excluded the relative frequency value for the region outside of the AOIs. Greenhouse-Geisser correction was also applied if any of the factors or interactions of a given ANOVA violated sphericity We mapped the spatial density of fixations during the study phase as a function of our experimental manipulations. Each fixation was plotted with equal density and spatial extent, and fixations were not weighted by the fixation duration (essentially the same qualitative pattern of results was obtained when this weighting function was applied). Fixations beyond the fifth fixation were excluded from the analysis to ensure an equal amount of data across trials. The first fixation was also excluded (See Results for motivation). To ensure that summation of fixation maps across different face trials produced spatially meaningful density maps, fixation maps for individual faces were first aligned to a common reference frame using simple translations only. This reference frame was defined by the internal facial features. Specifically, the alignment minimized the sum of the squared differences between the center of the AOIs for each face and the average centers of the AOIs across all 96 faces. Within this common reference frame, fixations were then plotted as Gaussian densities with a mean of 0 and a standard deviation of 0.3° of visual angle in both the x and y dimensions. These density plots were then averaged across trials and across participants. A small proportion of analyzed fixations (< 2% during study, < 1% during test) fell outside of the bounds of the stimulus image region (i.e. onto the black background outside the square frame of the face stimulus). These fixations were not excluded from the analyses, but are simply not visible in plots. The resulting maps show the spatial fixation densities, using a color scale from zero to the maximum density value observed, with values approaching zero being deep blue. All maps within a single figure contain the same total number of fixations and so are scaled the same to allow for direct comparison.
Spatial Density Contrasts: Difference Maps: In order to view differences in the spatial fixation density between two conditions, a pixel-wise subtraction between two spatial density maps was performed for each participant and then averaged across participants. Spatial Density Contrasts: Statistical Maps: In order to produce maps of statistically significant differences in the spatial density map contrasts, a Monte Carlo permutation test was performed on fixation locations between the contrasted conditions. A Monte Carlo permutation test (also called an approximate permutation test or a random permutation test) is a standard, accurate and robust method of performing a significance test on data that is not known to have a parametric (e.g. normal) distribution of values, such as our data. We have used this type of statistical analysis method on eye-tracking data in a previous study [34], based on methods applied to the analysis of functional brain imaging data [40] and similar to that used in a prior study of eye tracking [41]. The null hypothesis in the Monte Carlo permutation tests was that the distributions of fixation locations of each ordinal fixation (i.e. fixation 2, fixation 3 etc.) were the same between the contrasted conditions (e.g. fixation 2 in Caucasian versus Chinese trials, or fixation 3 in right start position versus left). Thus, exchangeability of fixation locations between the given contrasted conditions was assumed only for fixations of the same ordinal value in the sequence of five fixations per trial. Only the first five fixations were analyzed for the same reasons that only the first five fixations were analyzed in the AOI analyses. 104,000 resampling iterations were performed for each statistical map. For each iteration, locations of fixations were resampled for each individual participant according to the assumed exchangeability, then a new resampled spatial density contrast was produced. These resampled maps were then averaged across participants to produce 104,000 group difference maps, the distribution of which was used to determine significance. Maps of p-values were computed pixel-wise based on the number of corresponding pixels in the resampling iterations that were greater than a given positively valued pixel (i.e. where condition 1 had a greater density) in the true spatial density contrast and that were less than a given negatively valued pixel (i.e. condition 2 greater) in the true spatial density contrast. The maps were thresholded at a pixel significance of p < 0.01 (equivalent to two-tailed p < 0.02). For eye-tracking data, our statistical analysis has advantages over other methods of performing significance tests on contrasted fixation maps. Statistical analysis upon contrasted fixation maps can present particular problems to which AOI analyses are much less susceptible due to the high degree of pooling of fixation data in AOI analysis and due to the much more limited number of statistical tests involved in AOI analysis. Specifically, a pixel-wise t-test on fixation maps is inappropriate because the within-subject differences in fixation density data often deviate extremely from a normal distribution at many pixels of a heatmap. Even pixel-wise non-parametric tests could create a large multiple comparisons problem, which grows as the pixel resolution of heatmaps grow. In our analysis, fixation locations are exchanged rather than pixels; therefore, increasing the resolution at which heatmaps are displayed does not exacerbate the multiple comparisons problem. Our analysis is an alternative to a another approach, which has been implemented by Caldara and colleagues in a free Matlab toolbox called iMap [42].
