AParametric Study of Prefrontal Cortex Involvementin Human Working Memory

TODD S. BRAVER,* JONATHAN D. COHEN*, LEIGH E. NYSTROM,* JOHN JONIDES,EDWARD E. SMITH, AND DOUGLAS C. NOLL

*Department of Psychology, Carnegie Mellon University, Pittsburgh, Pennsylvania 15213;Departments of Psychiatry and Radiology, University of Pittsburgh, Pittsburgh, Pennsylvania 15260;

and Department of Psychology, University of Michigan, Ann Arbor, Michigan 48109

Received July 11, 1996

Although recent neuroimaging studies suggest thatprefrontal cortex (PFC) is involved inworkingmemory(WM), the relationship between PFC activity andmemory load has not yet been well-described in hu-mans. Here we use functional magnetic resonanceimaging (fMRI) to probe PFC activity during a sequen-tial letter task in which memory load was varied in anincremental fashion. In all nine subjects studied, dorso-lateral and left inferior regions of PFC were identifiedthat exhibited a linear relationship between activityand WM load. Furthermore, these same regions wereindependently identified through direct correlationsof the fMRI signal with a behavioral measure thatindexes WM function during task performance. A sec-ond experiment, using whole-brain imaging tech-niques, both replicated these findings and identifiedadditional brain regions showing a linear relationshipwith load, suggesting a distributed circuit that partici-pates with PFC in subserving WM. Taken together,these results provide a doseresponse curve describ-ing the involvement of both PFC and related brainregions in WM function, and highlight the benefits ofusing graded, parametric designs in neuroimagingresearch. r 1997Academic Press

Cognitive theorists have posited that higher functionssuch as language, planning, and problem-solving allrely on working memory (WM), which acts to tempo-rarily maintain and manipulate task-relevant informa-tion (Baddeley, 1986; Just and Carpenter, 1992; Shal-lice, 1988). Neurophysiological data from nonhumanprimates have suggested that prefrontal cortex (PFC)is an important component of the neural substrate forWM. In single-cell recording studies, neurons havebeen identified in dorsolateral PFC that remain activeduring delay periods in tasks that require the subject toactively maintain an internal representation of a targetstimulus (Fuster, 1989; Goldman-Rakic, 1987). In hu-mans, findings of WM impairment in patients with

damage to the frontal cortex support the involvement ofthis brain region in WM function (Shimamura, 1994;Stuss et al., 1994). Recently, neuroimaging data havealso begun to contribute to our understanding of theneurobiology of human working memory. Studies usingboth positron emission tomography (PET) and func-tional magnetic resonance imaging (fMRI) have pro-vided strong evidence of PFC involvement in a widevariety of working memory tasks (Cohen et al., 1994;Courtney et al., 1996; Fiez et al., 1996; Jonides et al.,1993; McCarthy et al., 1994; Petrides et al., 1993;Swartz et al., 1995). Further, activation of additionalbrain regions has often been observed during perfor-mance of these tasks (e.g., posterior parietal cortex,Brocas area), suggesting the possibility that WM func-tion is subserved by multiple brain regions.These neuroimaging studies have all identified WM-

related brain activity through the use of the subtractivemethod. According to this method, activity generatedby a control task is subtracted from that observed inan activation task. The control task is designed toengage all of the processes of the activation task exceptfor the cognitive process of interest (Posner et al., 1988).Thus, subtraction of the control images from the activa-tion images should reveal activity that is associatedwith the process of interest. Although the findings fromneuroimaging studies of WM have been reasonablyconsistent and informative there are, nevertheless, anumber of well-recognized limitations to the subtrac-tive approach (McClelland, 1979; Sternberg, 1969). Themost important of these is the assumption that theirrelevant processesthose common to both tasksand supposed to be subtracted outare performed inthe same way in the two tasks. In neuroimagingstudies, this also assumes that the irrelevant activityoccurs in the same anatomic regions across both tasks.Since it is impossible to verify this assumption directly,misleading interpretations can ensue regarding theobserved patterns of activation. For example, in all of

NEUROIMAGE 5, 4962 (1997)ARTICLE NO. NI960247

49 1053-8119/97 $25.00Copyright r 1997 by Academic Press

All rights of reproduction in any form reserved.

the neuroimaging studies of WM mentioned above, theincreased demand on processing resources in the activa-tion condition may cause subjects to adopt new strate-gies, and may even change the nature of task processesthought to be unrelated to WM, such as encoding andresponse execution. If these changes in task processinglead to increased activity, or to changes in the anatomiclocation of activated regions, they may be misinter-preted as reflecting structures involved in WM.In behavioral studies, cognitive scientists frequently

deal with the limitations of the subtractive approach byvarying the process of interest in a parametric, ratherthan a discrete, fashion. Thus, tasks are designed toproduce a specific pattern of response in the process ofinterest. While other processes may still vary acrossconditions, the likelihood is lower that they will showthe same pattern of response as is predicted for theprocess of interest. In other words, the parametricapproach places tighter constraints on the pattern ofresponses predicted, and thus provides greater selectiv-ity in measuring processes of interest. Sternbergs(1969) studies of short-term memory are a classicexample of a parametric experimental design. In theseexperiments, the number of items kept active inmemorywas varied incrementally. Changes observed in reac-tion time were then used to selectively probe andcharacterize the cognitive processes involved in re-trieval.Parametric designs can also be used in neuroimaging

studies, to more selectively identify regions of brainactivity associated with cognitive processes of interest.This approach has already begun to see application inthe study of basic sensory processes (Engel et al., 1994;Schneider et al., 1994; Tootell et al., 1995) and cognitiveprocesses such as long-term memory (Grasby et al.,1994). Here we present the results of two experimentsin which we exploit the advantages of parametricdesigns to characterize the relationship between brainactivation and WM function. In both, fMRI was used toprobe brain activity during performance of a sequentialletter memory task in which WM load was varied in agraded fashion. The speed and noninvasiveness offMRI allowed us to obtain large numbers of imagesfrom multiple task conditions within a single imagingsession, necessary for a parametric design. In the firstexperiment, we examined this relationship specificallywithin PFC. In the second, we used whole-brain imag-ing both to confirm our results and to identify addi-tional, nonfrontal regions that may contribute to WMfunction.

EXPERIMENT 1

Our initial study focused on PFC, since it has been amajor focus of neurophysiological studies of WM. Ouranalysis procedure identified PFC regions whose activ-

ity varied reliably with changes in memory load. Addi-tionally, behavioral data were analyzed to directlyvalidate findings concerning the relationship betweenPFC activity and WM function. Specifically, reactiontime (RT) was used as an empirically defined measureof load, providing a convergent method for identifyingregions of brain activity associated with WM function.

Methods

Subjects

Informed consent was obtained from nine neurologi-cally normal right-handed subjects (eight males, onefemale; 1831 years of age). All subjects were given apretesting session, in which they practiced the task,and were included only after reaching a criterion levelof performance (.75% accuracy on all conditions).

