Fear perception: Can objective and subjective awareness ...lce.umd.edu/publications_files/fear...

Fear perception: Can objective and subjective awarenessmeasures be dissociated?

Department of Psychological and Brain Sciences,Indiana University, Bloomington, IN, USARemigiusz Szczepanowski

Department of Psychological and Brain Sciences,Indiana University, Bloomington, IN, USALuiz Pessoa

Whereas previous studies of fearful-face perception have probed visual awareness according to either objective or subjectivecriteria, in the present study, we probed the perception of briefly presented and masked fearful faces by assessing bothtypes of perception within the same task. Both objective and subjective sensitivity measures were assessed within acommon signal detection theory framework. To evaluate single-participant awareness, we employed a nonparametricreceiver operating characteristic (ROC) analysis of the behavioral data, which involved collecting a large number of trialsover multiple sessions. Our findings revealed that nearly all subjects could reliably detect 17-ms fearful-face targets, thusexhibiting above-chance objective perception at this target duration. Reliable subjective sensitivity was also observed for33-ms fearful-face targets and, for some subjects, even for 17-ms targets. The analysis of single-session data suggeststhat previous experiments may have lacked sufficient statistical power to establish above-chance performance. Takentogether, our findings are consistent with a dissociation of fear perception according to objective and subjective criteria,which could be assessed for each individual participant. The determination of such a dissociation zone may help inunderstanding the conditions linked to aware and unaware fear perception.

Keywords: fear perception, dissociation, awareness

Citation: Szczepanowski, R. & Pessoa, L. (2007). Fear perception: Can objective and subjective awareness measures bedissociated? Journal of Vision, 7(4):10, 1–17, http://journalofvision.org/7/4/10/, doi:10.1167/7.4.10.

Introduction

Visual backward masking is a powerful technique thathas been used in the study of visual awareness during theviewing of emotion-laden visual information (Esteves &Ohman, 1993). In backward masking, the processing of abriefly presented target stimulus is degraded by asubsequent mask stimulus. Stimulus visibility is manipu-lated by varying both target and mask strength and/ortiming. Until recently, it was generally accepted thatemotional faces are processed during Bsubliminal[ con-ditions, namely, when subjects are unaware of the maskedemotional expression (Dolan, 2002; Ohman, 2002). How-ever, recent studies have challenged this view and claimedthat visual awareness may be critical for the processing ofsuch stimuli (Pessoa, 2005).

One factor that may help explain the discrepancy ofprevious results concerns the use of different criteria todetermine whether or not a subject is aware of perceiving astimulus. On the one hand, behavioral and neuroimagingstudies that report unaware perception of emotional faceshave often evaluated awareness according to subjectivecriteria (e.g., Whalen et al., 1998). According to subjectivecriteria, unaware perception occurs when subjects reportnot having seen target stimuli or being unable to performthe task better than chance (independent of their actual

performance). Subjective criteria hold that only thesubjects themselves have access to their inner states andthat their introspection is a reliable source of informationabout conscious experiences (Merikle, Smilek, & Eastwood,2001). On the other hand, objective criteria have been usedin studies that have suggested that awareness may benecessary for the processing of emotional faces (Pessoa,Japee, & Ungerleider, 2005; Pessoa, Japee, Sturman, &Ungerleider, 2006). According to objective criteria,unaware conditions occur when a subject’s performancein a yes/no or forced-choice task is at chance, such as whensubjects fail to detect alternative stimulus states (presencevs. absence of visual targets). Thus, the perception of visualstimuli is objectively unaware when the sensitivity measureof awareness (such as dV) is at the null levelVthat is,subjects exhibit null sensitivity. Under such conditions,behavioral effects of unaware stimuli (e.g., faster reactiontime for undetected fearful faces), as well as associatedphysiological or neuroimaging signals, would constitutecorrelates of unaware perception.

Unfortunately, no consensus exists regarding the Bbest[approach to measure and characterize awareness. In fact,there is a mounting tension between objective andsubjective threshold approaches (Merikle et al., 2001;Snodgrass, Bernat, & Shevrin, 2004). Historically, animportant concern with subjective procedures is that theycan be quite sensitive to response bias (Eriksen, 1960). For

Journal of Vision (2007) 7(4):10, 1–17 http://journalofvision.org/7/4/10/ 1

doi: 10 .1167 /7 .4 .10 Received August 19, 2006; published March 27, 2007 ISSN 1534-7362 * ARVO

Downloaded From: http://jov.arvojournals.org/pdfaccess.ashx?url=/data/Journals/JOV/933521/ on 05/18/2015 Terms of Use:

http://www.indiana.edu/%7Elceiub


mailto:[email protected]?subject=http://journalofvision.org/7/4/10/






http://journalofvision.org/7/4/10/

instance, subjects may be reluctant to indicate having seena stimulus when the available evidence is very weak. Suchconcern came into sharp focus with the development ofsignal detection theory (SDT; Green & Swets, 1966). Asshown by SDT models, subjective threshold effects arehighly vulnerable to alternative conscious perceptionexplanations. At the same time, the importance ofsubjective measures of awareness has been that theyresonate with the idea that an intuitively appealingdefinition of awareness should be based on introspectivereports of individuals’ inner states (James, 1890). In fact,from the subjective awareness perspective, at times, it hasbeen stated that the utilization of objective measureslogically precludes the existence of unconscious percep-tion (Bowers, 1984; Ohman, 1999).

As further addressed in the Discussion section, we arguethat it is important to investigate both objective andsubjective measures of perception. Accordingly, in thepresent study, to investigate the perception of minimallyvisible stimuli, our approach was to employ both types ofmeasures. In the past, the comparison between objectiveand subjective measures has been hampered by the use ofdifferent tasks to gauge them. For example, althoughsubjects may perform a forced task during studiesmeasuring objective thresholds, they typically employsome form of nonforced, self-report during studiesmeasuring subjective thresholds. Here, our goal was toobtain objective and subjective measures of perceptionwithin a single task. Initially, we will focus on character-izing the perception of minimally visible stimuli with bothtypes of measures, without necessarily drawing conclu-sions concerning what they might mean in terms of awareversus unaware perception. We will discuss this issue inthe Discussion section, where we address the potentialimplications of our findings.

In the present study, in each trial, subjects viewed a targetface that was fearful or nonfearful, which was followed by amask stimulus (Figure 1). Subjects were instructed to firstprovide a response that indicates whether or not a fearfulface was present in the display at any time. Suchcomponent of the task allowed us to compute a standard

sensitivity index (such as dV) that could be used to assessobjective perception. In addition, subsequently, subjectsindicated the level of confidence of their yes/no responses.As others in the past, we reasoned that we could use thesubject’s ability to discriminate between correct andincorrect responses based on their confidence ratings asan index of subjective perception (Kolb & Braun, 1995;Kunimoto, Miller, & Pashler, 2001). Thus, if subjectscould access information about the stimulus, higherconfidence trials should be linked with correct responsesmore often than with incorrect responses; likewise, lowerconfidence trials should be linked with incorrect responsesmore often than with correct responses. In other words, therelationship between response confidence and accuracyshould provide an index of subjective perception.

Figure 2 provides a representation of the two tasksperformed by the subject. The top part depicts the standarddetection task, namely, a yes/no choice between two typesof event. The yes/no detection task can be used as anindex of objective perception. The bottom part depicts thesecond task, which is known as a Type 2 task (Galvin,Podd, Drga, & Whitmore, 2003; Pollack & Decker, 1958),and can be viewed as an extension of the standarddetection problem. One of two events, correct or incor-rect, occurs when the observer makes the yes/no decision.The observer is then asked to decide BYes, it was correct[or BNo, it was not correct[ during the decision interval.Alternatively, as in our experiment, the observer is askedto provide a confidence rating during the decision interval,which can be used to gauge subjective perception.

A useful representation of the trial types of the rating taskcan also be formulated within the SDT framework:Discrimination accuracy (correct vs. incorrect) can bepaired with response confidence (Blow[ vs. Bhigh[) to yielda 2 � 2 table (Table 1). Within this representation, a Bhit[corresponds to a correct trial with high confidence; aBmiss[ corresponds to a correct trial with low confidence;a Bfalse alarm[ corresponds to an incorrect trial with highconfidence; and a Bcorrect rejection[ corresponds to anincorrect trial with low confidence. Thus, correct trials forwhich the subject rates with Bhigh[ confidence would

Figure 1. Backward masking paradigm. Fearful, happy, or neutral target faces were shown for 17, 25, 33, or 41 ms and were immediatelyfollowed by a neutral face mask that was shown such that the target plus the mask lasted 100 ms. Subjects were first asked whether fearappeared at any time in the display. They were then asked to rate the confidence of their initial response by using a 1-to-6 scale (from lowto high confidence).

