Running head: TRUE AND FALSE FACE RECOGNITION Component Structure of Individual Differences in True and False Recognition of Faces

Principal Component Analyses (PCA) of four face-recognition studies uncovered two independent components. The first component was strongly related to false-alarm errors with new faces as well as facial “conjunctions” that recombine features of previously studied faces. The second component was strongly related to hits as well as to the conjunction/new difference in false-alarm errors. The pattern of loadings on both components was impressively invariant across the experiments which differed in age range of participants, stimulus set, list length, facial orientation, and the presence versus absence of familiarized lures along with conjunction and entirely new lures in the recognition test. Taken together, the findings show that neither component was exclusively related to discrimination, criterion, configural processing, featural processing, context recollection, or familiarity. Rather, the data are consistent with a neuropsychological model that distinguishes frontal and occipito-temporal contributions to face recognition memory. Within the framework of the model, our PCA findings showed that frontal and occipito-temporal contributions are discernable from the pattern of individual differences in behavioral performance among healthy young adults. Component Structure 3 Component Structure of Individual Differences in True and False Recognition of Faces False memory phenomena are receiving much attention (Brainerd & Reyna, 2005), but the voluminous research on this topic has been limited in its scope. Firstly, most work in this area has focused on words or easily verbalized stimuli. Less attention has been paid to hard-to-verbalize stimuli such as human faces. False memories for faces are ecologically important when they occur in eyewitness identification (Charman & Wells, 2007; Penrod & Bornstein, 2007), and they are relevant to theories of memory because faces may differ from verbal stimuli in the way they are processed and represented (Tanaka & Farah, 1993, 2003; McKone, Kanwisher & Duchaine, 2007). Secondly, individual differences in false recognition are rarely addressed in current research. This is unfortunate, because, as has been pointed out repeatedly, examining the pattern of individual differences can aid in “. . . deciphering the basic structure and processes of memory” (Bors & MacLeod, 1996, p. 436; see also Melton, 1967; Underwood, 1975). In order to improve our understanding of false facial recognition, we used Principal Component Analysis (PCA, e.g., Jolliffe, 2002) to analyze individual differences in correct and false recognition of faces. In the experiments reported here, study lists of unfamiliar faces were followed by recognition tests that included old faces, entirely new faces, and one or two types of difficult lure known to elicit frequent false recognitions. PCAs performed on the recognition data revealed two orthogonal components that consistently differed in their relations to hits and false alarms with the different lure-types: Whereas one component was related to false alarms but not to hits, the other was primarily related to hits and to the pattern of false alarms across different Component Structure 4 lure-types. The theoretical importance of these components is best viewed in light of three lines of prior work which we review in turn: The first line involves the lineup task, the second concerns memory impairments linked to frontal-lobe damage and to aging, and the third concerns the “conjunction effect,” a well established phenomenon of false recognition that is robust with faces. Sequential Presentation of Lineups False identifications in the lineup task are affected by a number of variables (Kassin, Tubb, Hosch & Memon, 2001; Penrod & Bornstein, 2007), with presentation format being particularly potent. For example, compared to the standard procedure of presenting all lineup faces simultaneously, presenting them sequentially reduces false identifications in “target-absent” lineups (i.e., when all faces are foils). Importantly, sequential presentation has this effect without reducing correct identifications in “targetpresent” lineups (Lindsey & Wells, 1985; and see Steblay, 1997, for a similar asymmetric effect produced by “unbiased” or “fair” lineup instructions). The effect is important, but the attractiveness of the sequential procedure to the forensic community has been questioned by several recent studies in which sequential presentation has reduced correct identifications in target present lineups as well as false identifications in target absent lineups (e.g., Memon & Bartlett, 2002, and see Steblay, Dysart, Fulero, & Lindsay, 2001, for a review and meta-analysis). We need to know more about when and how a variable such as sequential presentation might influence false alarms without affecting hits. Face recognition deficits by brain-damaged persons and older adults are informative in this regard, and we turn to these deficits next. Face Recognition in Neuropsychological Patients and Healthy Older Adults Component Structure 5 It is widely known that damage to the right occipito-temporal brain region is linked to deficits in face recognition memory (Milner, 1968; Damasio, Tranel & Damasio, 1990; De Renzi, Perani, Carlesimo, Silveri & Fazio, 1994; Sergent, & Signoret, 1992; Tovée & Cohen-Tovée, 1993). It is less widely known that right frontal lobe damage is linked to face-recognition deficits as well (see, e.g., Rapcsak, Reminger, Glisky, Kaszniak & Comer, 1999; Rapcsak, Nielsen, Littrell, Glisky, Kaszniak & Laguna, 2001). Importantly, deficits due to right occipito-temporal damage differ qualitatively from deficits due to frontal damage in that the former involve decreased correct recognition whereas the latter involve increased false recognition. For example, in tests of old/new recognition of previously unfamiliar faces, Rapcsak et al. (1999) found that five patients with right medial-temporal lobe damage produced far fewer hits to old faces, but only slightly more false alarms to new faces (as compared to healthy normal controls). By contrast, two other patients with right frontal lobe damage were comparable to healthy normal controls in hits, but showed extremely high false-alarm rates. This latter finding was subsequently confirmed with 10 additional patients with right frontal damage, and has been shown to extend to “retrograde” tests that require discrimination between famous and non-famous faces (Rapcsak et al., 1999; 2001). Healthy older adults resemble patients with frontal lobe damage in the deficits they show in face recognition. Several studies found that older persons, as compared to young adults, have inflated false-alarm rates but similar hit rates (see Bartlett, 1993 and Searcy, Bartlett, & Memon, 1999, for reviews). These findings extend to the lineup task, where false identifications from target-absent lineups are usually higher among older eyewitnesses than among younger eyewitnesses while correct identifications from targetComponent Structure 6 present lineups generally show little or no age difference (see Bartlett & Memon, 2007). Because a subset of the elderly appear to suffer deficits in frontal processing (see, e.g., Anderson & Craik, 2000), these findings along with the neuropsychological evidence suggest that an isolable component of face recognition memory (and perhaps recognition memory in general) works selectively to oppose false alarms. Based on their observations, Rapcsak