Cognitive Tasks

Subjects performed a variant of a sequential lettertask that has been used previously in a number ofneuroimaging experiments to study working memory(Awh et al., 1996; Cohen et al., 1994; Gevins andCutillo, 1993; Smith et al., 1996). However, theseprevious experiments used only two task condi-tionsWM and control. In the current study, fourconditions were used that varied WM load incremen-tally from zero to three items (see Fig. 1).Subjects observed stimuli on a visual display and

responded using a hand-held response box with fiber-optic connections to a Macintosh computer in thecontrol room running PsyScope software (Cohen et al.,1993). In the 0-back condition, subjects responded to asingle prespecified target letter (e.g., X). In the 1-backcondition, the target was any letter identical to the oneimmediately preceding it (i.e., one trial back). In the2-back and 3-back conditions, the target was any letterthat was identical to the one presented two or threetrials back, respectively. Thus, WM load increasedincrementally from the 0-back to the 3-back condition.Stimuli were pseudorandom sequences of consonants(randomly varying in case), presented centrally (500-msduration, 2500-ms interstimulus interval). Subjectsresponded to each stimulus with their dominant hand,pressing one button for targets (33% of trials) andanother for nontargets. In each sequence, a number ofstimuli were nontarget repeats that were included asfoils (e.g., 2-back repeats occurring in the 3-back condi-tion). Conditions were run in blocks of approximately75 s (25 stimuli), during which scans were acquired.Ten repetitions of each of the four conditions were run,with condition order randomized across blocks andsubjects (two subjects received only five repetitions,with no difference in results), subject to the constraintthat all four conditions were sampled in every set offour blocks. A short delay between blocks (1020 s) was

50 BRAVER ET AL.

provided as a rest break for subjects and to allow thehemodynamic response to recover from the previousblock.

MRI Scanning Procedures

Images were acquired with a conventional 1.5-T GESigna whole-body scanner. Two 5-in. surface coils,mounted in parallel over the frontal area, were used tomaximize signal-to-noise ratio in the region of interest.Subject movement was constrained with a bite-barconsisting of an individually molded dental impressionattached to a rigid plastic frame. All scans were 7 mmthick, with 1.88 3 1.88-mm pixels in plane, acquired ateight contiguous locations, perpendicular to theACPCline, and extending rostrally from the anterior commis-sure (see Fig. 2a). Structural images were obtained atthese locations prior to functional scanning, using astandard T1-weighted sequence. Functional imageswere acquired using a 10-interleave spiral scan pulsesequence (Noll et al., 1995) with a TR of 640 ms, agradient echo TE of 35ms, a flip angle of 45, and a fieldof view of 24 cm. A set of images from all eight slicelocations was obtained every 6.4 s. Eleven sets wereobtained during each 75-s task block.

Image Analysis Procedures

Two analyses were performed, one using memoryload and the other reaction time as independent vari-ables. Both analyses were conducted according to thefollowing set of procedures. All functional images werefirst scaled to a common mean (to reduce the effect ofscanner drift or instability), and then registered in-plane using an automated algorithm (Woods et al.,1992). Statistical maps of activity were generated foreach slice (see below for details) and then thresholdedfor significance using a cluster-size algorithm thattakes account of the spatial extent of activation in

correcting for multiple comparisons (Forman et al.,1995). A contiguity threshold of seven voxels and per-pixel false-positive rate of 0.03were chosen, correspond-ing to a corrected pixel-wise a protection of 0.0006.Potential venous artifact was removed using a mag-netic resonance angiography filtering procedure (Co-hen et al., 1995).In the memory-load analysis, regions of memory-

related activity were identified using a voxel-wiseANOVA, with memory load (4 levels) as the experimen-tal factor and block repetition (5 or 10 levels) treated asa random factor. Treating block as a random variableallows the variance associated with spurious transientsin the signal (block 3 load interactions) to be separatedout from the task-related effects of interest (main effectof load). Statistical maps were generated for each sliceusing the F ratio obtained for the main effect of memoryload at each voxel. A cross-subject comparison was thenperformed, in order to identify foci of activity that werereliable across subjects. To do so, the statistical activa-tion maps for each subject were first standardized intoTalairach stereotactic coordinates (Talairach and Tour-noux, 1988) by rescaling images relative to the posi-tions of the anterior and posterior commissure and theedges of the brain (determined manually on a regis-tered anatomical volume image) using AFNI software(Cox, 1996). The number of subjects showing signifi-cant activation at each voxel of the aligned images wascomputed for the data set after blurring the imageswith a 5-mm FWHM Gaussian filter (to accommodatebetween-subject anatomical variability). As a conserva-tive criterion, only regions that showed significantactivation in all nine subjects were selected for moredetailed analyses. In the reaction time analysis, re-gions of behaviorally associated activity were identifiedthrough voxel-wise correlations of the mean activitywith the subjects mean RT (correct responses only) for

FIG. 1. Adiagram of the four memory conditions of the sequential letter task.

51PREFRONTAL CORTEX AND WORKING MEMORY

52 BRAVER ET AL.

each block. Statistical maps were generated for eachslice using the correlation coefficient at each voxel.

Results

Memory-Load Analysis

The cross-subject comparison revealed several re-gions within PFC that showed significant and reliablemain effects of memory load (see Fig. 2b). These werelocalized to the middle frontal gyrus (MFG; Brod-manns area 46/9) bilaterally, the left inferior frontalgyrus (LIFG; Brodmanns area 44/45), another leftinferior site which was more anterior (Brodmanns area47/10), and the anterior cingulate gyrus (Brodmannsarea 32). However, these main effects of load could beproduced by a variety of activation patterns across thefour task conditions. In this study, we were particularlyinterested in testing whether any PFC regions showedprogressively increasing activity as a function of load.For this reason, we performed a planned contrast onthe above four regions that tested whether signalintensity showed a pattern of monotonic increase acrossthe four load conditions (Braver and Sheets, 1993). Thiscontrast revealed that of the four regions showing amain effect of load, only MFG and LIFG showed thismonotonic pattern.The PFC response to changes in WM load within the

MFG and LIFG regions was examined in greater detailby plotting the fMRI signal as a function of load acrossthe four task conditions (Fig. 3). This revealed a linear

relationship between the signal and load (for the lineartrend, P , 0.001 for both regions; quadratic and cubictrends, NS). From this plot it was possible to determinethe average change in activity per item of load in theseregions. In MFG, fMRI signal intensity increased at arate of 0.34% per item, while in LIFG the increase was0.47%. Although effects of this magnitude may be toosmall to detect reliably through individual compari-sons, the parametric design of the study made itpossible to determine that the fMRI signal can, in fact,consistently track graded changes in cognitive load.

Reaction Time Analysis

As noted earlier, it has long been recognized that RTis sensitive to manipulations of WM load (Baddeley,1986; Just and Carpenter, 1992; Sternberg, 1969).Thus, RT can be used as an independent and empiri-cally defined measure of the effect of load that isuniquely derived for each subject. Significant positivecorrelations between the blockwise RT and fMRI signalwere observed in all nine subjects (average r 5 0.49,P , 0.01), but were most consistent only in the samelocations within PFC (MFG and LIFG) as were previ-ously shown to linearly track memory load (see Fig. 4).Thus, correlation with behavioral performance pro-vided a convergent source of evidence that these areasof PFC are related to WM function. Such a correlationmight have been predicted a priori, since RT alsoexhibited an expected linear relationship to load (0-back, 484 6 24 ms; 1-back, 539 6 35 ms; 2-back,628 6 45 ms; 3-back, 701 6 45 ms; linear trend,P , 0.001; quadratic and cubic trends, NS). However, alinear regression with memory load only partiallyaccounted for the variance in both RT (R2 5 0.36) andfMRI signal (R2 5 0.24). Thus, the two measures mighthave represented different, nonoverlapping compo-nents of variance and hence be uncorrelated.