Journal of Vision (2007) 7(4):10, 1–17 Szczepanowski & Pessoa 2


comprise Bhit[ trials, whether or not they actually containa target stimulus. At the same time, incorrect trials forwhich the subject rates with Bhigh[ confidence wouldcomprise Bfalse alarm[ trials, irrespective of the type ofincorrect trial (i.e., incorrect trials containing or notcontaining a target stimulus). Given such trial categoriza-tion, it is then possible to compute an SDT-like measureof sensitivity that reflects the distance between the meansof the distributions representing correct and incorrectresponses (Figure 3). We can then see that, in this context,a dV index would indicate the subjects’ ability todiscriminate between their correct and incorrect responses.

In this sense, it provides a (subjective) measure of thesubjects’ own assessment of their performance.

Previous studies of threshold perception of emotionalfaces have evaluated the perception of minimally visiblestimuli at the group level (Esteves & Ohman, 1993). Thus,for instance, an unaware condition is one in which, onaverage, subjects cannot detect emotional faces better thanchance. One result that has emerged from this frameworkis that participants appear to be unaware of maskedemotional faces shown for 30 ms or so. In practice, suchduration has often been treated in the literature as a Bhardthreshold[ that would apply universally (i.e., for allparticipants). Such approach is undesirable because itimplicitly assumes that subjects are homogeneous in termsof emotional face perception. We suggest that a morefruitful strategy is to assess threshold vision at the single-participant level, which is especially important givenrecent evidence of intersubject differences in the ability todetect fearful-face targets (Pessoa et al., 2006). Thus, inthe present study, we investigated threshold vision for allindividual participants separately.

In summary, in the present study, we investigated theperception of emotional faces by employing a backwardmasking paradigm. Both objective and subjective indexesof perception were computed based on the same fear-detection task. To investigate the minimal critical stimulusconditions for perceiving fearful faces according to ourcriteria, we parametrically varied the exposure duration oftarget stimuli. In addition, we examined potential changesin threshold vision by having participants undergo multiplebehavioral sessions (Experiment 1). Such approach

Discrimination accuracy

Confidence

High Low

Correct Hit MissIncorrect False alarm Correct rejection

Table 1. Trial type categorization for subjective perceptionsensitivity (based on Kunimoto et al., 2001).

Figure 2. Abstract representation of the tasks performed by thesubject. The first task (top) corresponded to the standarddetection task. During the observation interval, either a signal(S) or a noise (N) event occurs with probability p(S) and p(N),respectively. During the decision interval, the subject indicates‘‘Yes, a signal occurred’’ or ‘‘No, a signal did not occur.’’ Theobservation and decision intervals of the detection task becomethe observation interval for the Type 2 task (bottom). In this case,either a correct (C) or incorrect (I) event occurs with probabilityp(C) and p(I), respectively. During the decision interval, thesubject indicates ‘‘Yes, it was correct’’ or ‘‘No, it was not correct.’’In the present experiment, detection signal trials comprised fear-containing trials (fearful–neutral target–mask pairs), and detectionnoise trials comprised non-target-containing trials (happy–neutraland neutral–neutral target–mask pairs). In addition, confidenceratings were used instead of ‘‘correct’’/‘‘incorrect’’ judgments(inspired by Figure 1 of Galvin et al., 2003).

Figure 3. Assessing subjective perception with SDT. Two eventtypes corresponding to correct and incorrect trials are representedvia Gaussian distributions. The criterion for reporting confidenceis indicated by the vertical line and dictates when the subjectreports ‘‘high’’ versus ‘‘low’’ confidence. A dVvalue greater thanzero reveals the subject’s ability to discriminate between correctand incorrect responses and can be used to index subjectiveperception. In this article, we employed a nonparametric measureof sensitivity based on ROC curves (inspired by Figure 1 ofKunimoto et al., 2001).



allowed us to test whether or not performance was stableacross time or, alternatively, whether subjects exhibitedimprovements in sensitivity.

Experiment 1: Mixed design

MethodsParticipants

Five volunteers (of whom three were males) participatedin this study (29 T 5 years old), which was approved by theInstitutional Review Board of Brown University. Subjectswere in good health with no past history of psychiatric andneurological diseases and gave informed consent. Subjectshad normal or corrected-to-normal vision. One of theauthors (R.S.) served as a subject.

Stimuli and design

In each trial (see Figure 1), a white fixation cross wasdisplayed for 300 ms on a black background, followed bya 50-ms blank screen, followed by a fearful, happy, orneutral target face, which was immediately followed by aneutral face that served as a mask. In a given trial, theidentity of the face used as the target or the mask wasalways different (as well the identities used in successivetrials). Face stimuli subtended 4- � 5- of visual angle. Tomanipulate stimulus visibility, we varied the duration ofthe target face parametrically: 17, 25, 33, and 41 ms. Thetotal duration of the target–mask pair was fixed at 100 ms.After the presentation of each target–mask pair, subjectshad 2 s to indicate Bfear[ or Bno fear[ via button press.They then had 2.5 s to rate the confidence in their responseon a scale of 1 to 6 (low to high confidence). The totalduration of each trial was 5 s. In each experimentalsession, subjects performed 160 trials shown in a randomorder, which followed a 3 (target–mask pairs: fearful–neutral, happy–neutral, and neutral–neutral) � 4 (targetdurations: 17, 25, 33, and 41 ms) design. Half of the trialscontained a fearful-face target and half did not (1/4happy–neutral and 1/4 neutral–neutral target–mask pairs).Subjects did not receive any information about stimulusdurations or trial types. Subjects performed a total of fiveindependent sessions; three subjects performed the ses-sions in consecutive days and two subjects performed thesessions within 4 days, with two sessions in 1 day(separated by at least 2 hr). Overall, all subjects performed3,200 trials (800 trials for each duration).

Face stimuli were obtained from the Ekman set (Ekman& Friesen, 1976), a second set elaborated by Ohman et al.(KDEF, Lundqvist, D., Flykt. A., and Ohman, A.;Karolinska Hospital, Stockholm, Sweden), and a third setvalidated by Alumit Ishai at NIMH (Bethesda, USA; Ishai,Pessoa, Bikle, & Ungerleider, 2004). Stimuli weredisplayed on a ViewSonic Professional Series P95f+

monitor (screen refresh rate: 120 Hz) fitted with a Radeon9800 AGP graphics card. Measurements using a photo-diode and a digital oscilloscope (Tektronix, TDS3000Bseries) confirmed that the targeted durations were keptconstant in our setup.

We employed, as targets, 40 instances each of fearful,happy, and neutral faces, and, separately, we used 80neutral faces as masks. In most masking studies involvingemotional faces, only neutral–neutral trials are employed asBnoise[ trials. In the present study, happy faces wereincluded to more closely match fearful faces in terms oflow-level features, such as brightness around the mouth andeye regions, because both fearful and happy faces tend to bebrighter than neutral in these regions. Thus, the inclusion ofhappy faces reduced a subject’s ability of detecting fearfulfaces by using such low-level cues. In addition, theinclusion of happy faces discouraged subjects from adopt-ing a strategy of indicating fear whenever features deviatedfrom those of a neutral face. Finally, to prevent theBdetection[ of fear based on subtle motion cues from thetransition between fearful and nonfearful faces, maskstimuli were randomly displaced so as to not perfectlyoverlap the target face. Specifically, on approximately halfof the trials (chosen randomly), the mask stimulus wasshifted along one of the four diagonal directions by a smallspatial offset of È0.5- of visual angle (Phillips et al., 2004).

SDT for the analysis of objective and subjectiveperception: Nonparametric receiver operatingcharacteristic approach

To evaluate individual objective and subjective meas-ures, we employed a nonparametric receiver operatingcharacteristic (ROC) analysis of the behavioral data. Suchapproach does not make normality assumptions about thesignal and noise distributions (Green & Swets, 1966;Hanley & McNeil, 1982; Macmillan & Creelman, 2005).ROC curves of both objective and subjective measureswere generated (see below), and the area under the curve(AV; Egan, 1975; Green & Swets, 1966) was employed asAobjectiveV and AsubjectiveV values, respectively.