et al. (1999) proposed a neuropsychological model of face recognition, shown in Figure 1. According to the model, a “frontal executive system” works to reduce false recognition through several processes including “strategic search,” “effortful recollection” of contextual information, monitoring the contents of memory retrieval, and decision-making processes (including the setting of criteria). In addition to frontally mediated processes that work to reduce false alarms, Rapcsak et al. (1999) proposed an occipito-temporal component, which they characterized as a “face recognition module.” The occipito-temporal component supports recognition of previously viewed faces by signaling resemblance between a face being currently viewed and stored memory representations of previously seen faces. Although a recognition judgment might be based on a strong resemblance signal, Rapcsak et al. (1999) assert that, “It is only when appropriate contextual information is recovered that we feel confident that a face has been successfully identified” (p. 285; and see Mandler, 1980, who first made this point). In the case of well-known faces, context retrieval refers to the recovery of biographical information. In the case of old/new recognition of previously unfamiliar faces, context retrieval refers to “activating information pertaining to the study episode” (see right-most box in Figure 1). In either case, context retrieval can occur in two ways. First, it can be triggered in an automatic, Component Structure 7 bottom-up fashion by the resemblance-signaling face recognition module. Second, it can be retrieved through strategic search and/or effortful recollection controlled by the frontal executive component. A key claim of the Rapcsak et al. (1999) model is that successful recognition of a previously seen face can occur with minimal executive control through a strong resemblance signal from the face recognition module, and/or bottom-up retrieval of contextual and/or biographic information triggered by the output of the module. Hence, successful recognition of previously viewed faces is more dependent on the face recognition module than on the frontal executive component. False recognition of new faces is a different matter. Because all faces share the same first-order configuration (hairline above eyes, eyes above nose, etc.; see Diamond & Carey, 1986) and faces of different persons can be highly similar in their features, the face recognition module could often signal strong resemblance even for new faces. The model holds that frontally mediated top-down control is critical for reducing false alarms to such faces. The Rapcsak et al. (1999) model may require updating (see, Rapcsak, 2003, and the General Discussion), but, nonetheless, it nicely handles the neuropsychological and aging data on which it was based, and provides a perspective on the question of why, in some conditions, a procedure such as sequential presentation might reduce false recognitions without affecting hits in the lineup task. It also may illuminate a third line of research to which we turn next. The Facial Conjunction Effect A “conjunction” is a specially constructed recognition-test-lure that recombines the parts of two previously studied items. In the case of faces, a conjunction might Component Structure 8 contain: (a) the inner features of one studied face and the outer features of another studied face, as shown in Figure 2 (e.g., Bartlett, Searcy, & Abdi, 2003), (b) the eyes and mouth of one face and the outer features and nose of another face (see Reinitz, Morrissey & Demb, 1994; Jones, Bartlett & Wade, 2006), or (c) the eyes and eyebrows of one face with the nose and mouth of another (with hair and other outer features removed, see McKone & Peh, 2006). On a recognition test that includes old faces, conjunctions, and entirely new faces, hits to old faces are generally more frequent than false alarms to conjunctions. However, the latter are considerably more frequent than false alarms to entirely new faces. For example, hits, conjunction false alarms and new-face false alarms occurred at rates of approximately .67, .53 and .33, respectively, with upright faces in Experiment 2 of McKone and Peh (2006). The conjunction/new difference in false-alarm errors (.20 in this case) constitutes the conjunction effect. 1 Rapcsak et al. (1999) did not address the conjunction effect in developing their model. Yet, the effect is precisely what one would expect if face recognition involves a component that 1) signals the amount (but not the nature) of resemblance between a currently viewed face and information in memory, and 2) can also trigger bottom-up retrieval of contextual information. A facial conjunction is likely to have partial resemblance to the memory representation of each of its two “parents” seen in the study list, such resemblance being based on matching abstract features (e.g., distinctive hookshaped nose) or matching low-level visual codes capturing the gradations in luminance values across facial regions (see Bartlett, Searcy & Abdi, 2003). In any case, the summed resemblance signal evoked by a conjunction is likely to be strong. A strong resemblance signal may frequently be accompanied by retrieval of contextual information, which can Component Structure 9 oppose false alarms in some situations (see Kelley & Jacoby, 2000). However, contextual information is unlikely to be helpful, and may even be harmful, in the case of conjunctions, as the parents of conjunctions were previously studied in the same general context . Hence, the Rapcsak et al. (1999) model correctly predicts that conjunction faces should be highly attractive lures, even if contextual information is frequently retrieved. The Present Research Our goal in the present research was to examine individual differences in correct recognition and false recognition in the facial conjunction paradigm, and to determine whether these differences might be explained by the Rapcsak et al. model or by alternative conceptions of recognition memory. Our point of departure was a PCA of the data from a previously published facial-conjunction study (Searcy et al., 1999, Experiment 2). Seventy-five young adults and 76 older adults viewed a study list of 16 photographs of faces, with each face presented twice. This was followed by a test containing 8 old faces, 8 new faces and 8 conjunctions. Each conjunction combined the internal features (eyes, nose, and mouth) of one study-list face with the external features (hair, ears, jaw, and chin) of another (see Figure 2). The task was to classify old faces as “old,” while rejecting both conjunctions and new faces as “new.” As in prior research with conjunction faces, we found that hit rates for old faces (M = .78) exceeded falsealarm rates for conjunctions (M = .41), which in turn exceeded false-alarm rates for new faces (M = .14). The PCA was performed on a table (3 columns and 151 rows) containing the hit rates for old faces, the false-alarm rates for conjunctions, and the false-alarm rates for new faces for each of the 151 participants. It produced two components with eigenvalues Component Structure 10 greater than 1.0. The upper graph in Figure 3 displays the loadings of each stimulus condition on each of the two components. Note that the first component, which accounted for 42% of the variance (eigenvalue = 1.25), produced a near-0 loading for hits, but strong negative loadings for conjunction false alarms and new-face false alarms. Hereafter, we refer to this first component as the “false-alarm-rate component” or simply as “Factor 1.” The second component accounted for 39% of the variance (eigenvalue = 1.15), and produced a strong positive loading for hits, a moderate positive loading for conjunction false alarms, and a moderate negative loading for new-face false alarms. Hereafter, we refer to this second component as the “hit-rate component” or simply “Factor 2.” The outcome indicated that hits and false alarms in face recognition memory reflect separable processing components, in line with the Rapcsak et al. (1999) model. Beyond this, the PCA suggested that the conjunction effect is linked to the hit-rate component (Factor 2) and not the false-alarm-rate component (Factor 1). 