Discussion

The findings from this study are consistent withprevious data suggesting the involvement of PFC inWM. Memory-related activity was observed bilaterallyin a middorsolateral area (MFG) that has been consis-tently implicated in WM tasks in both humans (Cohenet al., 1994; McCarthy et al., 1994; Petrides et al., 1993)and nonhuman primates (Fuster, 1989; Goldman-Rakic, 1987). Additionally, we observed activity in a leftinferior region (LIFG) consistent with Brocas area thatis typically activated in tasks requiring verbal re-hearsal (Frackowiak, 1994; Paulesu et al., 1993; Pe-

FIG. 2. (a) A midsagittal T1-weighted image, indicating the location and plane of functional scans. (b) The results of the memory-loadgroup analysis, with the regions of activation overlaid on the corresponding stereotactically normalized and averaged structural images.Colored pixels designate areas showing a main effect of memory load that was significant in all nine subjects.

FIG. 3. Plots of the MR signal, expressed as a percentage ofdifference from the 0-back condition, across the four load conditions.Values are plotted for the MFG region (Talairach coordinatesL38,30,22 and R35,22,27) and the IFG region (Talairach coordinatesL40,6,26), the only two regions which showed a significant monotoniceffect of load. The activity was averaged across the nine subjects;however, each subject showed essentially the same linear relation-ship to load.

53PREFRONTAL CORTEX AND WORKING MEMORY

tersen et al., 1989). Both the anterior cingulate and theleft anterior area have also been previously observed inWM tasks (Awh et al., 1996; Cohen et al., 1994;Courtney et al., 1996; Jonides et al., 1993; McCarthy etal., 1994), although for the latter region activity wasobserved in the right rather than the left hemisphere.Our results go beyond previous studies, however, by

defining the cognitive analog of a doseresponse curvefor the relationship between WM load and the activityof circumscribed regions of PFC. The identification ofthis relationship to load relied on the use of a paramet-ric design that produced a specific response pattern inboth the fMRI signal and the behavioral performance.This response pattern was observed in both MFG andLIFG as a progressive, linear increase in signal witheach increment of load. Furthermore, the parametricnature of the study made it possible to detect theseincreases in the fMRI signal in these two regions eventhough they were relatively small (,0.5% per item ofload). In contrast, we did not observe this pattern ofprogressive signal increase in the anterior cingulateand the left anterior inferior region. Post hoc inspectionof these areas revealed that activity actually decreasedacross the four load conditions. Although we do nothave a clear explanation of this activity pattern, itnevertheless suggests that there are multiple brainresponses to increasing load.Finally, our ability to independently identify both the

MFG and the LIFG regions through correlations withreaction time provides an important validation of theobserved findings, while avoiding many of the theoreti-cal assumptions implicit in the first analysis (i.e., byusing observed behavior to index WM function ratherthan the predefined task manipulation). Specifically, itdemonstrates that block-by-block fluctuations in PFCactivity closely track fluctuations in observable behav-ior, suggesting that behavioral performance can beused as a psychologically relevant reference function(cf., Bandettini et al., 1993; DeYoe et al., 1994) by whichto isolate activity related to WM. Indeed, this type ofanalysis may apply well to other neuroimaging studiesof higher cognitive processes.

Alternative Explanations

One alternative interpretation of the findings fromneuroimaging studies of WM is that the patterns ofactivity represent the operation of processes associatedwith error detection and compensation, rather thanthose involved in the storage and manipulation ofinformation within WM. The fact that subjects typi-cally make more errors in the more demanding WMconditions of these studies makes this argument aplausible one. Adding further support are recent find-ings from event-related potential studies suggestingchanges in frontal activity immediately following errortrials (Dehaene et al., 1994; Gehring et al., 1993). Thisexplanation could apply to the current study as well,

because even though subjects were generally veryaccurate on the task (.90%), error rates increased withmemory load (F(3,24) 5 12.4, P , 0.001). However, areanalysis of the imaging data using only error-freetrials (by removing all scans occurring within 7 s beforeor after an error) yielded virtually identical results inboth the areas of activity and the linear relationship toload observed. Thus, it is unlikely that the observedpattern of results was due solely to the activity of atransient frontal error-detection process. However, itmust be acknowledged that our reanalysis does not ruleout the possibility that tonic error-monitoring pro-cesses may be engaged in the task by the greaterdifficulty imposed by increasing load.Another alternative interpretation of these findings

is that they reflect the operation of more general, anddiffuse, processes associated with mental effort, ratherthan processes specific to WM. However, the circum-scribed anatomic localization of the identified regionsmakes this argument less plausible. Moreover, wetested the effort hypothesis directly in another recentlycompleted experiment, in which WM and effort wereindependentlymanipulated (Braver et al., 1996). Dorso-lateral PFC activated in memory-demanding condi-tions relative to control conditions matched for diffi-culty and, furthermore, did not activate when difficultywas manipulated independently of WM demands (bydegrading the stimuli). Thus, activation of this struc-ture seems to reflect processes specific to WM, and notprocesses associated with more general forms of mentaleffort.

Intersubject Variability

In this study, our primary focus was on regionswithin PFC in which activity most reliably trackedchanges in memory load. For this reason, we performedan analysis that selected only those voxels showing asignificant effect of load in all of the subjects studied.Although this criterion ensures high intersubject aswell as intrasubject reliability, it has the effect ofdetecting only the areas of maximum overlap in thePFC regions activated across subjects. Thus, it mightbe argued that our results present a misleading pictureof the focality of the PFC regions involved in this task,by neglecting: (1) the full extent of the identifiedregions (i.e., MFG and LIFG) beyond their areas ofhighest overlap and (2) additional PFC regions thatwere activated in a subset, but not all of the subjects(and thus fell below our selection criterion). Indeed, theregions of MFG and LIFG identified by our analysisprocedure do not fully reflect the substantial amount ofintersubject variability in the location and size of theactivated regions (e.g., compare Fig. 2b with Fig. 4).Furthermore, when the selection criterion was loweredfrom the nine-subject threshold, additional regions ofactivity were revealed. These included bilateral frontaloperculum (insular cortex; seven subjects), bilateral

54 BRAVER ET AL.

superior frontal gyrus (BA 6 and 8; seven subjects),frontopolar cortex (BA 10; six subjects), and a moreposterior region of the anterior cingulate (BA 32; fivesubjects).Understanding and quantifying these sources of inter-

subject variability are important issues in neuroimag-ing and are currently areas of active research (Herholzet al., 1996; OLeary et al., 1996; Wessinger and Gazza-niga, 1996). The high spatial resolution and noninva-siveness of fMRI relative to othermethods have broughtthis issue into sharp focus. With fMRI it is possible tolocalize areas of activity reliably to specific anatomicalstructures (e.g., cortical gyri) within individual sub-jects. As a consequence, both intra- and intersubjectvariability in location and size of activated regions canbe more easily observed. These technological advancesare of great benefit to studies specifically focusing onindividual differences or the idiosyncrasies of func-tional neural systems, as well as to studies utilizingsingle-subject and longitudinal designs. These types ofefforts are directly analogous to ones in behavioralresearch, where distinct subfields and specializedmeth-ods have developed to study issues surrounding intra-and intersubject variability (e.g., psychophysics andindividual differences studies). However, the goal ofmost behavioral and neuroimaging studies is to func-tionally characterize the commonalities across a groupof subjects, so that inferences can reliably be drawnregarding the characteristics of a population. The cur-rent study falls into this latter category, in that our goalwas to identify the PFC regions activated in commonacross the subject group, in order to identify thoseneural components that are fundamental to WM func-tion in the normal adult population.In neuroimaging studies with goals such as ours, the