ROC curves for objective perception were estimatedbased on yes/no decisions (Bfear[/Bno fear[) and thecorresponding confidence ratings on a six-point scale(Macmillan & Creelman, 2005), which were used to obtaincumulative conditional probabilities (Green & Swets,1966; Macmillan & Creelman, 2005). The hit rate wasdefined as the probability of reporting Bfear[ given that afear-containing stimulus was shown ( p(Bfear[ | fear-containing stimulus)), whereas the false alarm rate wasdefined as the probability of reporting Bfear[ given that anon-fear-containing stimulus was shown (p(Bfear[ | non-fear-containing stimulus)).

ROC curves for subjective perception were estimatedbased on response correctness and confidence as motivatedin the Introduction (see Table 1). Correct responses involvedBfear[ report for fear-containing trials and Bno fear[ report



for non-fear-containing trials. Incorrect responses includedBno fear[ report for fear-containing trials and Bfear[ reportfor non-fear-containing trials. In the context of subjectiveawareness, we considered Bhits[ as correct trials associatedwith high-confidence ratings and Bfalse alarms[ as incorrecttrials with high-confidence ratings (Kunimoto et al., 2001;Pollack & Decker, 1958). We considered high-confidencetrials as those with ratings between 4 and 6 (inclusive);critically, other partitioning schemes, such as those basedon the median (i.e., Bmedian split[), yielded nearlyindistinguishable results. The hit rate was defined as theprobability of reporting Bhigh confidence[ given that thetrial was correct ( p(Bhigh confidence[ | correct)), whereasthe false alarm rate was defined as the probability ofreporting Bhigh confidence[ given that the trial wasincorrect ( p(Bhigh confidence[ | incorrect)). These defini-tions follow those employed in Type 2 SDT analyses(Galvin et al., 2003; Kunimoto et al., 2001; Macmillan &Creelman, 2005; Pollack & Decker, 1958).

When plotting ROC curves (Figure 4), pairs of hit andfalse alarm rates derived from 12 alternative yes/nodecisions (6 levels for reporting Bfear[ and 6 levels forreporting Bno fear[) were used to plot objective ROC

curves; pairs of hits and false alarms derived from 6confidence ratings (3 levels for reporting Bhigh con-fidence[ and 3 levels for reporting Blow confidence[)were used to plot subjective ROC curves. All ROCsoriginated at (0,0) and ended at (1,1).

To assess the statistical significance of the area under theROC curve, we employed two methods. First, we employedbootstrap resampling methods (Efron & Tibshirani, 1993)for the estimation of confidence intervals around AVvalues, as implemented in the AccuROC software package(Accumetric Corporation, Montreal, Canada). We deemedsubjects as objectively or subjectively above chance whenAVmeasures exceeded 0.5 (i.e., the area under the y = xline when the hit and false alarm rates are equivalent) andthe 95% confidence interval did not overlap with 0.5; infigures, error bars indicate 95% bootstrap confidenceintervals based on 2,000 iterations. In addition, to reportp values, we employed the method of DeLong, DeLong,and Clarke-Pearson (1988). In practice, the two methodsprovided equivalent results.

Because multiple statistical tests were applied simulta-neously in the context of Tables 2, 3, and 4, without anexplicit correction for multiple comparisons, the probability

Figure 4. ROC curves for two representative subjects. Top: Subject S3 exhibited above-chance objective and subjective sensitivity at alltarget durations. Bottom: Subject S4 exhibited above-chance objective sensitivity at all target durations and above-chance subjectivesensitivity for the 25-, 33-, and 41-ms durations. For 17-ms targets, Subject S4 could detect fearful faces better than chance but failed toreliably discriminate between correct and incorrect responses. Such pattern of results is suggestive of a dissociation between objectiveand subjective perception (see Figure 10).



of a Type I error may have attained unacceptable levels.Statistical significance was thus corrected for multiplecomparisons according to the Bonferroni method. ForTable 2, the ! level was .0013 (.05/40, where .05 is thestandard ! level and 40 corresponded to the number ofsimultaneous tests [5 subjects � 4 durations � 2 perceptionmeasures]). For Table 3, the ! level was .0003 (.05/150,where .05 is the standard ! level and 150 corresponded to thenumber of simultaneous tests [5 subjects � 3 durations � 2perception measures � 5 sessions]); in Table 3, we onlyconsidered the 17-, 25-, and 33-ms target durations, as41-ms targets were essentially suprathreshold. For Table 4,the ! level was .0021 (.05/24, where .05 is the standard !

level and 24 corresponded to the number of simultaneoustests [6 subjects � 2 durations � 2 perception measures]).For completeness, uncorrected p values are provided inTables 2, 3, and 4, and color coding is used to indicate whenvalues were significant when multiple comparison wereemployed (green), when they were significant when multi-ple comparisons were not employed (pink), and when theywere not significant (red).

In an important paper, Pollack and Hsieh (1969) showedthat the sampling variability of the area under the ROCcurve is well approximated by a binomial probability of amean proportionVthis is especially the case when valuesare not near the extremes of the distribution (e.g., when AV

Table 2. Experiment 1: Objective and subjective sensitivity values (AV): pooled data. Note: p values are enclosed in parentheses. Green:statistically significant, corrected for multiple comparisons; pink: statistically significant, not corrected for multiple comparisons; red: notstatistically significant. p values of .000 indicate values less than .001.

Table 3. Experiment 1: Objective and subjective sensitivity values (AV): single-session data. Note: p values are enclosed in parentheses.Green: statistically significant, corrected for multiple comparisons; pink: statistically significant, not corrected for multiple comparisons;red: not statistically significant. p values of .000 indicate values less than .001.



does not approach 1). To investigate the issue of samplesize, we performed a power analysis under the binomialassumption (Figure 7).

ResultsResults for trials pooled across sessions

We initially investigated our results by pooling trialsacross the five sessions. To determine individual objectiveand subjective measures during the fear-detection task, wecomputed sensitivity values (AV) for the following targetdurations: 17, 25, 33, and 41 ms. Examples of bothobjective and subjective ROC curves for two representa-tive subjects are presented in Figure 4. Subject S3 couldreliably detect fearful faces, as well as reliably discrim-inate between correct and incorrect responses, for allstimulus durations. On the other hand, Subject S4 coulddetect fearful faces at all stimulus durations but wasunable to discriminate correct/incorrect responses for17-ms targets (the ROC curve for 17-ms targets fell alongthe diagonal). Thus, for 17-ms targets, Subject S4 coulddetect fear but did not exhibit evidence of accessinginformation concerning his or her performance. Table 2displays AVvalues for all subjects and conditions.

Somewhat surprisingly, our results revealed that AobjectiveVvalues reliably exceeded 0.5 (the value expected bychance) for all subjects even during the 17-ms condition;thus, all subjects could detect 17-ms targets. In terms ofAsubjectiveV values, all subjects discriminated correct/incorrectresponses for target durations of 33 and 41 ms; all subjects,except for Subject S1, were above chance for the 25-msduration; Subjects S2 and S3 exhibited above-chanceperformance even for very brief 17-ms targets (see the Nullsensitivity and statistical power section).

The above results indicate potential dissociations betweenobjective and subjective values that can be further visualized

by considering Figures 5A and 5B. When error bars ofsubjective measures (red) cross the 0.5 line and error barsof objective measures (blue) do not cross this line, evidenceof a dissociation occurs (see arrows). Note that the errorbars do not take into account any correction for multiplecomparisons; thus, although they are useful measures ofvariability for an individual subject, they are too liberalwhen one considers the series of results shown in Table 2(which take into account multiple comparisons).

To rule out the possibility that the potential dissociationsbetween objective and subjective measures stemmed fromour method of computing AV, we also computed concurrentdVsensitivity measures (which assume that signal and noisedistributions are normally distributed; see Appendix). Asseen by considering Figures 5A and 5B, both methodsmade relatively consistent assessments of dissociations (forthe 17-ms condition of Subject S1, the dsubjectiveV valuealso crossed the chance line, consistent with the results ofTable 2). Combined, the two analyses were consistent withan objective/subjective dissociation for 17-ms targets forSubjects S4 and S5 and a dissociation for 25-ms targets forSubject S1 (here, we refer to the dissociations suggested byboth methods).