2 This last point is best illustrated in Figure 4, which shows the mean proportions of “old” judgments made to each of the three item-types as a function of Factor 1 scores (left-hand graph) and Factor 2 scores (right-hand graph). 3 In each case, we sorted the participants by their scores on the factor, and then placed them into six different groups (n = 25 or 26 per group) based on these factor scores. The lowest scoring group was labeled Group 1, the next-lowest scoring group was labeled Group 2, and so on through Group 6 (the highest-scoring group). Note in the left-hand graph that hits were largely unrelated to Factor 1 scores, but that false alarms fell as Factor 1 scores increased. This effect was apparent with both types of lure. By contrast, the right-hand graph shows that hits rose steeply with Factor 2 scores. Furthermore, while conjunction false alarms rose Component Structure 11 with Factor 2 scores, new-face false alarms decreased. As a result, the conjunction effect (i.e., the conjunction/new difference in false-alarm rates) strengthened from .05 in the lowest Factor 2 group to .57 in the highest Factor 2 group. To statistically evaluate this observation, we computed the correlation between individual participants’ Factor 2 scores and the conjunction/new difference in their false-alarm rates. The correlation was robust (r = .65, df = 149, p < .001). By contrast, the corresponding correlation of Factor 1 scores with the conjunction/new difference was nil (r = .10, df = 149, ns). In summary, the PCA suggested that two components underlie individual differences in face recognition memory, one that is linked to false recognition of both conjunctions and new items, and a second that is linked to hits as well as to the conjunction effect. The outcome was of interest because the two components appeared to map onto the false-alarm-controlling frontal executive system and the hit-enhancing occipito-temporal face recognition module in the neuropsychological model of Rapcsak et al (1999). That hit rates and the conjunction effect were linked to the same component appeared particularly impressive, as this outcome fit the model’s claim that the face recognition module — which is important for hits—signals the amount but not the nature of resemblance between a test face and information in memory. In the experiments that follow, we sought to replicate and extend our original PCA results and, also to characterize the hit and false-alarm components in theoretical terms. It was important to replicate the dual-component outcome with a sample of healthy young adults because the Searcy et al. (1999) study included older adults, and so our first component might simply reflect that elderly participants make more false alarms.. Assuming that the two-component outcome extends to young-only samples, it Component Structure 12 was important to examine alternative theoretical characterizations of the two components. Experiments 1 and 2 addressed two such alternative hypotheses. The first hypothesis was that our two components reflect uncorrelated variance in discrimination and criterion in the sense of signal detection theory (Macmillan & Creelman, 2005). The data in Figure 4 appear to rule against this notion, as both factors are related to differential responding to old faces versus new faces and conjunctions. However, it was important to gather more definitive data using signal-detection-theory measures, and we did so in Experiment 1. The second hypothesis was that the two components reflected the separable roles of configural and featural information in facial recognition (e.g., Bartlett et al., 2003; Cabeza & Kato, 2000; Rakover, 2002). Facial inversion (i.e., presenting faces upsidedown) impairs configural processing more than featural processing (Bartlett & Searcy, 1993; Searcy & Bartlett, 1996; Freire, Lee & Symons, 2000), and, in recognition memory, it impairs discrimination of old faces from both conjunctions and new items (McKone & Peh, 2006; Bartlett et al., 2003). Hence, a reasonable hypothesis was that our first component _ which had strong same-sign loadings for conjunction false alarms and new-face false alarms _ reflected configural processing, while our second component _ which had reversed-sign loadings for the two types of false alarm_ reflected featural processing. The clear prediction was that facial inversion should weaken our first component much more than our second. We tested this prediction by presenting all faces upside-down to half of the participants in both Experiments 1 and 2. Experiments 1a and 1b Experiment 1a was briefly described in Bartlett et al. (2003), though without the PCA results. Experiment 1b has not been published in any form. The two studies were Component Structure 13 identical in design, and both included between-groups variables (presentation frequency and test delay) that were expected to influence recognition accuracy and insure sufficient variance for an informative PCA. The two experiments differed only in that all faces at study and at test were shown upright in Experiment 1a and inverted in Experiment 1b. Method Participants. Following procedures approved by the Institutional Review Board of Texas Woman’s University, 192 participants were recruited from the University community. Their ages ranged from 18 to 53 years (M = 30 years). Half of the participants served in Experiment 1a while the remainder served in Experiment 1b. Materials. The materials were made from digitally scanned black-and-white, fullfrontal, head-and-shoulders view photographs of 48 young-adult males from the 1965 Texas A&M; yearbook. All had short-cropped hair, a dark coat, white shirt and dark tie. Each face was photographed against the same gray background. Pairs of faces that shared such characteristics as approximate age, face shape, and complexion were identified. Each face in a pair served as a parent “original” from which two “synthetic” faces were created using Aldus Photostyler and Corel Photo-Paint. Each synthetic face combined the internal facial region (i.e., eyes, nose, and mouth) of one original face with the external facial region (i.e., hair, ears, and jaw line) of the other. The resulting stimulus set comprised 24 “quadruplets” each including two original faces and two synthetic faces (see Figure 2 for an example of a quadruplet). Design and Procedure. A 32-item study list (plus two filler faces at the start and two at the end) was followed by a 48-item recognition test that included 16 old faces, 16 conjunctions, and 16 entirely new faces. The stimuli were shown on a computer screen Component Structure 14 for 5 s each, with an interstimulus interval of 2 s at study and 3 s at test. Both experiments included factorial, between-groups manipulations of presentation frequency (one vs. three study-list presentations) and test delay (10 minutes vs. 24 hours), with 24 participants receiving