issue of intersubject variability poses a different prob-lem, in that no standard method has emerged forreliably comparing activity across subjects. Between-subject differences in both structural anatomy andexperimental effect size make it difficult to determinewhen a specific brain region has been reliably affectedby task manipulations. Thus, a number of differentapproaches have been used across laboratories, and weourselves have tried different approaches in previouswork (Cohen et al., 1994; Schneider et al., 1993). Wechose the analysis procedure used in the current studyin order to select only those regions that reflected themost statistically reliable tendencies of the data. Unfor-tunately, with this procedure, there did not appear to beany nonarbitrary way of determining the appropriatelevel of Type I/Type II balance. Thus, we opted toincrease our Type I protection as much as possible inorder to be highly confident of the significance in anyactivated regions. However, our approach may havebeen too conservative, and may not have had sufficientpower to detect subtle but real effects.A more principled solution to this problem might be

found by comparing the situation with that found inbehavioral studies, where similar issues have alreadybeen addressed. When population-wide inferences areto be drawn in a behavioral experiment, the analysisprocedure treats each subject studied as a randomvariable. Because of this conceptualization, an investi-gator will not typically expect to find that every subjectin the sample shows the hypothesized effect. Instead,what is sought is statistical confidence that the effect ispresent as a central tendency of the sample, such thatstatistical inferences can be drawn regarding the popu-lation of interest. Likewise, in neuroimaging studies,the expectation that every subject show an effect (as wehad in the current study) is likely to be too demandingand unrealistic, especially when the effect size is small(or variability is high).However, in behavioral studies, comparing data across

subjects is much simpler than in neuroimaging, sincethe dependent variables of interest are usually well-defined (e.g., accuracy and RT) and easily comparableacross subjects. As mentioned above, comparing activ-ity in specific brain regions across subjects is compli-cated by the variability in both the size and thethree-dimensional geometry of individual brains. Themost common approach to this problem is to account forthis variability by normalizing brains into a standard-ized stereotactic space and using a smoothing proce-dure to reduce remaining anatomical differences be-tween subjects. Following this normalization procedure,neuroimaging data can be pooled across subjects inexactly the same fashion as in behavioral studies: treateach subject as a random factor in the design andexamine each brain region individually to identify anystatistically reliable effects due to experimentalmanipu-lations. This cross-subject pooling procedure is alreadywidely used for PET studies, but has been less com-monly adopted in those using fMRI. Although theprocedure has its own associated problems (neitherappropriate normalization techniques nor the optimalstereotactic space have been universally agreed upon;e.g., Dale and Sereno, 1993; Drury et al., 1996; Fristonet al., 1991; Woods et al., 1993), it does seem to be thebest current method for formally determining the reli-ability and magnitude of experimental effects across agroup of subjects, and for comparing effects betweengroups. For this reason, we adopted such an approachin Experiment 2, described below.

Spatial Distribution of WM

Our finding that activity in PFC is correlated withWM load provides a strong demonstration of the involve-ment of this brain region in human WM function.Moreover, these results are consistent with the neuro-physiological literature, which has directly demon-strated in nonhuman primates that sustained neuronal

55PREFRONTAL CORTEX AND WORKING MEMORY

FIG. 4. Acoronal slice plane (slice 4, approximately 24 mm anterior to theAC) in each of the nine subjects, showing regions in dorsolateralPFC (MFG) that were correlated with the subjects blockwise mean RT (color scale indicates degree of significance). Correlated activity wasalso reliably found in IFG (not shown).

FIG. 5. Axial slices showing the results of Experiment 2, with regions showing significant load sensitivity (as determined through afocused contrast) overlaid on a stereotactically normalized reference brain. Note the replications of activity in bilateral PFC (135 mm) andBrocas area (116 mm), as well as the large bilateral region of activated parietal cortex (150 mm, 135 mm). Tailarach coordinates of allsignificant regions of activation are given in Table 1.

56 BRAVER ET AL.

activity in dorsolateral PFC is crucial for the short-term maintenance of task-relevant information (Fus-ter, 1989; Goldman-Rakic, 1987). However, by limitingour focus to PFC, the current study provides no informa-tion about whether other brain regions also show loadsensitivity. A number of lines of evidence suggest that,in fact, this type of response might also be found inother cortical and subcortical structures. Neurophysi-ological studies have shown sustained neural activityin premotor, posterior parietal, and inferior temporalcortex as well as in subcortical structures, such as thebasal ganglia and thalamus (Fuster and Alexander,1973; Fuster and Jervey, 1981; Gnadt and Anderson,1988; Schultz and Romo, 1988; Weinrich and Wise,1982). Neuropsychological and neuroimaging data alsolend support to the idea that WM is subserved bymultiple neural structures. WM impairments havebeen observed following damage to both parietal cortexand the basal ganglia (Sagar et al., 1988; Vallar andShallice, 1990), and whole-brain neuroimaging studiestypically activate a constellation of areas during perfor-mance of WM tasks (Courtney et al., 1996; Smith et al.,1996; Swartz et al., 1995). If information is maintainedin a widely distributed manner, or through the bidirec-tional flow of activity among these various structures,then one might expect similar patterns of load sensitiv-ity among them.Cognitive theories of WM provide an additional

reason to predict a brain-wide distribution of activatedareas during performance of the task we have studied.The dominant view of WM in cognitive psychology is ofa multicomponent system involving two types of pro-cesses: executive, which are subserved by a centralcontroller, and maintenance, which are subserved bymodality-specific slave buffers (Baddeley, 1986). Accord-ingly, experimental tasks designed to engage WM ofteninvolve the manipulation of as well as short-termstorage of information. The sequential letter task usedin the current study is a good example of this. Inaddition to maintaining the identity of each stimulus, itis also necessary to assign the stimulus a specifictemporal order (e.g., remember that B was the letter2-back in this trial, while C was the letter 1-back), andupdate these assignments on each trial (e.g., C, whichwas the letter 1-back in the previous trial is now theletter 2-back). These functions are also likely to beload-sensitive, since the complexity and number ofassignments increases withmemory load. Neural struc-tures subserving these and other executive processesengaged by the task should show activity profiles thatreflect this involvement. In the following study, weexamined whether additional brain regions would showa similar pattern of load sensitivity to the ones identi-fied in the first experiment.

EXPERIMENT 2

The first goal of this experiment was to replicate thefindings of Experiment 1 in a new group of subjects andfollowing modifications to the experimental protocol.Although the same behavioral paradigm was used, anumber of changes were made to specific task param-eters, scanner acquisition, and analysis procedures.Most importantly, we used the pooling procedure de-scribed above to combine data across subjects. A secondgoal of the experiment was to use whole-brain imagingtechniques to identify brain regions outside of PFC thatshow sensitivity to memory load and to gain a broaderview of the functional circuit underlying WM function.

Methods

Subjects

Informed consent was obtained from eight neurologi-cally normal right-handed subjects (six males, twofemales; 1825 years of age). The same pretestingprocedures were conducted as in Experiment 1.