Spatial Density Contrasts: Correction for Multiple Comparisons on Statistical Maps: In order to reduce the chance of false positives in our statistical maps due to multiple comparisons, we utilized False Discovery Rate (FDR) control, which enables setting the statistical thresholds to those at which a given estimated rate of false positives can be attained. The AFNI (http://afni.nimh.nih.gov) function 3dFDR was applied to each of the statistical maps. Because approximately half of the pixels in the statistical maps did not correspond to face stimulus pixels and because our aim was to detect fixation differences over internal facial features, the same non-face region mask was applied to all statistical maps before FDR correction so that those pixels would be ignored in the 3dFDR algorithm. Our FDR threshold was set to q < 0.05, at which it would be estimated that 5% of surviving pixels are false positives. Cluster size correction is an alternative method to FDR control for multiple comparisons correction, though in the context of this study, where fine-grained mapping of highly significant regions is preferred to detection of larger area regions, we chose to employ FDR control.
Because AOI analyses can be criticized for requiring a highly subjective a priori segmentation of visual features [42], but spatial statistical maps can be criticized for lacking sensitivity, we conducted additional exploratory analyses that were meant to increase sensitivity without subjective segmentation. In particular, we calculated profile densities (i.e. densities summed along a single dimension of a heatmap) for the different conditions during the study phase. The y-profile plots were the result of summing along the horizontal dimension (x-axis) of a spatial density heatmap. The y-profile plots visualize fixation density over specific facial features without respect to laterality or fine differences in horizontal position. Since the primary effects of interest here focused on which facial features were fixated (eyes, nose, mouth), we report only y-profile plots.
Profile Density Contrasts: Difference Plots: In order to visualize potential differences in profile density between two conditions, spatial density difference maps were summed along the vertical dimension to produce x-profile density difference plots and summed along the horizontal dimension to produce y-profile density difference plots. X-profile density difference plots visualize potential differences in left-right face laterality between contrasted conditions, and y-profile density difference plots visualize potential differences in density over specific facial features.
Profile Density Contrasts: Profile Statistical Maps: To find regions of statistically significant difference in the profile density difference plots, we re-used the 104,000 resampled iterations from the spatial density contrast statistical map analyses to perform a Monte Carlo permutation test on the contrasted profile plots. All resampled iterations from the relevant spatial density Monte Carlo permutation test were summed along the vertical dimension to produce the resampled iterations of the x-profile Monte Carlo permutation test, and were summed along the horizontal dimension to produce the resampled iterations of the y-profile Monte Carlo permutation test. P-values were computed pixel-wise (i.e. at each pixel along the relevant dimension) based on the number of corresponding pixels in the resampling iterations that were greater than a given positively valued pixel (i.e. where condition 1 had a greater profile density) in the true profile density difference plot and that were less than a given negatively valued pixel (i.e. condition 2 greater) in the true profile density difference plot. Maps visualizing the results were thresholded at a pixel significance of p < 0.025 (equivalent to two-tailed p < 0.05). The threshold p-value for the uncorrected profile density maps was numerically higher than that for the uncorrected spatial density maps, in part to accommodate for the fewer multiple comparisons, but also for demonstrative purposes to qualitatively compare effects with those of the AOI analyses. In these maps, pixels along the entire orthogonal dimension were highlighted where the dimension of interest had a significantly different profile density between contrasted conditions.
Profile Density Contrasts: Correction for Multiple Comparisons on Profile Statistical Maps: FDR control with threshold of q < 0.05 was again employed, but this time utilizing all pixels for each profile statistical map.
|
![[This page as RDF]](http://data.gesis.org/somesci/static/rdf-icon.gif){kind=icon}