each combination. The study and test lists were constructed such that half the faces in each condition were originals while the rest were synthetics. Most importantly, half of the “conjunctions” on the recognition test were actually original photographs. They served as conjunctions because their inner and outer features were seen at study, as parts of different faces (the synthetics from their quadruplets). Thus, any subtle difference between the original and synthetic faces, such as a difference in naturalness of appearance, could not be used as a sound basis for recognition judgments. Following presentation of the study list of faces, participants were asked to make “old” or “new” judgments to faces on the recognition test using a six-point confidence scale (on which 1 = “sure new,” and 6 = “sure old”). Participants were shown examples of conjunction faces and were instructed to reject them as “new.” Results and Discussion Face-orientation was varied across the two experiments, but, for ease of exposition, we treat inversion as a between-groups variable along with presentation frequency (one vs. three study-list presentations) and test delay (10 minutes vs. 24 hours). We first report how inversion, presentation frequency and test delay affected recognition performance, and then turn to the PCA results. Considering our large sample (n = 192), we set alpha level for all statistical tests at .01. Component Structure 15 Hit and false-alarm rates. As shown in Figure 5, the hit rate for old faces exceeded the false-alarm rate for conjunctions, which in turn exceeded the false-alarm rate for entirely new faces (M’s = .74, .52 and .29, respectively). The figure suggests that while hit rates were more sensitive to presentation frequency, false-alarm rates were more sensitive to inversion. Supporting this interpretation, an analysis of variance (ANOVA) revealed a main effect of item-type [F (2,368) = 504.3, MSe = .02], a main effect of inversion [F (1,184) = 46.0, MSe = .04] and reliable interactions between itemtype and inversion and item-type and presentation frequency [F’s (2,368) = 68.0 and 31.7, respectively, MSe’s = .02]. Follow-up ANOVAs revealed that hits were affected by both presentation frequency and inversion, though the frequency effect was considerably stronger [F’s (1,184) = 39.4 and 12.2, respectively, MSe’s = .02]. Conjunction false alarms were affected by inversion [F (1,184) = 47.3, MSe = .03], but not by presentation frequency, while new-face false alarms were affected by inversion [F (1,184) = 105.5, MSe = .03] more strongly than by frequency [F (1,184) = 14.2, MSe = .03]. We note that the original ANOVA produced a weak item-type × test-delay interaction [F (2, 368) = 5.19, MSe = .02], but test delay was without reliable effects in the follow-up ANOVAs. PCA. Table 1 shows the intercorrelations among our three basic measures (hit rates for old faces, false-alarm rates for conjunctions and false-alarm rates for new faces) across the two experiments and between-groups conditions. It also shows the loadings of each measure on each of the two components revealed by the PCA. Except where noted, components are not discussed unless their eigenvalues exceed 1.0. The two components in Table 1 had eigenvalues of 1.46 and 1.15. Component Structure 16 The loadings of our measures on the two components were very similar to those observed in our reanalysis of the Searcy et al. (1999) data (compare top and bottom graphs in Figure 3). As before, Factor 1 showed a near-0 loading for hits, along with strong negative loadings for both types of false alarms. And, once again, Factor 2 showed a strong positive loading for hits, a smaller positive loading for conjunction false alarms, and a negative loading for new-face false alarms. Figure 6 shows plots of hit and false-alarm rates for 8 subsets of 24 participants each, grouped according to their Factor 1 scores (left graph) and Factor 2 scores (right graph). As with the Searcy et al. (1999) data, Factor 1 scores were strongly related to false-alarm errors with both conjunctions and new items, but they were not systematically related to the conjunction/new difference in false alarms. By contrast, Factor 2 scores were strongly related to hits and to the conjunction/new difference in false alarms: This difference rose from .01 in the lowest Factor 2 group to .52 in the highest Factor 2 group. Supporting this observation, we found a robust correlation between individual participants’ Factor 2 scores and the conjunction/new difference in their false-alarm rates (r = .64, df = 190), and a near-0 correlation between Factor 1 scores and the conjunction/new difference (r = .11, df = 190, ns). Factor scores and signal-detection measures. To determine if the PCA components might be characterized as criterion and discrimination, we computed d' and C scores (see Macmillan & Creelman, 2005) from each participant’s hit rate for old faces and false-alarm rate for new faces, and then computed the correlation of each measure with the participants’ Factor 1 and Factor 2 scores. As can be seen in Table 2 (columns 1 and 2), the old/new d' scores were positively (and reliably) correlated with Factor 1 and Component Structure 17 2 scores (with df = 190, | r’s | ≥.19 are reliable at p < .01). The C scores were also reliably correlated with both factor scores, though in this case the correlation was positive for Factor 1 and negative for Factor 2. The same pattern was obtained when the d' and C scores were derived from hit rates for old faces and false-alarm rates for conjunctions (Table 2, columns 3 and 4). We conclude that neither component can be characterized as exclusively “criterion” or “discrimination.” Effects of inversion on factor scores. To determine if one of our two components reflected the use of orientation-specific configural processing, we examined the distribution of Factor 1 and Factor 2 scores from the 96 participants from Experiment 1a (upright presentation) and the 96 participants from Experiment 1b (inverted presentation). The outcome is shown in Figure 7, which displays Factor 1 and 2 scores for participants in the one-presentation condition (triangles) and three-presentations conditions (circles) of Experiment 1a (upright faces/filled symbols) and Experiment 1b (inverted faces/empty symbols). The large symbols are the mean factor scores for each of the four conditions. The empty symbols generally fall to the left of the filled symbols, indicating that inversion lowered Factor 1 scores. However, the empty symbols also tend to fall lower on the graph than do the filled symbols, indicating that inversion reduced Factor 2 scores as well. ANOVAs confirmed that inversion reduced both Factor 1 scores [F (1,184) = 113.7, MSe = .61] and Factor 2 scores [F (1,184) = 11.1, MSe = .79], though the former effect was stronger. By contrast, study-list repetition increased Factor 2 scores [F (1,184) = 39.2, MSe = .79], without reliably affecting Factor 1 scores [F (1,184) = 4.5, MSe = .61, p > .10]. Component Structure 18 In summary, Experiments 1a and 1b replicated the finding that two separate components underlie performance in the face conjunction paradigm. One component was linked to false alarms (with both conjunctions and new items) and appeared unrelated to the conjunction effect (i.e., the conjunction/new difference in false alarms). By contrast, a second component was strongly related to hits as well as to the conjunction effect. Both components were related to discrimination and criterion, and so it is impossible to characterize either