Cognitive Tasks

Subjects performed the sequential letter task de-scribed above with the same four increments of WMload (zero to three items). However, two changes weremade to task parameters: (1) Block duration wasshortened to 57 s (19 stimuli) and (2) only sevenrepetitions of the four conditions were run (along withseven repetitions of another four conditions that werethe focus of a different study).

MRI Scanning Procedures

Images were acquired with a conventional 1.5-T GESigna whole-body scanner, using a multislice spiral-scanning procedure similar to the one described forExperiment 1. A number of changes were made to thescanning protocol to enable whole-brain acquisition.Functional images were acquired using a 4-interleavespiral-scan pulse sequence (Noll et al., 1995) with a TRof 640 ms, a gradient echo TE of 35 ms, a flip angle of40, and a field of view of 24 cm. A quadrature head coilwas used for homogeneous signal-to-noise ratio acrossthe brain. Scans were composed of isotropic voxels (3.75mm3) and acquired at 27 contiguous locations, perpen-dicular to the ACPC line, thus covering over 10 cm ofthe superiorinferior extent of the brain. Slice locationswere prescribed according to a procedure designed tomaximize reproducibility across subjects (Noll et al.,1996). Four scans were obtained from all 27 slicelocations during each task block. Structural imageswere obtained at these locations prior to functionalscanning, using a standard T1-weighted sequence.

57PREFRONTAL CORTEX AND WORKING MEMORY

Image Analysis Procedures

Prior to analysis, images for each subject were normal-ized to a common mean and movement-corrected usingan automated registration algorithm (Woods et al.,1992). Images were then coregistered and pooled acrosssubjects using a procedure similar to one standardlyused in PET studies (Wiseman et al., 1996; Woods et al.,1993). This was done by registering the structuralimages for each subject to a common reference brainusing a 12-parameter version of the registration algo-rithm. The set of parameters derived from this proce-dure were used to coregister the functional scans,which were then smoothed with an 8-mm FWHM 3DGaussian filter to accommodate between-subject differ-ences in anatomy. A focused contrast was used todirectly identify pixels showing a monotonic increase inactivity as a function of load. Focused contrasts aregenerally considered more powerful statistical teststhan simpleANOVAswhen a specific theoretical hypoth-esis is being examined (Rosenthal and Rosnow, 1985).Statistical maps of activity were thresholded for signifi-cance using a contiguity threshold of eight voxels andper-pixel false-positive rate of 0.005, ensuring an effec-tive pixel-wise a protection of 0.005. Anatomic localiza-tion was determined by overlaying the activation maponto the reference structural image and transformingthe data into Talairach coordinates (Talairach andTournoux, 1988) usingAFNI software (Cox, 1996).

Results

Replications

Our findings replicated the results of the first study,with activation present in both bilateral MFG andLIFG (see Fig. 5 and Table 1). The replication of activity

in these areas was noteworthy, given the changes madeto task parameters, image acquisition, and analysisprocedures. The use of a focused contrast ensured thatthe areas in PFC showed the same monotonic loadsensitivity as was exhibited previously. As shown inTable 1, the response function in MFG was similar inslope (0.27% increase in signal per item of load) to thatfound in the first experiment. In contrast, the LIFGregion showed a shallower slope, increasing at a rate of0.08% per item of load (compared to 0.47% in Experi-ment 1). The response profile in both regions was againfit well by a linear function (in both regions P , 0.001for the linear trend; quadratic and cubic trends, NS).

Other Regions of Activation

The analysis also revealed load-sensitive activity infrontal areas not observed in the first experiment: theright homologue of the LIFG region (BA 44), the leftfrontal operculum (insular cortex), and a number ofmotor, premotor, and supplementary motor regions(BAs 4 and 6). Nonfrontal activity was also found, inbilateral posterior parietal cortex (BA 40/7) and the leftcaudate nucleus of the basal ganglia (see Fig. 5 andTable 1). All of these regions have been implicated inthe distributed circuitry underlying WM, and mosthave been previously identified in PET studies using asimple subtractive version of this paradigm (Awh et al.,1996; Smith et al., 1996).

Discussion

The use of whole-brain imaging and cross-subjectpooling in this experiment allowed us to replicate ouroriginal findings and demonstrate their robustness toacquisition protocols and analysis procedures. More-

TABLE 1

Brain Areas Showing Monotonic Increases in Activity as a Function of Memory Load

Region ofinterest

Brodmannarea xa ya za

Volume(mm3)

1-backincreaseb(SEM)

2-backincreaseb(SEM)

3-backincreaseb(SEM)

Zscore

R superior frontal gyrus 8/6 13 8 54 6,611 0.12 (0.09) 0.33 (0.04) 0.55 (0.07) 4.03L middle frontal gyrus 46/9 242 23 39 1,527 0.12 (0.15) 0.44 (0.05) 0.81 (0.09) 4.03R middle frontal gyrus 46/9 37 37 33 4,273 0.18 (0.11) 0.48 (0.08) 0.81 (0.13) 5.24L inferior frontal gyrus 44/6 247 6 15 468 0.08 (0.05) 0.17 (0.05) 0.24 (0.04) 3.76R inferior frontal gyrus 44 43 6 31 583 0.17 (0.03) 0.26 (0.05) 0.31 (0.06) 3.05L frontal operculum Insula 232 20 8 520 0.09 (0.05) 0.15 (0.04) 0.30 (0.04) 3.43L precentral gyrus 6 239 21 29 241 0.10 (0.05) 0.25 (0.03) 0.29 (0.06) 3.19L precentral gyrus 6 247 22 43 350 0.05 (0.11) 0.45 (0.05) 0.50 (0.08) 3.19L precentral gyrus 4 230 26 57 1,350 0.09 (0.11) 0.42 (0.03) 0.57 (0.10) 3.27L inf/sup parietal lobule 40/7 226 260 45 8,737 0.17 (0.20) 0.53 (0.09) 0.75 (0.08) 4.22R inf/sup parietal lobule 40/7 32 256 43 12,643 0.15 (0.12) 0.52 (0.09) 0.78 (0.10) 4.29R caudate nucleus Basal ganglia 18 1 23 417 0.10 (0.06) 0.18 (0.05) 0.24 (0.04) 3.25

a x, y, z are coordinates relative to standard stereotactic space.b Increase expressed as percentage of change in signal relative to 0-back condition.

58 BRAVER ET AL.

over, these methods enabled the detection of additionalregions of load-sensitive activity both within PFC andin other cortical and subcortical areas. Of these addi-tional regions of activity, the parietal cortex was espe-cially notable. The volume of activity in this region wasthe largest of any of the areas identified, and coveredmost of the superior and inferior lobules (BA 40/7). Inmagnitude of effect, both PFC and parietal cortexdisplayed significantly stronger load-response func-tions than any other regions (P , 0.001). In general,however, the slopes of the load-response functions inthe activated areas were shallower in this experimentthan in the previous one. It is likely that this decreaseis due to the smoothing filter applied to the data prior toanalysis. The smoothing reduces variability across thebrain, but also attenuates any regional peaks in activ-ity. Thus, the magnitude of the response decreases, butsince variability is also decreased the effect is stillstatistically significant.