Objective and subjective measures as a function ofnoise trials

In the present study, Bsignal[ trials involved fearful–neutral target–mask pairs, whereas noise trials involvedtwo types of trials: happy–neutral and neutral–neutral pairs.In the preceding analyses, these two trial types were pooledtogether in our estimates of AobjectiveV and AsubjectiveV values.It is instructive to compute these measures as a function ofnoise type (i.e., after sorting trials based on noise type).The results are shown in Figure 6; only the 17- and 25-mstarget durations are shown given that they wereBperithreshold[ conditions. First, let us consider AobjectiveVvalues (blue) for the 17-ms duration (top). It is interestingto note that all values were suprathreshold (i.e., 90.5)when neutral–neutral trials (right) comprised the noise.When happy–neutral trials were considered (left), SubjectsS4 and S5 did not exceed 0.5 and Subject S1 barely did so(as before, the error bars may be thought as Bliberal[given that they do not incorporate a correction for multiplecomparisons). These results suggest that these subjectscould reliably detect fearful targets only insofar as thesediffered from neutral stimuli; in other words, they mayhave solved the task by adopting a strategy of detectingdeviations from neutral and not necessarily by detectingfear per se. However, Subjects S2 and S3 were able toreliably detect fearful targets even in the context ofhappy–neutral noise, suggesting that they were able todetect fear (i.e., not only deviations from neutral).

Now, let us consider AsubjectiveV values for the 17-msduration (Figure 6; top, red). Three subjects (S1, S4, andS5) did not exhibit above-chance behavior in the context ofhappy–neutral noise trials (left), and two subjects (S3 and

Table 4. Experiment 2: Objective and subjective sensitivity values(AV): pooled data. Note: p values are enclosed in parentheses.Green: statistically significant, corrected for multiple comparisons;red: not statistically significant. p values of .000 indicate valuesless than .001.



S4) did not exhibit above-chance behavior in the context ofneutral–neutral noise trials (right; see Table 2). In theformer case, the elimination of subjective perception waspretty complete (the AsubjectiveV values were 0.5, 0.47, and0.51, respectively), suggesting that the correct versusincorrect judgment was more difficult during happy–neutraltrials (left) than during neutral–neutral trials (right).

Finally, as shown in Figure 6, objective/subjectivedissociations appeared to occur when happy or neutralfaces were considered as noise targets. For instance, for17-ms targets, dissociations were observed for Subjects S3and S4 (and possibly for Subject S5) when neutral–neutraltrials (top, right) were considered as noise. For 25-mstargets, dissociations appeared to occur for Subject S1(and possibly Subject S4) when happy–neutral trials(bottom, left) were considered as noise.

It should be noted that because subjects performed trialsin a mixed fashion (all trial types were randomized), it isconceivable that participants found it difficult to maintainstable response criteria for both yes/no and confidencerating responses across the experiment (see Experiment 2).Thus, some of the differences between happy and neutralnoise trials reported here may be due to the adoption ofdifferent response criteria for these trial types.

Perception measures as a function of experimentalsession

Given that subjects participated in five 1-hr sessions, weprobed how objective and subjective sensitivity valuesvaried as a function of experimental session.

Objective perception

First, let us discuss the 17-ms duration (Table 3). On theone hand, Subjects S2 and S3 consistently exhibitedsuprathreshold AobjectiveV values for single sessions (forSubject S3, this was the case when a correction formultiple comparisons was not considered). On the otherhand, Subjects S1, S4, and S5 often failed to exhibitsuprathreshold AobjectiveV values for single sessions. Theseresults, taken together with those obtained when allsessions were pooled together (Table 2), were evaluatedin the context of both (a) statistical power and (b) learningeffects (i.e., changes in performance across sessions).

During single sessions, for 17-ms targets, for Subjects S1,S4, and S5, AobjectiveV values (Table 3) that exceeded 0.5were relatively modest (with the exception of Session 5 forSubject S1 for which AobjectiveV was 0.68). Consequently,single sessions may have been underpowered to revealsuprathreshold AobjectiveV values (even when corrections for

Figure 5. Objective and subjective measures of perception. Values of AobjectiveV (blue) and AsubjectiveV (red) are plotted as a function ofsubject for Experiment 1 (A and B) and Experiment 2 (C and D). (A and C) Perception measures based on the nonparametric analysis.(B and D): Perception measures based on the parametric analysis. The arrows indicate conditions for which there was evidence for adissociation between objective and subjective perception. See text for further details. Error bars indicate 95% bootstrap confidenceintervals.



multiple comparisons are not considered). To furtherinvestigate this issue, we performed a power analysis(Methods section), the results of which are shown inFigure 7. For an individual session, the critical value fordetecting a true deviation from 0.5 with 80% power(! = .05) was slightly more than 0.6. Such results reveal thatindividual sessions are probably underpowered to detecttrue suprathreshold perception at short durations, which aretypically associated with modest sensitivity values (i.e.,G0.6). At the same time, the critical value of AobjectiveV wasapproximately 0.55 when trials from the five experimentalsessions were pooled together (80% power; ! = .05),suggesting that the present experimental design hadadequate statistical power for the detection of true objectivesensitivity at short target durations (see values in Table 2).

To investigate potential changes in performance, weplotted objective sensitivity values as a function of sessionnumber (relative to first-session values). Figure 8 illustratesthat, on average, objective measures increased acrosssessions; improvements were observed for all stimulusdurations. More formally, a significant linear trend wasobserved (F = 68.9, p = .001).

Subjective perception

For 17-ms targets, few individual sessions exhibitedAsubjectiveV values significantly greater than chance (Table 3).However, for at least Subjects S2 and S3, results weresignificant when the data from all five sessions were pooledtogether. See the Discussion section for further elaborationon statistical power.

We also investigated learning effects for subjectivemeasures. An interesting pattern emerged when we consid-ered the evolution of AsubjectiveV values across sessions.

Although values were relatively flat for the 17- and 25-msdurations, an increase in sensitivity was evident for 33- and41-ms targets (Figure 8). Overall, the linear trend onlyapproached significance (F = 6.088, p = .07), possibly dueto differences in the trends for the two shorter and twolonger durations.

Experiment 2: Blocked design

In Experiment 1, participants performed trials in arandom (intermixed) order, with each of the 12 differentstimuli (3 target–mask pairs � 4 durations) occurring withequal probability. Thus, it is conceivable that participantsdid not adopt stable response criteria for both yes/no andconfidence rating responses across the experiment. Suchscenario complicates the interpretation of AobjectiveV andAsubjectiveV values in general, as well as potential dissoci-ations between these measures, in particular. Accordingly,in Experiment 2, we adopted a Bblocked[ design for whichstimulus conditions were held constant throughout blocks,maximizing the likelihood that participants would adopt astable criterion.

MethodsParticipants

Six volunteers (of whom one was male) participated inthis study (29 T 7 years old), which was approved by theInstitutional Review Board of Indiana University, Bloo-

Figure 6. Objective and subjective perception measures as a function of noise trials. Objective measures are shown in blue, andsubjective measures are shown in red. Top: Results for 17-ms targets based on happy–neutral noise trials (left) or based on neutral–neutral noise trials (right). Bottom: Results for 25-ms targets based on happy–neutral noise trials (left) or based on neutral–neutral noisetrials (right). Error bars indicate 95% bootstrap confidence intervals.



mington. Subjects were in good health with no past historyof psychiatric and neurological diseases and gave informedconsent. Subjects had normal or corrected-to-normalvision.

Stimuli and design

Stimuli and design were very much like those in Experi-ment 1; here, we focus on the differences between the twoexperiments. Target faces were either fearful or neutraland were shown for 17 or 25 ms. Thus, Experiment 2employed a 2 � 2 factorial design. During each session,subjects performed 800 trials, which occurred in a blockedfashion. During each block, a single target duration wasemployed, which was indicated via an initial Binstructionscreen[ at the beginning of the block. Blocks contained 80trials presented in a random order, with half of themcontaining fearful–neutral target–mask pairs and the otherhalf containing neutral–neutral target–mask pairs. Blocktype was randomized across the session. Participantsperformed a total of two independent sessions (separatedby 1–2 days for five subjects; one subject performedsessions during the same day, which were separated by2 hr). Overall, subjects performed 1,600 trials (800 trialsat each target duration).