component as exclusively discrimination or criterion. Finally, inversion affected Factor 1 scores more strongly than Factor 2 scores. Although this finding might be viewed as supporting the hypothesis that Factor 1 reflects configural processing, the findings of Experiment 2 undermine the hypothesis. Experiments 2a and 2b These two experiments further examined the effects of inversion as well as studylist repetitions in the conjunction paradigm (Experiment 2a was briefly described in Bartlett et al., 2003). Participants viewed a study list in which eight faces appeared one time, and eight additional faces appeared eight times. The subsequent test list included four old faces that had appeared once at study, four old faces that had appeared eight times at study, four conjunctions whose “parents” had appeared once at study, and four conjunctions whose “parents” had appeared eight times at study. It also included eight entirely new faces. Half of the participants in each experiment saw all faces upright, while the remainder saw all faces inverted (hence, orientation was a true independent variable). The experiments assessed the generality of the two-component outcome across two very different levels of learning (one versus eight presentations at study), and they Component Structure 19 provided a second look at the effects of inversion on Factors 1 and 2. In addition, Experiment 2a included a manipulation of retroactive interference to examine how forgetting might alter Factors 1 and 2. Experiment 2b examined whether proximal presentation of the parents of conjunctions might alter the conjunction effect, and affect the factor scores. Method The participants in both experiments (n’s = 112 and 96, respectively) were undergraduate students at the University of Texas at Dallas (approximately 67% female) who participated as one alternative means of fulfilling a course requirement. The stimulus set comprised 12 quadruplets, each consisting of two original and two synthetic faces. The study list included the two originals from each of four quadruplets and the two synthetics from each of four additional quadruplets, totaling 16 faces in all. Half of the faces were presented eight times in the study list, with at least two faces separating a study-list face and its nearest repetition. The remaining study-list faces were presented just once. The 24-item recognition test included: (a) four old faces shown one time at study (1X-old condition), (b) four old faces shown eight times at study (8X-old condition), (c) four conjunctions whose parents were shown one time at study (1Xconjunction condition), (d) four conjunctions whose parents were shown eight times at study (8X-conjunction condition), and (e) eight new faces drawn from four quadruplets that were not used at study. While the faces that appeared in the new condition were the same for all participants, counterbalancing insured that each of the 16 faces that appeared in the old and conjunction conditions at test served equally often in the 1X-old, 8X-old, 1X-conjunction, and 8X-conjunction conditions. Component Structure 20 In each experiment, participants were randomly assigned to one of four experimental groups that were defined by the factorial combination of orientation (all faces shown either upright or inverted) and one additional independent variable. In Experiment 2a, this variable was retroactive interference. Half of the participants viewed an interference set of 48 faces—similar in appearance to the study-list faces—immediately following the study list. The other half of the participants took a vocabulary test requiring the same amount of time (six minutes). In Experiment 2b, the independent variable was parent proximity. In the proximal condition, the two study-list faces taken from each quadruplet were always presented in successive list positions. Therefore all the conjunctions in the subsequent test were proximal conjunctions (i.e., their parents has been studied close together). In the nonproximal condition, the two study-list faces from a given quadruplet were never presented in successive list positions, and thus all the conjunctions in the subsequent test were non-proximal conjunctions. Parent proximity can inflate conjunction errors (Hannigan & Reinitz, 2000; Kroll et al., 1996; McKone & Peh, 2006; Underwood, Kapelak, & Malmi, 1976), but this effect is found primarily with simplified face stimuli, and when the proximal condition entails either simultaneous (e.g., side-by-side) or alternating presentation (A, B, A, B; C, D, C, D . . ) of the conjunctions’ parents (see Jones et al., 2006). The stimuli used here were full-face photographs, and the proximal condition involved merely proximal presentation—not simultaneous or alternating presentation—of the conjunctions’ parents. Thus, we did not obtain a standard proximity effect. However, proximity interacted with facial inversion, clarifying the relation of configural processing to Factors 1 and 2. Component Structure 21 The facial stimuli were viewed on a 50-cm-wide Panasonic video monitor at a distance of approximately 1.5 m. All other aspects of design and procedure were the same as in Experiment 1. Results and Discussion Hit and false-alarm rates. Table 3 displays hit rates for old items presented one versus eight times at study (1Xand 8X-old items), along with false-alarm rates for conjunctions whose parents had been viewed one versus eight times at study (1Xand 8X-conjunctions), and false-alarm rates for entirely new faces. The data from Experiment 2a are collapsed over the retroactive interference condition (which had minimal effects), but the data from Experiment 2b are shown separately for the proximal and non-proximal conditions. Both hits and false alarms rose with presentation frequency, though the increase from the 1X condition to the 8X condition was greater with hits (M difference = .36) than with conjunction false alarms (M difference = .15). Thus, repetition at study served to strengthen the conjunction effect, but at the same time, it increased discrimination between old faces and conjunctions. 4 Both of these repetition effects appeared with inverted faces (M differences = .35 and .17, respectively) and upright faces (M differences = .37 and .12, respectively). The effects of parent proximity in Experiment 2b are shown at the bottom of Table 3. As expected conjunction false alarms were more frequent in the proximal condition than in the non-proximal condition. Puzzlingly, however, the effect occurred only with upright faces, and there, the proximal condition showed an increase in hits as well as conjunction false alarms. Component Structure 22 A useful perspective on this unexpected pattern can be gained by examining the effects of inversion in Experiment 2a, in the methodologically similar non-proximal condition of Experiment 2b, and in the proximal condition of Experiment 2b (see rows 3, 6 and 9 of Table 3). Note that in the first two cases, the effect of inversion is largely restricted to false-alarm rates: Inversion increased conjunction false alarms as well as new-face false alarms with minimal effects on hits. By contrast, in the proximal condition of Experiment 2b, inversion largely reduced hits. In light of much evidence that inversion impairs processing of configural information, the pattern suggests that configural information can be used either to reduce false alarms or increase hits, depending on conditions. The PCA presented below substantiates this claim. Before turning to the PCA, we note that all preceding observations were