Laterality

We found no definitive evidence of lateralization inactivity in this experiment. Although the total acti-vated volume was greater in the right hemisphere thanin the left (24,527 mm3 vs 13,193 mm3), it cannot bedetermined whether this reflects greater right hemi-sphere engagement or a difference which is not func-tional in nature (e.g., unequal anatomical variabilityacross the hemispheres). Moreover, in a number of keyregions, such as middle frontal gyrus, inferior frontalgyrus, and posterior parietal cortex, we observed bilat-eral activation and no evidence of hemisphere 3 loadinteractions in activity (P . 0.1 for all regions). Thesefindings are in contrast with two previous studies ofverbal working memory, including one using a similarparadigm, in which subtle, but reliable left lateraliza-tion was found (Smith et al., 1996).There are a number of possible reasons for this

failure to replicate a laterality effect. First, whenlateralization is present, it is usually a subtle effect.Thus, the differences between the findings of the cur-rent study and that of Smith et al. (1996) could reflectvariability in subject populations. Given that bothstudies only included eight subjects, this explanation isa plausible one. Second, our study differed from previ-ous ones in the total time spent performing the task. Inthe 3-back task used in Smith et al. (1996), subjectsspent a total time of approximately 8 min performingthe letter conditions. In our n-back task, subjectsspent a total time of approximately 28 min performingthe letter conditions. Smith et al. (1996) suggested thatright hemisphere activation during this task mightreflect the functional recruitment of contralateralmechanisms to assist in task performance under highlydemanding conditions. This explanation, if correct,might account for the current findings, since greater

right hemisphere activity would be expected as thetime spent performing the task increases. Finally, theparametric nature of the n-back task used here mightlead to increased sensitivity in detecting brain regionssubserving the WM functions required by the task. Asdiscussed above, parametric designs are considered tohave both greater sensitivity and selectivity for detect-ing experimentally induced changes in a process ofinterest. Thus, our task design may have allowed moresensitive detection of an intrinsically bilateral neuralsystem underlying WM.

Functional Interpretations

As mentioned above, all of the additional regionsidentified in this experiment have previously beenimplicated in WM function. Parietal activation hasbeen observed in numerous WM studies, and has beeninterpreted as a buffer for modality-specific informa-tion (Paulesu, 1993). The basal ganglia form strongreciprocal connections with PFC (Alexander et al.,1986), and these frontostriatal loops may be a keycomponent of WM circuitry. The left-hemisphere motorareas (SMA, premotor, and primary motor cortex) havebeen identified in a 2-back version of this paradigm(Awh et al., 1996). These regions, in cooperation withBrocas area, might mediate the subvocal articulatoryprocesses that are thought to subserve verbal re-hearsal. The load sensitivity observed in these areas isconsistent with the increase in number of rehearsalitems associated with each increase in load. Alterna-tively, in a subset of themotor areas, especially primarymotor cortex, the activity may reflect changes in motorreadiness. Since higher loads are likely to demandmore processing time, this leaves less time for responsepreparation. The systemmay adapt to this bymaintain-ing a higher level of motor readiness, producing a tonicincrease in activity within primary motor areas. Theleft operculum was also activated in the Awh et al.study. Interestingly, this region has also been observedin tasks requiring retrieval of verbal information fromlong-term memory (Buckner and Petersen, 1996). Bothsituations require a comparison of internally storedrepresentations of stimuli with externally presentedones. Thus, it is possible that the operculum may beinvolved in matching or recognition operations associ-ated with these sorts of comparisons. These matchingoperations may become more difficult (or take longertime) as the size of the memory set increases (e.g.,Sternberg, 1966).

GENERAL DISCUSSION

Active Maintenance

Our observations of load-sensitive activity in dorsolat-eral PFC, parietal cortex, and basal ganglia are consis-

59PREFRONTAL CORTEX AND WORKING MEMORY

tent with the neurophysiological literature in nonhu-man primates. Single-cell recordings in these regionshave observed sustained neural activity during thedelay periods of short-term memory tasks, and thispattern of activity has been typically interpreted as theneural mechanism for the active maintenance of infor-mation (Goldman-Rakic, 1987). Consequently, the mostobvious interpretation of our findings is that the ob-served regions of activity reflect the operation of main-tenance processes associated with WM. The load ma-nipulations in the task directly impact both the numberof items maintained (zero to three items) and theduration of maintenance (up to 9 s for 3-back stimuli).However, this confounding of delay and load in ourparadigmmakes it difficult to clearly interpret whetherthe observed activity selectively reflects active mainte-nance rather than other WM-related processes. Asmentioned above, there are a number of nonmainte-nance processes operating in this task that are alsolikely to be affected by load manipulations. Theseinclude matching/comparison operations and executiveprocesses, such as assignment of sequential order andmemory updating. Thus, the question of whether theload-sensitive neural structures identified in this taskare associated with maintenance vs executive or otherWM-related processes cannot be answered definitivelyfrom this study alone; additional work is needed tofunctionally decompose the relevant task components.A potential approach to selectively engaging brain

regions subserving active maintenance is to hold WMload constant and vary only the delay over whichinformation must be maintained. We have taken thisapproach in a recently completed study, and found thatonly dorsolateral PFC, Brocas area, and parietal cortexshowed increases in activity associated with delay(Braver et al., 1996). A second approach is to examinethe time course of activity in the various regions duringtask performance. This temporal information shouldhelp to dissociate regions subserving maintenance vsexecutive processes, insofar as the former should ex-hibit sustained activity over the trial duration whilethe latter should be stimulus-locked and transient.fMRI can provide the temporal resolution necessary toobtain such time course information, given an adequateinterstimulus interval in which to accommodate thephase lag of the hemodynamic response and to allow itto revert to baseline (Savoy et al., 1995; Vazquez andNoll, 1996). We have also begun to utilize this approachand have found that we can differentiate the responseprofiles of the various regions activated in this task(Cohen et al., 1996). In particular, we have observedthat activity in PFC is sustained throughout the reten-tion interval. This finding is particularly significant,since it has been proposed that PFC houses the execu-tive control functions of WM, while active maintenanceis carried out by other brain regions (Gathercole, 1994).

Shape of Load-Response Functions

In both of the experiments discussed here, we wereable to plot the activity of specific brain regions inresponse to parametric changes in WM load. Althoughour analysis procedure was designed to select onlyregions showing amonotonically increasing response toload, it did not specify the particular form of theresponse. There are a number of different types ofmonotonic functions that could have been observed,including nonlinear functions such as exponentiallyaccelerating, exponentially decelerating, or sigmoidal.Moreover, each of these types could be significant forthe study of WM. For example, an exponentially decel-erating function could indicate the saturation/plateau-ing of WM mechanisms above some level of load, whilea sigmoid or step function might indicate the presence/emergence of qualitatively distinct WMmechanisms atdifferent levels of load. In contrast, a linear functionmight be interpreted as indicating a single neuralmechanism that additively increases by a fixed amountwith each additional item of load. Likewise, the slope ofsuch a function might specify the level of neuralactivation needed to process and/or maintain each itemin a particular brain region. There is a precedent forthese sorts of interpretations from behavioral studies,which have attributed functional significance to linearRT functions (e.g., Shepard and Metzler, 1971; Stern-berg, 1966). Thus, it is of potential theoretical interestto determine the specific load-response functions foundin the current data.As we have mentioned above, a linear function

seemed to best fit the PFC response and most of theother identified brain regions as well. This characteriza-tion was determined through statistical tests of lineartrend as well as through visual inspection. At the sametime, we believe that it would be premature to drawtheoretical inferences from this finding for severalreasons. First, the test of linear trend does not differen-tiate between linear and nonlinear monotonic func-tions; it is only useful for rejecting the null hypothesisthat the best fitting line through the data has zero slope(Braver and Sheets, 1993). Statistical tests that coulddifferentiate between monotonic functions would re-quire amuch larger sample and/or a wider range of loadintervals to have sufficient power. Second, in other,more recent studies of this task, we have observed anumber of regions showing load-response functionswhich appeared more step-like than linear (Cohen etal., 1996; Jonides et al., 1996). This suggests that theshape of these functions may be variable and/or sensi-tive to slight differences in experimental factors. Fi-nally, even if the response functions could be reliablydetermined to be linear, it is not yet clear that theoreti-cal inferences could be drawn without convergent datafrom other sources. For example, the apparent linearityof the load-response function could be an artifact of the