Comparing observed and predicted objective andsubjective values

To further explore potential dissociations between objec-tive and subjective measures, we compared observed valuesto predicted values, as computed via computer simulations.Following an ideal-observer framework, samples weredrawn from pairs of overlapping Gaussian distributionswith dVseparations from 0.0 to 2.5 (in steps of 0.1). A total

of 400 independent random samples were drawn (200 noisesamples and 200 signal + noise samples). As done in thepast, we assumed that the distributions had a fixed standarddeviation ratio of 1:1 (Hajian-Tilaki, Hanley, Joseph, &Collet, 1997; Obuchowski, 1994). A six-point confidencescale was simulated by considering equidistant cutoffs.Note that this procedure is resistant to departures from thenormality assumption of the sensory distributions. Thus,although the ideal-observer model does not explicitly takeinto account different configurations of pairs of overlappingnoise and signal + noise distributions, ROC-based measuresare robust to departures from normality and potential biasesintroduced in their estimation are negligible (Hajian-Tilakiet al., 1997). Both objective and subjective ROC curveswere then computed, and the area was evaluated (AVvalues).For subjective measures, we considered high-confidencetrials as those with ratings between 4 and 6 (inclusive), asdone with actual data (see above). For both objective andsubjective measures, the procedure was repeated 100 timesto obtain average sensitivity values, which are plotted inFigure 9. A separate set of simulations was performedfollowing methods suggested by Pollack and Hsieh (1969)and yielded very similar results.

Results

The results of Experiment 2 (Table 4; Figures 5C and5D), which employed a blocked design, were largelyconsistent with the findings of Experiment 1, whichemployed a mixed design. Five of the six subjects could

Figure 7. Statistical power curves. The number of trials is plottedas a function of the critical value of AV at several levels ofstatistical power (! = .05, two-tailed test). Thus, for example, todetect a true sensitivity of AV= 0.6 with 80% power, approximately300 (È2 sessions) trials would be required. To detect a truesensitivity of AV= 0.54 (vertical dashed line) with 80% power,approximately 1,300 (È8 sessions) trials would be required.

Figure 8. Learning effects. Sensitivity measures are plotted as afunction of session number (as a ratio relative to Session 1values). Top: Objective values increased across sections (lineartrend, p G .01). Bottom: Subjective values showed some evidenceof increase across sessions, but the results were not significant(linear trend, p = .07).



reliably detect 17- and 25-ms fearful faces (objectivemeasure). At the same time, Subjects S2, S4, and S5 werenot able to accurately discriminate between correct andincorrect responses for 17-ms targets; Subjects S4 and S5also did not exhibit above-chance subjective values for25-ms targets. The results of Experiment 2 also revealedobjective/subjective dissociations as observed in Experi-ment 1, such as, for instance, those exhibited by SubjectsS2 and S5 during 17-ms targets. Taken together, thereplication in Experiment 2 of the general pattern observedin Experiment 1 suggests that the assessment of objectiveand subjective measures in the first experiment was notstrongly contaminated by unstable response criteria.

Dissociation between objective and subjectivemeasures in Experiments 1 and 2

The results of Experiments 1 and 2 suggested a potentialdissociation between objective and subjective measures.In general, correct/incorrect discrimination scores areexpected to be lower than yes/no sensitivity valuesVfor

instance, subjects may typically find the latter task harderthan the first (Kunimoto et al., 2001; Galvin et al., 2003).This pattern of results was clearly observed in our data(Figures 5A and 5B). It is important then to investigate theobserved differences between objective and subjectivemeasures and how these relate to expected differencesbetween the two measures, as shown in Figure 9. In thisfigure, the diagonal lines represent the case in whichobjective and subjective values for a given condition areexpected to be equal; the horizontal lines represent thesubjective chance line; finally, the blue lines providepredicted values based on simulations (Methods section).In Experiment 1 (Figure 9A), the error bars of severalobserved values overlapped with the predicted line. At thesame time, for Subject S1, the 25-ms target duration didnot overlap with the prediction line but crossed the chanceline, consistent with a stronger dissociation betweenobjective and subjective measures. In Experiment 2(Figure 9B), most observed values were actually lowerthan predicted ones. In particular, for Subjects S2 (17 ms)and S5 (17 and 25 ms), observed values did not overlapwith the prediction line but crossed the chance line, againconsistent with a stronger dissociation. If we consider theresults of Experiments 1 and 2 collectively, 15 of 20points fell below the predicted values, a deviation that isstatistically significant (binomial test, p G .05; here, weexcluded Subject S4 of Experiment 2, who was unable toreliably perform the task). Thus, overall, the comparisonof observed and predicted data is consistent with adissociation between objective and subjective measures.

Naturally, the relationship between observed and simu-lated data critically depends on the assumptions of thecomputational method (see the Methods section for adiscussion of the procedure and assumptions). Forinstance, like other similar models in the past (Hajian-Tilaki et al., 1997; Obuchowski, 1994), we did not assumeany loss of information. In particular, no memory trace forthe stimulus was assumed (Gold, Murray, Sekuler,Bennett, & Sekuler, 2005). However, because confidenceratings were made after the detection of the yes/noresponse, the existence of some form of memory tracefor the stimulus event could lead to confidence ratingsbeing made on noisier representations. Thus, in futurework, it would be desirable to investigate a broader rangeof models.

Discussion

In the present experiment, subjects performed a singledetection task for which they were required to make twoevaluations: an evaluation of whether or not a fearful facewas present and an evaluation of their response confidence.As discussed in the Introduction, the first evaluationcorresponds to a yes/no choice during a standard detection

Figure 9. Simulated and observed objective and subjectivevalues. (A) Experiment 1. (B) Experiment 2. Simulated resultsare shown in blue, whereas specific data points are shown in red(17 ms) and green (25 ms). Evidence of stronger dissociationbetween objective and subjective measures is given by points thatdo not overlap with the prediction line but cross the horizontal‘‘chance’’ line. Error bars indicate 95% bootstrap confidenceintervals.



task. The second evaluation can be thought of as adiscrimination between correct and incorrect responses(by providing Blow[ and Bhigh[ confidence ratings).Whereas the detection task explicitly probed subjects’perception of fear, confidence ratings provided an indirectassessment of the accessibility of information about fear.Thus, our experiment allowed us to test for potentialdissociations between these two measures of fear percep-tion, for instance, whether the successful detection offearful faces would be accompanied by random discrim-ination of response correctness.

Summary of findingsObjective perception

Until recently, it was largely assumed that, undermasking conditions similar to those employed here,participants were unable to detect emotional target facesfor durations of È30 ms. In a recent report (Pessoa et al.,2005), we challenged this view given that, in our maskingstudy, 7 of 11 subjects were able to reliably detect 33-msfearful-face targets; in addition, 2 of 11 subjects wereable to reliably detect 17-ms-target faces (see alsoMaxwell & Davidson, 2004). The present findings con-firm and extend our previous results. Indeed, when thedata from multiple experimental sessions were pooledtogether (Experiments 1 and 2), all but one subjectexhibited above-chance detection of fearful faces for allstimulus durations, including 17- and 25-ms targets. For17-ms targets, it is important to note, however, that whensingle sessions of Experiment 1 were considered inisolation, AobjectiveV values often failed to differ signifi-cantly from 0.5 (see further discussion in the Nullsensitivity and statistical power section).

Subjective perception

Our results revealed that subjects not only couldobjectively detect fearful faces at short stimulus durationsbut also demonstrated above-chance subjective perform-ance. Specifically, all subjects in Experiment 1 exhibitedreliable, correct versus incorrect discrimination for dura-tions of 33 and 41 ms. In addition, most subjects exhibitedabove-chance discrimination for 25-ms targets and severalsubjects exhibited subjective sensitivity even for verybrief 17-ms targets (Experiments 1 and 2). Taken together,our results demonstrate that sensitivity for briefly pre-sented fearful faces is not limited to detection tasks but thatit can be demonstrated for subjective perception as well.

Individual differences

Although our experiment was not designed to investigateindividual differences, our results revealed that subjectsvaried considerably in their ability to detect fearful facesand to perform correct/incorrect discriminations. For

example, in Experiment 1, Subjects S2 and S3 were betterat detecting fearful faces than the remaining subjects. Infact, objective measures from all individual sessions ofSubject S2 were statistically significant even after correct-ing for multiple comparisons; the same is true for the lasttwo sessions of Subject S3.