statistically supported. An ANOVA of hit rates in Experiment 2a supported the main effect of presentation frequency [F (1,108) = 114.0, MSe = .06], while an ANOVA of false-alarm rates supported an effect of inversion [F (1,108) = 32.6, MSe = .09] as well as frequency [F (2,216) = 62.4, MSe = .04], with each pairwise comparison of frequency conditions reliable at p < .01 (Tukey’s HSD test). In Experiment 2b, the ANOVAs showed proximity × inversion interactions with both hits [F (1,92) = 6.64, MSe = .05], and false alarms [F (1,92) = 8.84, MSe = .09]. In the non-proximal condition, an ANOVA of hits supported the main effect of presentation frequency [F (1,46) = 71.2, MSe = .06], while an ANOVA of false-alarm rates supported an effect of inversion [F (1,46) = 19.7, MSe = .08] as well as frequency [F (2,92) = 11.7, MSe = .04], just as in Experiment 2a. However, in the proximal condition, the ANOVA of hits showed main effects of both inversion [F (1,46) = 18.3, MSe = .06] and frequency [F (1,46) = 49.0, MSe = .05], while Component Structure 23 an ANOVA of false alarms showed no effect of inversion [F (1,46) < 1] along with a strong effect of frequency [F ( 2,92) = 31.7, MSe = .04]. PCA. Table 4 shows the intercorrelations among our five basic measures (rates of “old” judgments for 1X-old faces, 8X-old faces, 1X-conjunctions, 8X-conjunctions, and new faces) across the two experiments and between-groups conditions. It also shows the loadings of each measure on each of the two components emerging from the PCA (eigenvalues = 1.88 and 1.14 for Factors 1 and 2, respectively). Figure 8 displays the factor loadings in a way that facilitates comparison with the data from Experiment 1. It is clear that the factor loadings for 8X-old faces, 8X-conjunctions and new faces nicely replicated the pattern that we previously obtained (see Figure 3). By comparison, the loadings for 1X-old faces, 1X-conjunctions and new faces produced a degraded pattern, suggesting that the appearance of the two components depends on some requisite level of learning. Effects of inversion on factor scores. Figure 9 plots the Factor 1 and 2 scores for each participant in both the upright (filled symbols) and inverted (empty symbols) conditions. In general, inversion had stronger effects on Factor 1 scores than on Factor 2 scores, as in Experiment 1. However, in the proximal condition of Experiment 2b (represented by circles), Factor 1 scores were generally low, and inversion affected only Factor 2 scores. ANOVAs confirmed that, in Experiment 2a, inversion affected both Factor 1 scores [F (1,108) = 32.7, MSe = .81] and Factor 2 scores [F (1,108) = 8.75, MSe = .80], though the former effect was stronger. By contrast, in Experiment 2b, there was an inversion × proximity interaction in Factor 1 scores [F (1,92) = 8.51, MSe = .88], and, at p < .05, in Factor 2 scores as well [F (1,92) = 5.24, MSe = .88, p < .03]. Follow-up Component Structure 24 ANOVAs of the Factor 1 scores showed that the inversion effect was reliable in the nonproximal condition [F (1,46) = 15.4, MSe = .75], but not in the proximal condition [F (1,46) < 1]. Follow-up ANOVAs of the Factor 2 scores showed precisely the opposite pattern. There, the inversion effect was reliable in the proximal condition [F (1,46) = 20.7, MSe = .71], but not in the non-proximal condition [F (1,46) < 1]. These data show that inversion can affect Factor 1 and/or Factor 2, depending on conditions. The outcome undermines the view that either component is exclusively related to orientation-specific configural processing. Further evidence on this point comes from two follow-up PCAs, one on the data from the upright condition and another on the data from the inverted condition. Each PCA identified two components with eigenvalues exceeding 1 (1.79 and 1.12 in the upright condition, 1.90 and 1.11 in the inverted condition). Because the pattern of factor loadings was generally degraded in the 1X condition (see Figure 8), we show only the factor loadings from the 8X and new-item conditions in Figure 10. It is clear that the two-component outcome obtained in the overall analysis and in prior studies generalizes nicely from upright faces to inverted faces. We conclude that both components can be weakened by inversion, so that neither is uniquely linked to orientation-specific configural information. Experiment 3 Experiments 1 and 2 replicated the two-component outcome of our initial PCA. Further, they established that neither component could be uniquely and invariably linked to discrimination, criterion, or orientation-specific configural processing. In Experiment 3, we asked whether Factors 1 and 2 might be linked to the two core concepts of dualComponent Structure 25 process theories of recognition memory: recollection and familiarity (e.g., Jacoby, 1991, 1999; Diana, Reder, Arndt & Park, 2006; Yonelinas, 2002). By the dual-process account, familiarity is the non-specific feeling that a stimulus has been encountered before. It varies only in strength, providing no information on the context in which the stimulus was encountered. By contrast, recollection involves conscious retrieval of more detailed information about one’s prior encounter of the stimulus, and can specify the context of this encounter. There is evidence that familiarity falls from old items to conjunctions, and from conjunctions to new items (Jones & Jacoby, 2001; Jones & Bartlett, 2009). Thus, it is reasonable to suggest that Factor 2—which shows a graded pattern of loadings for the three item-types—reflects familiarity. It also is plausible that Factor 1 might reflect recollection, but the experiment did not directly test this idea. 5 To test the dual-process hypothesis for Factors 1 and 2, we introduced into the study another kind of lure, which we refer to as a “familiarized lure” (FL). FLs are testlist faces that were seen previously by participants, but outside of the context of the study list (e.g., Jennings & Jacoby, 1997). Because FLs differ from study-list faces primarily with respect to their presentation context, a context-free familiarity component should respond to such lures much like truly old faces. Hence, if Factor 2 reflects context-free familiarity, it should show a strong loading for FL false alarms, in the same direction as the loading for hits. The FLs in Experiment 3 were exact repetitions of new faces presented several items earlier in the recognition test (cf. Jennings & Jacoby, 1997). Participants were instructed to reject these FLs as “new.” Pilot testing indicated that FL false alarms would be comparable in frequency to conjunction false alarms. Component Structure 26 Method Ninety-six participants were recruited as in Experiment 2. The method was the same as that of Experiment 1 except that: (a) the materials were made from a new set of 180 monochromatic, full-frontal view photographs of faces of male and female students taken five or more years before the study (see Figure 2), (b) the study list consisted of 120 faces (plus two fillers at the start and two at the end), and it was presented three times, each time in a different randomized sequence, (c) the stimuli were viewed as a PowerPoint slideshow on a 46-cm-wide computer monitor, with study-list faces shown for 2 s each and followed by a 2 s interstimulus interval, and test-list faces shown for 2 s each followed by a 3 s inter-stimulus interval, and (d) the test list consisted of 240 items including 60 old faces, 60 conjunctions, 60 new faces, and 60 familiarized-lures (FLs). Each FL was an exact repetition of a new face presented 9 to 