60 BRAVER ET AL.

way neural activity is detected with fMRI methods.Since it is known that activity changes are convolvedwith a nonlinear hemodynamic response function(Kwong et al., 1992; Vazquez and Noll, 1996), it ispossible that this convolution may act to blunt orotherwise distort subtle features of the actual response.Taken together, it seems clear that neuroimaging re-sponse functions require additional study before inter-pretations based on their specific form are warranted.However, even with these caveats in mind, our

finding of monotonic load sensitivity in PFC and otherbrain areas provides important new information regard-ing the involvement of these regions in WM. Further,both the parametric design and the correlations withbehavioral performance ameliorate many of the meth-odological concerns which surround simple subtractivedesigns. Finally, these results highlight the utility ofparametric cognitive neuroimaging studies, by demon-strating that it is possible to detect slight, but reliablechanges in the activity of circumscribed brain regionsin response to subtle task manipulations (e.g., the2-back vs 3-back conditions).

ACKNOWLEDGMENTS

This work was supported in part by grants from the McDonnellFoundation, the National Institute of Mental Health (J.D.C.), and theU.S. Office of Naval Research (J.J. and E.E.S.). The authors thankEdward Awh, Erick Lauber, Christy Marshuetz, Robert Orr, FredSabb, and Eric Schumacher for their technical assistance, andDeanna Barch, Marlene Behrmann, Sanford Braver, B. J. Casey,Steve Forman, and Prahlad Gupta for their thoughtful commentsand helpful suggestions.

REFERENCES

Alexander, G. E., Delong, M. R., and Strick, P. L. 1986. Parallelorganization of functionally segregated circuits linking basal gan-glia and cortex. Ann. Rev. Neurosci. 9:357381.

Awh, E., Jonides, J., Smith, E. E., Schumacher, E. H., Koeppe, R., andKatz, S. 1996. Dissociation of storage and rehearsal in verbalworking memory: Evidence from PET. Psychol. Sci. 7:2531.

Baddeley, A. D. 1986. Working Memory. Oxford Univ. Press, NewYork.

Braver, S. L., and Sheets, V. L. 1993. Monotonic hypotheses inmultiple group designs: A Monte Carlo study. Psychol. Bull.113:379395.

Barch, D. M., Braver, T. S., Nystrom, L. E., Forman, S. D., Noll, D. C.,and Cohen, J. D. 1996. Dissociating working memory from effort inhuman prefrontal cortex. Submitted for publication.

Buckner, R. L., and Petersen, S. E. 1996. What does neuroimagingtell us about the role of prefrontal cortex in memory retrieval?Semin. Neurosci., 8:4755.

Cohen, J. D., Forman, S. D., Braver, T. S., Casey, B. J., Servan-Schreiber, D., and Noll, D. C. 1994. Activation of prefrontal cortexin a nonspatial working memory task with functional MRI. Hum.Brain Map. 1:293304.

Cohen, J. D., MacWhinney, B., Flatt, M. R., and Provost, J. 1993.PsyScope: A new graphic interactive environment for designingpsychology experiments. Behav. Res. Methods Instruments Com-put. 25(2):257271.

Cohen, J. D., Noll, D. C., and Nystrom, L. 1995. Qualitative andquantitative assessment of testretest reliability of functional MRIdata. Paper presented at the Third Scientific Meeting, Nice,France.

Cohen, J. D., Perstein, W. M., Braver, T. S., Nystrom, L. E., Noll,D. C., Jonides, J., and Smith, E. 1996. Temporal dynamics ofcortical activity in verbal working memory. Submitted for publica-tion.

Courtney, S. M., Ungerleider, L. G., Keil, K., and Haxby, J. V. 1996.Object and spatial visual workingmemory activate separate neuralsystems in human cortex. Cereb. Cortex 6:3949.

Cox, R. W. 1996. AFNI: Software for analysis and visualization offunctional magnetic resonance neuroimages. Comput. Biomed.Res. 29:162173.

Dehaene, S., Posner, M. I., and Tucker, D. M. 1994. Localization of aneural system for error detection and compensation. Psychol. Sci.5:303306.

Engel, S. A., Rumelhart, D. E., Wandell, B. A., Lee, A. T., Glover,G. H., Shadlen, M. N., and Chichilinsky, E.-J. 1994. fMRI of humanvisual cortex.Nature 369:525.

Fiez, J. A., Raife, E. A., Balota, D. A., Schwarz, J. P., Raichle, M. E.,and Peterson, S. E. 1996. A positron emission tomography study ofthe short-term maintenance of verbal information. J. Neurosci.16(2):808 822.

Forman, S. D., Cohen, J. D., Fitzgerald, M., Eddy, W. F., Mintun,M. A., and Noll, D. C. 1995. Improved assessment of significantactivation in functional magnetic resonance imaging (fMRI): Use ofa cluster-size threshold.Magn. Reson. Med. 33:636647.

Fuster, J. M. 1989. The Prefrontal Cortex: Anatomy, Physiology andNeuropsychology of the Frontal Lobe. Raven Press, NewYork.

Fuster, J. M., andAlexander, G. E. 1973. Firing changes in cells of thenucleus medialis dorsalis associated with delayed response behav-ior. Brain Res. 61:7991.

Fuster, J. M., and Jervey, J. 1981. Inferotemporal neurons distin-guish and retain behaviorally relevant features of visual stimuli.Science 212:952955.

Gathercole, S. E. 1994. Neuropsychology and working memory: Areview.Neuropsychology 8(4):494505.

Gehring, W. J., Goss, B., Coles, M. G. H., Meyer, D. E., and Donchin,E. 1993. A neural system for error detection and compensation.Psychol. Sci. 4:385390.

Gevins, A. S., and Cutillo, B. C. 1993. Neuroelectric evidence fordistributed processing in human working memory. Electroencepha-logr. Clin. Neurophysiol. 87:128143.

Gnadt, J. W., and Anderson, R. R. 1988. Memory related motorplanning activity in the posterior parietal cortex of macaque. Exp.Brain Res. 70:216220.

Goldman-Rakic, P. S. 1987. Circuitry of primate prefrontal cortex andregulation of behavior by representational memory. InHandbook ofPhysiologyThe Nervous System (F. Plum and V. Mountcastle,Eds.), Vol. 5, pp. 373417. Am. Physiol. Soc., Bethesda.

Grasby, P. M., Frith, C. D., Friston, K. J., Simpson, J., Fletcher, P. C.,Frackowiak, R. S. J., and Dolan, R. J. 1994.Agraded task approachto the functional mapping of brain areas implicated in auditoryverbal memory. Brain 117:12711282.