Null sensitivity and statistical power

Naturally, establishing null sensitivity for any measureof perception faces extreme methodological difficulties(Macmillan, 1986; Reingold & Merikle, 1988). In thepresent study, we attempted to mitigate this problem byhaving subjects perform a relatively large number of trialsacross multiple experimental sessions (i.e., much largerthan related studies). In the present section, we focus onthe results of Experiment 1 for which single- and multiple-session results were investigated more closely; comparablepoints can be made in the context of Experiment 2. Forobjective perception, we did not find evidence for nullsensitivity even for the shortest duration of 17 ms.However, in four cases, we obtained evidence forsubjective null sensitivity: 17 ms for Subjects S1 (whena correction for multiple comparisons was employed), S4,and S5; 25 ms for Subject S1.

How should these results be interpreted? We tentativelysuggest that the AsubjectiveV values of 0.51 (Subject S4,17 ms) and 0.52 (Subject S1, 25 ms) do, in fact, reflectnull sensitivity (Experiment 1). In these cases, it isextremely unlikely that performing additional sessions(for instance, doubling the number of sessions) wouldhave allowed us to detect true sensitivity at a reasonablelevel of statistical power (see Figure 7). At the same time,the value of 0.54 (Subjects S1 and S5, 17 ms) is lessstraightforward to interpret. The power analysis suggestedthat true sensitivity at a level of 0.54 would requireapproximately eight sessions (which is not too far fromwhat we employed) at an 80% level of power (see dashedline in Figure 7). Thus, we may have failed to reject nullsensitivity because of relatively low statistical power(from our plot, we seemed to have had approximately60% statistical power).

The above power considerations assumed that a standard! level of .05 was employed. However, in our experiment,we deemed necessary to correct p values given that asizeable number of simultaneous comparisons was made.In this context, the probability of making at least one TypeI error would be [1 j (1 j !)C], where C is the number ofcomparisons. Such probability would assume unaccept-able values when, for instance, the 40 comparisons ofTable 2 were made (p = .87); that is, at least one Type Ierror would have occurred with Bnear certainty.[ Theseconsiderations suggest that using an ! level of .05 forpower calculations may be Btoo liberal[ and that thenumber of trials indicated by the power analysis may bean Bunderestimate.[



In summary, as extensively discussed in the past,establishing null sensitivity is surrounded by great meth-odological difficulties (Macmillan, 1986; Reingold &Merikle, 1988). The present results suggest that subjectivenull sensitivity occurred for two subjects (Subject S1,25 ms; Subject S4, 17 ms) and, possibly, for a thirdsubject (Subject S5, 17 ms). Our results also suggest thatsingle-session measures from the present study, or fromother studies in the literature, need to be interpretedcarefully because they may have been underpowered todetect true sensitivity.

Learning effects

The five-session format of Experiment 1 allowed us toinvestigate potential changes in task performance as afunction of session. Such Blearning effects[ were observedin a more reliable manner for objective measures; forsubjective measures, only a statistical trend was found.Because we characterized performance in terms of SDTinstead of, for instance, a percent-correct index, it is likelythat increases in sensitivity reflected enhancements in theability to detect fear per se. Such learning may beconceptualized as increasing the distance between thesignal and noise distributions as subjects performed moretrials. Although further experiments are needed to inves-tigate this question, it is possible that performanceimprovements relied on the ability to focus on, or moreeffectively employ, discriminative facial features involv-ing the eyes, the mouth, or both, which are known to beimportant features for the perception of emotional expres-sions (Adolphs et al., 2005; Gold, Sekuler, & Bennett,2004; Schyns, Bonnar, & Gosselin, 2002; Schyns & Oliva,1999; Spezio, Adolphs, Hurley, & Piven, 2007).

Experimental logic revisited: What dopotential dissociations mean in terms ofvisual awareness?

A common experimental logic adopted in awarenessstudies has been to seek to demonstrate a dissociationbetween two measures of perception (Merikle et al., 2001;Snodgrass, 2004; Snodgrass et al., 2004). In our experi-ment, subjects performed a single task for which theywere required to make detection and discriminationevaluations. Our findings were consistent with a dissoci-ation between these two measures of fear perception suchthat, for some subjects, the successful detection of fearfulfaces was accompanied by random discrimination ofresponse correctness. Such dissociation was supported byconverging analyses of our data, including power analysis(Figure 7) and simulation results (Figure 9). Figure 10illustrates the potential relationship between the criticaldurations associated with objective and subjective mea-sures. The shorter duration represents the critical duration

for reliably detecting a fearful-face target, and the longerduration represents the critical duration for reliablydiscriminating between correct and incorrect performance.The shaded area indicates a potential dissociation zone inwhich the subject is objectively above threshold butsubjectively below threshold. The critical question thatwe next turn to is the following: What does suchdissociation zone imply? There are at least three answersto this question.

Subjective unawareness with objective awareness

In the past, proponents of the importance of subjectivemeasures of awareness have suggested that such dissocia-tions be interpreted as evidence for unaware perception. Inthe present context, because participants could success-fully detect fearful faces but were unable to reliablydiscriminate response correctness, they would be subjec-tively unaware of stimuli that were linked to above-chancedetection performance (i.e., objectively aware).

Historically, the importance of subjective measure ofawareness has been that it resonates with the idea that anintuitively appealing definition of awareness should bebased on introspective reports of individuals’ inner states

Figure 10. Dissociation between objective and subjective percep-tion measures. Illustration of perception dissociation in which theobjective threshold (OT) is smaller than the subjective threshold(ST). The dissociation zone refers to stimulus durations for whichthe subject would be able to reliably detect a fearful-face targetbut for which the subject could not reliably discriminate betweencorrect and incorrect responses. The arrows around the durationvalues suggest that, when a dissociation occurs, the values willvary for different subjects. In addition, the arrows call attention tothe fact that the OTand the ST were not determined in the presentexperiment given that a fixed set of durations was employed. Thesituation depicted in the graph could represent, for example,objective and subjective perception for Subject S4 in Experiment 1(see Figure 4 and Table 2). A possible interpretation of our resultsis that a dissociation zone reveals that the subject would besubjectively unaware but objectively aware of the stimulus.However, other interpretations that do not imply awarenessdissociations are also possible (see text). Abbreviations: OU =objectively unaware; SU = subjectively unaware.



(James, 1890). As argued by Kunimoto et al. (2001), thepresent measure of subjective perception would bephenomenologically valid because it depends cruciallyon the subjects’ own introspective assessment of theirperformance. According to this view, the subjects’ con-fidence cannot reflect accuracy unless they are partiallyaware of the information on which they based theirdiscriminative responses, that is, Bhigh[ vs. Blow[confidence (Kunimoto et al., 2001). At the same time,under conditions in which subjects have no awareness ofthe stimuli, their confidence ratings should be unrelated toaccuracy.

It is well known that when awareness is probed via verbalreports, both detection- and discrimination-type tasks areheavily influenced by individual strategies (Eriksen,1960). It is not surprising, therefore, that many previousstudies of awareness based on verbal reports have beenstrongly criticized in the past (Eriksen, 1960; Holender,1986). In this article, we sought to analyze subjectiveperception in a manner that would be less susceptible toindividual strategies than verbal reports. Our measure wasbased on early proposals of discrimination betweencorrect and incorrect responses, which was originallynamed the Type 2 task (Clarke, Birdsall, & Tanner, 1959;Pollack & Decker, 1958), with the corresponding Type 2SDT analysis. Such approach allowed us to define aperception measure that can be analyzed within the SDTframework, thereby reducing the response bias problem.

Subjective measures likely reflect weakly consciousprocesses

Although the notion of a subjective threshold has beenpersuasive to many (Ohman, 1999; Merikle et al., 2001),others have argued that it may be problematic (e.g.,Snodgrass & Shevrin, 2006). In particular, the SDT critiquesuggests that, in general, subjective threshold effects can becaused by weakly conscious perception (although belowthe subjective criterion). Strictly speaking, the SDT frame-work is Bagnostic[ regarding the awareness questionbecause consciousness (or lack thereof) has no role in thetheory. Accordingly, in the present context, our findingsshould not be considered as strong evidence for unawareprocessing because a subjective approach is inherentlyproblematic.