12 items earlier in the test. The sequencing of faces in the test list was randomized with the constraints that (a) each test half included equal numbers of faces in all four conditions, and (b) the last new item appeared at least ten items from the end of the list so that it was possible to repeat the item (to make a FL) with a lag of at least nine. As in Experiments 1 and 2, the test instructions were to endorse exact matches to study-list items as “old” and reject all other types of item as “new” using a six-point confidence scale. Results and Discussion Hit and false-alarm rates. The upper panel of Figure 11 displays the hit rates for old faces and false-alarm rates for conjunctions, new faces, and FL faces. As in prior studies, hits were more frequent than conjunction false alarms, which in turn were more frequent than new-face false alarms. FL false alarms occurred at about the same rate as Component Structure 27 conjunction false alarms. An ANOVA of the proportion of “old” responses produced a strong main effect of item-type [F (3, 285) = 215.1, MSe = .02]. All pairwise differences were reliable by Tukey’s HSD test except for the difference between conjunctions and familiarized lures. PCA. The upper half of Table 5 displays the intercorrelations among dependent variables along with their loadings on the first two principal components (eigenvalues = 2.19 and 1.14, respectively). Figure 12 (upper portion) displays these loadings, along with the mean hit and false-alarm rates for 8 subsets of 12 participants when grouped by their Factor 1 scores, and also when grouped by their Factor 2 scores. As in prior studies, Factor 1 emerged as a false-alarm-rate component with strong negative loadings for conjunction false alarms (-.85) and new-face false alarms (-.85). Factor 1 also showed a negative loading for FL false alarms (-.76), further supporting its status as a false-alarmrate component. We note also a moderate negative loading for hits (-.42). This is similar to the Factor 1 hit-rate loading found in the low-learning (1X-old) condition of Experiment 2, and may reflect the fact that list length was considerably longer in Experiment 3 (120 study-list faces) than in Experiments 1 and 2 (32 and 16 study-list faces, respectively). In any event, the sub-group analysis (upper-right quadrant of Figure 12) shows that only the very highest Factor 1 participants showed an appreciable reduction in hits. Factor 2 showed a strong loading for hits (.85) and moderate loadings in opposite directions for conjunction false alarms and new-face false alarms (.32 and _.36, respectively), replicating Experiments 1 and 2. As in these prior studies, the sub-group analysis (Figure 12, upper-right quadrant) reveals that Factor 2 scores were strongly Component Structure 28 related not only to hits, but also to the conjunction effect (the conjunction/new difference in false-alarm rates; compare filled squares and empty circles in Figure 12). Again, we quantified the latter relation by computing the correlation between individual participants’ Factor 2 scores and the conjunction/new difference in their false-alarm rates. As before, the correlation was robust (r = .72, df = 94, p < .01). The most important new finding in Experiment 3 was that Factor 2 showed a loading for FL false alarms (.43) that was similar to that for new-face false alarms (_.36) and reversed in sign from that for hits (.85) and conjunction false alarms (.32). As a result, the Factor 2 groupings reveal a striking crisscrossed pattern such that FL falsealarm rates exceeded conjunction false-alarm rates among the lower Factor 2 participants, while the reverse was true among higher Factor 2 participants (compare empty circles with filled squares in Figure 12). Note in addition that Factor 2 was quite strongly related to discrimination between old and FL faces, as corroborated by a .89 correlation (df = 94, p < .01) between individual participants’ Factor 2 scores and the differences between their hit rates for old faces and false-alarm rates for FLs. 7 This robust correlation contradicts the notion that Factor 2 is a context-free familiarity component, and suggests that this factor is linked to context retrieval. Experiment 4 Although the findings of Experiment 3 weigh against the hypothesis that Factor 2 reflects context-free familiarity, they do so only on the assumption that participants distinguished old and FL faces based on context retrieval. This assumption is plausible, but is open to question on the grounds that old and FL items may have differed somewhat in familiarity. On the one hand, old faces might have been more familiar than FLs Component Structure 29 because old items were shown three times at study while FLs had been shown only one time before (earlier in the test). On the other hand, old faces might have been less familiar than FLs because FLs were seen more recently (9 to 12 items back versus an average of 180 items back). In either case, participants may have used familiarity information to distinguish old faces from FLs, and therefore it might be argued that, after all, Factor 2 reflects familiarity. To obtain more definitive data, we sought to replicate the findings of Experiment 3 in two conditions, one which allowed a familiarity difference between old faces and FL items, and one designed to minimize any such difference. The FLs in Experiment 4 were faces viewed in a familiarization set prior to the study list (cf. Jacoby, 1999). This familiarization set was presented only once to half of the participants (1X-FL condition), but it was presented three times to the others (3X-FL condition). The study list was presented three times to all participants. In both conditions, the recognition test included study-list faces and FLs, along with conjunctions and entirely new faces. Our thinking was that, in the 1X-FL condition, the old/FL discrimination might be based to some extent on familiarity differences, as the FL items were shown only once before the test while the old items were shown three times. In the 3X-FL condition, however, familiarity should be approximately the same for FL and old items (both presented three times). Hence, we assumed that the old/FL discrimination in the 3X-FL condition would be based primarily on context retrieval. To test this assumption, we planned to examine receiver operating characteristics (ROCs) relating hits and false alarms across five levels of criterion in each experimental condition. Although ROCs usually have a curvilinear, convex shape, Yonelinas (1997, 1999) found linear ROCs Component Structure 30 when recognition judgments were based on context retrieval. Based on this work, we expected linear ROCs for hits-versus-FL false alarms in the 3X-FL condition though not necessarily in the 1X-FL condition. The key question was whether Factor 2 would show a robust correlation with old/FL discrimination not only in the 1X-FL condition (where familiarity might support discrimination to some degree), but also in the 3X-FL condition (where its contribution should be minimal). A strong correlation in the 3X-FL condition would indicate that Factor 2 is linked to context retrieval. Method The 144 participants in Experiment 4 were recruited as in Experiment 3. The design and procedure were similar as well, except that: (a) the study and test lists were shorter (64 and 128 faces, respectively, as opposed to 120 and 240, respectively, in Experiment 3), and (b) the study list (plus one filler at the start and one at the end) was preceded by a familiarization set of 32 faces. The