Herholz, K., Thiel, A., Pietrzyk, U., von Stockhausen, H.-M., Kessler,J., and Heiss, W.-D. 1996. Individual functional anatomy of Brocasarea.NeuroImage 3: S134.

Jonides, J., Schumacher, E. H., Smith, E. E., Lauber, E. J., Awh, E.,Minoshima, S., and Koeppe, R. A. 1996. Verbal working memoryload affects regional brain activity as measured by PET. J. Cognit.Neurosci., in press.

Jonides, J., Smith, E. E., Koeppe, R. A., Awh, E., Minoshima, S., and

61PREFRONTAL CORTEX AND WORKING MEMORY

Mintun, M. A. 1993. Spatial working memory in humans asrevealed by PET.Nature 363:623625.

Just,M.A., and Carpenter, P.A. 1992.Acapacity theory of comprehen-sion: Individual differences in working memory. Psychol. Rev.99(1): 122149.

Kwong, K. K., Belliveau, J. W., Chesler, D. A., Goldberg, I. E.,Weisskoff, R. M., Poncelet, P., Kennedy, D. N., Hoppel, B. E., S., C.M., Turner, R., Cheng, H. M., Brady, T. J., and Rosen, B. R. 1992.Dynamic magnetic resonance of human brain activity duringprimary sensory stimulation. Proc. Natl. Acad. Sci. USA 89:56755679.

McCarthy, G., Blamire,A. M., Puce,A., Nobre,A. C., Bloch, G., Hyder,F., Goldman-Rakic, P., and Shulman, R. G. 1994. Functionalmagnetic resonance imaging of human prefrontal cortex during aspatial working memory task. Proc. Natl. Acad. Sci. USA 91:86908694.

McClelland, J. L. 1979. On the time relations of mental processes: Anexamination of systems of processes in cascade. Psychol. Rev.86:287330.

Noll, D. C., Cohen, J. D., Meyer, C. H., and Schneider, W. 1995. SpiralK-space MR imaging of cortical activation. J. Magn. Reson. Imag-ing 5(1):4956.

Noll, D. C., Genovese, C. R., Nystrom, L. E., Vazquez, A., Forman,S. D., Eddy, W. F., and Cohen, J. D. 1996. Estimating testretestreliability in functional MR imaging. II. Application to motor andcognitive activation studies. Submitted for publication.

OLeary, D. S., Andreason, N. C., Hichwa, R. D., Cizadlo, T. J., Hurtig,R. R., Boles Ponto, L. L., Watkins, L. G., and Kesler, M. L. 1996.Variability in brain activation: The relationship between indi-vidual and group PET images of regional cerebral blood flow.NeuroImage 3:S86.

Petrides, M. E., Alivisatos, B., Meyer, E., and Evans, A. C. 1993.Functional activation of the human frontal cortex during theperformance of verbal working memory tasks. Proc. Natl. Acad.Sci. USA 90:878882.

Posner, M. I., Petersen, S. E., Fox, P. T., and Raichle, M. E. 1988.Localization of cognitive operations in the human brain. Science240:16271631.

Rosenthal, R., and Rosnow, R. L. 1985. Contrast Analysis. CambridgeUniv. Press, Cambridge, UK.

Sagar, H., Sullivan, E., Gabrieli, J., Corkin, S., and Growden, J. 1988.Temporal ordering and short-term memory deficits in Parkinsonsdisease. Brain 111:525539.

Savoy, R. L., Bandettini, P. A., OCraven, K. M., Kwong, K. K., Davis,T. L., Baker, J. R., Weisskoff, R. M., and Rosen, B. R. 1995. Pushingthe temporal resolution of fMRI: Studies of very brief visualstimuli, onset variability and asynchrony, and stimulus-correlatedchanges in noise. Paper presented at the Society of MagneticResonance, 3rd Meeting.

Schneider, W., Casey, B. J., and Noll, D. C. 1994. Functional MRImapping of stimulus rate effects across visual processing stages.Hum. Brain Map. 1:117133.

Schneider, W., Noll, D. C., and Cohen, J. D. 1993. Functional

topographic mapping of the cortical ribbon in human vision withconventional MRI scanners.Nature 365:150153.

Schultz, W., and Romo, R. 1988. Neuronal activity in the monkeystriatum during the initiation of movements. Exp. Brain Res.71:431 436.

Shallice, T. 1988. From Neuropsychology to Mental Structure. Cam-bridge Univ. Press, Cambridge, UK.

Shepard, R. N., and Metzler, J. 1971. Mental rotation of three-dimensional objects. Science 171:701703.

Shimamura, A. P. 1994. Memory and frontal lobe function. In TheCognitive Neurosciences (M. S. Gazzaniga, Ed.), pp. 803815. MITPress, Cambridge, MA.

Smith, E. E., Jonides, J., and Koeppe, R. A. 1996. Dissociating verbaland spatial working memory using PET. Cereb. Cortex 6:1120.

Sternberg, S. 1966. High-speed scanning in human memory. Science153:652654.

Sternberg, S. 1969. The discovery of processing stages: Extensions ofDonders method. In Attention and Performance II (W. G. Koster,Ed.). North-Holland, Amsterdam.

Stuss, D. T., Eskes, G. A., and Foster, J. K. 1994. Experimentalneurospychological studies of frontal lobe function. InHandbook ofNeuropsychology (F. Boller and J. Grafman, Eds.), Vol. 9. Elsevier,Amsterdam.

Swartz, B. E., Halgren, E., Fuster, J. M., Simpkins, E., Gee, M., andMandelkern, M. 1995. Cortical metabolic activation in humansduring a visual memory task. Cereb. Cortex 5:205214.

Talairach, J., and Tournoux, P. 1988. Co-Planar Stereotaxic Atlas ofthe Human Brain. Thieme, NewYork.

Tootell, R. B. H., Reppas, J. B., Kwong, K. K., Malach, R., Born, R. T.,Brady, T. J., Rosen, B. R., and Belliveau, J. W. 1995. Functionalanalysis of human MT and related visual cortical areas usingmagnetic resonance imaging. J. Neurosci. 15:32153523.

Vallar, G., and Shallice, T. 1990. Neuropsychological Impairments ofShort TermMemory. Cambridge Univ. Press, Cambridge, UK.

Vazquez, A. L., and Noll, D. C. 1996. Non-linear temporal aspects ofthe BOLD response in fMRI. International Society of MagneticResonance in Medicine, Proceedings, 4th Meeting, p. 1765.

Weinrich, M., and Wise, S. P. 1982. The premotor cortex of themonkey. J. Neurosci. 2:13291345.

Wessinger, C. M., and Gazzaniga, M. S. 1996. Between subjectvariability in the location of the calcarine fissure. NeuroImage3:S147.

Wiseman, M. B., Nichols, T. E., Dachille, M. A., and Mintun, M. A.1996. Working towards an automatic and accurate method forPET-MR alignment. J. Nucl. Med. 37:224.

Woods, R. P., Cherry, S. R., and Mazziotta, J. C. 1992. Rapidautomated algorithm for aligning and reslicing PET images. J.Comput. Assisted Tomogr. 16:620633.

Woods, R. P., Mazziotta, J. C., and Cherry, S. R. 1993. MRIPETregistration with automated algorithm. J. Comput. Assisted To-mogr. 17(4):536546.

62 BRAVER ET AL.