Objective and subjective measures index qualitativelydifferent processes

In a sense, the interpretation of the present findings islargely dependent on the overall conceptual framework ofthe study. On the one hand, the objective SDT approachprovides a parsimonious account of a large body of datawithout resorting to a Bthreshold[ that provides a dividingline between consciousness and its absence (Macmillan &Creelman, 2005). Instead, the response criterion describeshow subjects trade off sensitivity and specificity during a

specific task (including task instructions and Binherent[biases). On the other hand, the subjective thresholdapproach argues that measures that index participants’Bintrospective[ states are required if one wishes toinvestigate consciousness.

In general, we propose that, instead of deciding whichapproach is Bsuperior,[ a more fruitful approach is tomeasure both objective and subjective measures of percep-tion in trying to understand the nature of visual awareness.Indeed, regardless of the ultimate status of how objectiveand subjective measures map to awareness, we argue thatboth are important because they can index two distinctprocesses. In the context of the present findings, we suggestthat the observed dissociation may reveal a qualitativedissociation in behavior.

In addition, objective awareness may constitute a moreelementary form of awareness than subjective awareness(Pessoa, 2005; Snodgrass & Shevrin, 2006). For instance,objective awareness may index phenomenal consciousness(experiential contents), whereas subjective awareness mayindex reflective or access consciousness (metacognitiveprocesses involving the evaluation of phenomenal contents;see also Block, 1995, 2005). On a more experimentalfashion, these two forms of awareness may index processesthat differ with respect to controlVfor instance, objectiveawareness effects may not be amenable to responsestrategies, whereas subjective awareness effects may be(as in attempting to exclude certain types of responses;Snodgrass & Shevrin, 2006; see also Yonelinas, 2001).

In concluding this section, we need to emphasize that theabove discussion concerning a potential dissociation zoneis limited to the specific context of our study. In otherwords, while we have illustrated the dissociation in terms ofthe duration of the target stimulus (Figure 10), in general,several other factors need to be taken into consideration.In particular, stimulus visibility will strongly depend onparameters such as contrast, size, and spatial frequency ofthe target stimulus, in addition to spatial and temporalcharacteristics of the mask stimulus. Furthermore, whilewe treated our face stimuli set as uniform, there may havebeen variability among faces in terms of the expressive-ness of fear that may have affected the results. Thus,further studies are needed to address how these and otherfactors contribute to dissociations such as those illustratedin Figure 10.

Neuroimaging studies

Our study is also relevant in the context of neuroimagingstudies that have sought to probe the neural correlates ofunaware perception (Liddell et al., 2005; Morris, Ohman,& Dolan, 1998; Pessoa et al., 2006; Phillips et al., 2004;Whalen et al., 1998). Some of these studies determinedbrain responses while subjects passively viewed brieflypresented (30 ms), masked fearful faces. Because subjects



reported not seeing the fearful faces upon debriefing, theywere deemed unaware of the fearful faces (at least in asubjective sense). The present study revealed that allparticipants exhibited subjective sensitivity for 33-mstargets when perception was gauged in terms of theirability to discriminate between correct and incorrectresponses (via their response confidence ratings). Inaddition, whereas previous studies did not control fordifferences in individual response strategies, our studyattempted to mitigate the response bias problem byemploying an index of subjective perception that couldbe analyzed within the SDT framework. In fact, webelieve that our study is the first to probe subjectiveperception of masked fearful faces in a manner thatattempts to more directly confront the response biasproblem.

Summary and implications

Our findings demonstrated that when the perception ofminimally visible stimuli is assessed by employing arelatively large number of trials, subjects reliably exhibitobjective sensitivity for fearful-face targets shown asbriefly as 17 ms. Reliable subjective sensitivity is alsoobtained for 33-ms targets and, for some subjects, even for17-ms targets. Our paradigm revealed a potential dissoci-ation of fear perception according to objective andsubjective criteria. By characterizing perception via ROCanalyses, we showed that such dissociation can be assessedfor each individual participant. The determination of adissociation zone offers the possibility of investigating thephysiological correlates (e.g., fMRI responses) of suchparameter range, which may offer new insights into theneural bases of aware and unaware perception.

Appendix A: SDT for the analysisof objective and subjectiveperception: Parametric ROCapproach

For comparison, we also analyzed our results in terms ofa parametric SDT approach (Figures 5B and 5D), whichassumed normal distributions for both signal and noise(Green & Swets, 1966; Macmillan & Creelman, 2005).Participants were deemed above chance when dV-typemeasures were significantly greater than 0. The p value forstatistical significance was .05.

For objective awareness, we employed the standard dVthat assesses the distance between the hit rate and the falsealarm rate in standard-deviation units. The variance of dVwas approximated with the Gourevitch and Galanter(1967) formula.

Subjective perception was also evaluated in terms of aparametric analysis (Clark et al., 1959; Macmillan &Creelman, 2005). Under the common assumption that theinitial Bfear[/Bno fear[ decision is unbiased, we canpartition the decision axis in such as way that the subjectwould respond Bhigh confidence[ if x e jk or x 9 k andrespond Blow confidence[ if jk e x G k, where kdetermines the cutoff points (Macmillan & Creelman,2005). The subjective dV index and the criterion value kare then determined by solving the following equations forthe conditional probabilities of Bhits[ (pBhit[) and Bfalsealarms[ (pBfalse alarm[):

p‘‘hit’’ ¼ p kxk 9 kkx 9 0ð Þ ¼ p x 9 kð Þp x 9 0ð Þ ¼

�ðdV=2jkÞ� dV=2ð Þ

p‘‘false alarm’’ ¼ p jxj 9 kkx G 0ð Þ¼ p x Gjkð Þp x G 0ð Þ ¼ �ðjdV=2jkÞ

� jdV=2ð Þ ;

ðA1Þ

where �(z) is the normal cumulative distribution functionevaluated at the z score (Macmillan & Creelman, 2005).As discussed in the Introduction, the definition of Bhits[and Bfalse alarms[ in the context of subjective perceptionfollowed the partitioning of Table 1.

The variance of dsubjectiveV was derived by employing aTaylor expansion with two variables, similar to deriva-tions of the same–difference discrimination method (Bi,2002):

var dVð Þ ¼varðp‘‘hit’’Þ

b21

a1

b1j a2

b2

� �2þ

varðp‘‘false alarm’’Þ

b22

a1

b1j a2

b2

� �2; ðA2Þ

where a1 =@p‘‘hit’’@dV

��k0;d0V

, a2 =@p‘‘false alarm’’

@dV

��k0;d0V

, b1 =@p‘‘hit’’@k

��k0;d0V

,

and b2 =@p‘‘false alarm’’

@k

��k0;d0V

are the partial derivatives with

respect to the dVindex and the cutoff k.

Acknowledgments

We thank the National Institutes of Mental Health for itssupport for this work (Award 1R01 MH071589 to L.P.). Wealso thank the reviewers for a number of suggestions thathave improved this article.

Commercial relationships: none.Corresponding author: Luiz Pessoa.Email: [email protected]: Department of Psychological and Brain Sciences,Indiana University, 1101 East Tenth Street, Bloomington,IN 47405, USA.



References

Adolphs, R., Gosselin, F., Buchanan, T. W., Tranel, D.,Schyns, P., & Damasio, A. R. (2005). A mechanismfor impaired fear recognition after amygdala damage,Nature, 433, 68–72. [PubMed]

Bi, J. (2002). Variance of dV for the same–differentmethod. Behavior Research Methods, Instruments,and Computers, 34, 37–45. [PubMed]

Block, N. (1995). On a confusion about a function ofconsciousness. Behavioral and Brain Sciences, 18,227–247. [Article]

Block, N. (2005). Two neural correlates of consciousness.Trends in Cognitive Sciences, 9, 46–52. [PubMed]

Bowers, R. M. (1984). On being unconsciously influencedand informed. In K. S. Bowers & D. Meichelbaum(Eds.), The unconscious reconsidered (pp. 227–272).New York: John Wiley & Sons.

Clarke, F. R., Birdsall, T. G., & Tanner, W. P., Jr. (1959).Two types of ROC curves and definition of parame-ters. Journal of the Acoustical Society of America, 31,629–630.

DeLong, E. R., DeLong, D. M., & Clarke-Pearson, D. L.(1988). Comparing the areas under two or morecorrelated receiver operating characteristic curves: Anonparametric approach. Biometrics, 44, 837–845.[PubMed]

Dolan, R. (2002). Emotion, cognition, and behavior.Science, 298, 1191–1194. [PubMed]

Efron, B., & Tibshirani, R. J. (1993). An introduction tothe bootstrap. New York: Chapman & Hall.

Egan, J. P. (1975). Signal detection theory and ROCanalysis. New York: Academic Press.

Ekman, P., & Friesen, W. V. (1976). Pictures of facialaffect. Palo Alto, CA: Consulting Psychologists Press.

Eriksen, C. W. (1960). Discrimination and learningwithout awareness: A methodological survey andevaluation. Psychological Review, 67, 279–300.[PubMed]

Esteves, F., & Ohman, A. (1993). Masking the face:Recognition of emotional facial expressions as afunction of the parameters of backward masking.Scandinavian Journal of Psychology, 34, 1–18.[PubMed]

Galvin, S. J., Podd, J. V., Drga, V., & Whitmore, J.(2003). Type 2 tasks in the theory of signaldetectability: Discrimination between correct andincorrect decisions. Psychonomic Bulletin & Review,10, 843–876. [PubMed]

Gold, J. M., Murray, R. F., Sekuler, A. B., Bennett, P. J., &Sekuler, R. (2005). Visual memory decay is determin-istic. Psychological Science, 16, 769–774. [PubMed]

Gold, J. M., Sekuler, A. B., & Bennett, P. J. (2004).Characterizing perceptual learning with externalnoise. Cognitive Science, 28, 167–207.

Gourevitch, V., & Galanter, E. (1967). A significance testfor one parameter isosensitivity functions. Psychome-trika, 32, 25–33. [PubMed]

Green, D. M., & Swets, J. A. (1966). Signal detectiontheory and psychophysics. New York: Wiley.

Hajian-Tilaki, K. O., Hanley, J. A., Joseph, L., & Collet, J. P.(1997). A comparison of parametric and nonparametricapproaches to ROC analysis of quantitative diagnostictests. Medical Decision Making, 17, 94–102.[PubMed]

Hanley, J. A., & McNeil, B. J. (1982). The meaning anduse of the area under a receiver operating character-istic (ROC) curve. Radiology, 143, 29–36.[PubMed]

Holender, D. (1986). Semantic activation without con-scious awareness in dichotic listening, parafovealvision, and visual masking: A survey and appraisal.Behavioral and Brain Sciences, 9, 1–23.

Ishai, A., Pessoa, L., Bikle, P. C., & Ungerleider, L. G.(2004). Repetition suppression of faces is modu-lated by emotion. Proceedings of the NationalAcademy of Sciences of the United States of America,101, 9827–9832. [PubMed] [Article]

James, W. (1890). Principles of psychology. New York:Holt.

Kolb, F. C., & Braun, J. (1995). Blindsight in normalobservers. Nature, 377, 336–338. [PubMed]

Kunimoto, C., Miller, J., & Pashler, H. (2001). Confidenceand accuracy of near-threshold discrimination re-sponses. Consciousness and Cognition, 10, 294–340.[PubMed]

Liddell, B. J., Brown, K. J., Kemp, A. H., Barton, M. J.,Das, P., Peduto, A., et al. (2005). A direct brainstem–amygdala–cortical Falarm_ system for subliminalsignals of fear. Neuroimage, 24, 235–243. [PubMed]

Macmillan, N. A. (1986). The psychophysics of sublim-inal perception. Behavioral and Brain Sciences, 9,38–39.

Macmillan, N. A., & Creelman, C. D. (2005). Detectiontheory: A user’s guide (2nd ed.). New York: Cam-bridge University Press.

Maxwell, J. S., & Davidson, R. J. (2004). Unequallymasked: Indexing differences in the perceptual sali-ence of Bunseen[ facial expressions. Cognition andEmotion, 18, 1009–1026.

Merikle, P. M., Smilek, D., & Eastwood, J. D. (2001).Perception without awareness: Perspectives fromcognitive psychology. Cognition, 79, 115–134.[PubMed]



http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=pubmed&cmd=Retrieve&dopt=AbstractPlus&list_uids=15635411&query_hl=90&itool=pubmed_DocSum


http://cogprints.org/231/00/199712004.html



http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=pubmed&cmd=Retrieve&dopt=AbstractPlus&list_uids=12424363&query_hl=6&itool=pubmed_docsum









http://www.pubmedcentral.nih.gov/articlerender.fcgi?tool=pubmed&pubmedid=15210952





Morris, J. S., Ohman, A., & Dolan, R. J. (1998). Consciousand unconscious emotional learning in the humanamygdala. Nature, 393, 467–470. [PubMed]

Obuchowski, N. A. (1994). Computing sample size forreceiver operating characteristic studies. InvestigativeRadiology, 29, 238–243. [PubMed]

Ohman, A. (1999). Distinguishing unconscious from con-scious emotional processes: Methodological consider-ations and theoretical implications. In T. Dalgleish &M. Power (Eds.), Handbook of cognition and emotion(pp. 321–352). Chichester: John Wiley & Sons.

Ohman, A. (2002). Automaticity and the amygdala:Nonconscious responses to emotional faces. CurrentDirections in Psychological Science, 11, 62–66.

Pessoa, L. (2005). To what extent are emotional visualstimuli processed without attention and awareness?Current Opinion in Neurobiology, 15, 188–196.[PubMed]

Pessoa, L., Japee, S., Sturman, D., & Ungerleider, L. G.(2006). Target visibility and visual awareness modu-late amygdala responses to fearful faces. CerebralCortex, 16, 366–375. [PubMed] [Article]

Pessoa, L., Japee, S., & Ungerleider, L. G. (2005). Visualawareness and the detection of fearful faces. Emotion,5, 243–247. [PubMed]

Phillips, M. L., Williams, L. M., Heining, M., Herba,C. M., Russell, T., Andrew, C., et al. (2004). Differ-ential neural responses to overt and covert presenta-tions of facial expressions of fear and disgust.Neuroimage, 21, 1484–1496. [PubMed]

Pollack, I., & Decker, L. R. (1958). Confidence ratings,message reception, and the receiver operating char-acteristic. Journal of the Acoustical Society ofAmerica, 30, 286–292.

Pollack, I., & Hsieh, R. (1969). Sampling variability of thearea under the ROC-curve and of dVe. PsychologicalBulletin, 71, 161–173.

Reingold, E. M., & Merikle, P. M. (1988). Using directand indirect measures to study perception without

awareness. Perception & Psychophysics, 44, 563–575.[PubMed]

Schyns, P. G., Bonnar, L., & Gosselin, F. (2002). Showme the features! Understanding recognition from theuse of visual information. Psychological Science, 13,402–409. [PubMed]

Schyns, P. G., & Oliva, A. (1999). Dr. Angry and Mr.Smile: When categorization flexibly modifies theperception of faces in rapid visual presentations.Cognition, 69, 243–265. [PubMed]

Snodgrass, M. (2004). The dissociation paradigm and itsdiscontents: How can unconscious perception ormemory be inferred? Consciousness and Cognition,13, 107–116. [PubMed]

Snodgrass, M., Bernat, E., & Shevrin, H. (2004). Uncon-scious perception: A model-based approach tomethod and evidence. Perception & Psychophysics,66, 846–867. [PubMed]

Snodgrass, M., & Shevrin, H. (2006). Unconsciousinhibition and facilitation at the objective detectionthreshold: Replicable and qualitatively different un-conscious perceptual effects. Cognition, 101, 43–79.[PubMed]

Spezio, M. L., Adolphs, R., Hurley, R. S., & Piven, J.(2007). Analysis of face gaze in autism usingBBubbles[. Neuropsychologia, 45, 144–151.[PubMed]

Whalen, P. J., Rauch, S. L., Etcoff, N. L., McInerney,S. C., Lee, M. B., & Jenike, M. A. (1998). Maskedpresentations of emotional facial expressions modu-late amygdala activity without explicit knowledge.Journal of Neuroscience, 18, 411–418. [PubMed][Article]

Yonelinas, A. P. (2001). Consciousness, control, andconfidence: The 3 Cs of recognition memory. Journalof Experimental Psychology: General, 130, 361–379.[PubMed]







http://cercor.oxfordjournals.org/cgi/content/full/16/3/366











http://www.jneurosci.org/cgi/content/full/18/1/411


Fear perception: Can objective and subjective awareness ...lce.umd.edu/publications_files/fear...

Documents

Transcript of Fear perception: Can objective and subjective awareness ...lce.umd.edu/publications_files/fear...