familiarization set was shown one time to half of the participants (1X-FL condition) and three times (in different randomized orders) for the remainder (3X-FL condition). The familiarization task in the 1X-FL condition was to rate faces for trustworthiness, while the task in the 3X-FL condition was to rate the faces for trustworthiness during the first presentation, attractiveness during the second presentation, and intelligence during the third presentation. The study list was presented three times to all participants (in a different sequential order each time) and the task was simply to study the faces (as in Experiment 3). The subsequent test included 32 faces of each of four types (old, conjunction, new, and FL), presented in a randomized order. The task was to judge whether each test face was or was not from the study list, using a six-point confidence scale. Component Structure 31 Results and Discussion Hit and false-alarm rates. The lower panel of Figure 11 displays the hit rates for old faces and the false-alarm rates for conjunction, new and familiarized-lure (FL) faces in the 1X-FL and 3X-FL conditions. As in Experiment 3, hits were generally more frequent than false-alarms, but false alarms were more frequent in response to conjunctions and FLs than in response to entirely new faces. The findings were similar in the 1X-FL and 3X-FL conditions, but note that the 3X-FL condition produced slightly lower hit rates for old faces and slightly higher false-alarm rates for FLs. The pattern suggests that, when the familiarity difference between old faces and FLs was minimized, discrimination between the two item-types was slightly impaired. To test this observation, we computed d' scores for discrimination between old faces and each of the other item-types. Old/FL discrimination was reliably higher in the 1X-FL condition than in the 3X-FL condition [M d' s = 1.24 and .69, SDs = .90 and 1.13, respectively; F (1, 42) = 10.3, MSe = 1.05]. By contrast, old/conjunction discrimination and old/new discrimination were not reliably affected by experimental condition. Converging evidence that the familiarity difference between old and FL faces was minimized in the 3X-FL condition comes from a ROC analysis. Using each participant’s recognition-confidence ratings, we computed hit-rates for old faces, and false-alarm rates for conjunctions, and false-alarm rates for FLs at each of five criterion levels, and the means across participants were used to plot ROC curves for old/conjunction discrimination and for old/FL discrimination in the 1X-FL and 3X-FL conditions. As shown in Figure 13, Old/FL discrimination in the 3X-FL condition was associated with a linear ROC, suggesting that it was based largely on recollection (context retrieval). By Component Structure 32 contrast, ROCs for old/FL discrimination in the 1X-FL condition and for old/conjunction discrimination in both 1X-FL and 3X-FL conditions were curvilinear, indicating that familiarity played at least a partial role. To quantitatively evaluate these findings, we compared the fit of the four grouplevel ROCs for two different two-parameter models, one assuming that familiarity is the sole basis of performance, and a second assuming that only recollection is involved. The familiarity-only model estimated parameters for d’ and the ratio the target and lure distribution variance, whereas the recollection-only model estimated parameters for recollection-based acceptance of targets and recollection-based rejection of lures (see Yonelinas & Parks, 2006 for the excel-based model-fitting program). The familiarityonly model produced a better fit than did the recollection-only model for old/FL discrimination in the 1X-FL condition (SSE = .00004 vs. .002), and for old/conjunction discrimination in the 1X-FL condition (SSE = .00007 vs. .002) and in the 3X-FL condition (SSE = .0001 vs. .002). By contrast, the recollection-only model produced a better fit for old/FL discrimination in the 3X-FL condition (SSE = .00003 vs. .006), suggesting that recollection was a major basis of performance in that condition. We also fitted the data to a three-parameter dual-process model that provided estimates of recollection-based acceptance, recollection-based rejection and familiarity (d’). The fits to the three-parameter model were good (SSE = .00001 and.000005 for old/FL discrimination in the 3X-FL and 1X-FL conditions, and .003 and .00008 for old/conjunction discrimination in the 3XF and 1XF conditions), and the parameter estimations were in line with expectations. Specifically, for old-FL discrimination in the 1X-FL condition, and old-conjunction discrimination in the 3X-FL and 1X-FL Component Structure 33 conditions, the d’ estimates were .79, .65, and .91, respectively. The recollectionacceptance estimates were .29, .22 and .13, respectively, and the recollection-rejection estimates were essentially null (.005 or less). By contrast, for old-FL discrimination in the 3X-FL condition, the familiarity estimate was near 0 (and numerically negative, d’ = -.10), whereas both recollection estimates were substantial (.21 for recollectionacceptance and .30 for recollection-rejection). Because averaging artifacts can affect group-level ROCs, we fitted the dual-process model to the old-FL data from each participant who showed at least a 4-point old-FL ROC and above-chance old-FL discrimination (n = 42 and 54 for the 3X-FL and 1X-FL conditions, respectively). The estimates of d’, recollection-acceptance, and recollection-rejection averaged .31, .36, and .14, respectively, in the 1X-FL condition and -.18, 26, and.36, respectively, in the 3X-FL condition (the SSE values averaged .002 in both cases). All parameters were reliably greater than 0 (p’s < .000001 by t test) except for the d’ parameter in the 3X-FL condition which was numerically negative (M = -.18) and not different from 0 (t (41) = 1.28, p > .20). PCA. The lower half of Table 5 displays the intercorrelations among the hit-rate and false-alarm-rate measures along with the factor loadings of each measure on the first two components emerging from the PCA (eigenvalues = 1.71 and 1.19, respectively). The lower-left quadrant of Figure 12 displays the factor loadings, and the lower-right quadrant displays hit and false-alarm rates for eight sub-groups of 18 participants each, sorted by their Factor 1 and 2 scores. The pattern is similar to that of Experiment 3, although floor effects appear to have weakened the relation between Factor 2 scores and the conjunction effect. Again, we quantified this relation by computing the correlation Component Structure 34 between individual participants’ Factor 2 scores and the conjunction/new difference in their false-alarm rates. The correlation was reliable (r = .40, df = 142, p<.01), albeit weaker than in Experiment 3 (where r = .72). The most important finding of this study is that it replicated the negative relation between Factor 2 scores and FL false alarms, and did so in the 3X-FL condition as well as in the 1X-FL condition. As shown in Figure 14, separate PCAs of the 1X-FL condition and the 3X-FL condition (n = 72 in each case) produced highly similar twocomponent outcomes (eigenvalues of 1.94 and 1.23 for Factors 1 and 2, respectively, in the 1X-FL condition, and 1.60 and 1.10, respectively, in the 3X-FL condition). Both PCAs produced the crisscrossed relation between conjunction false alarms and FL false alarms across the Factor 2 groups (see right side of upper-right and lower-right graphs in Figure 14). Further, in both conditions there was a strong correlation between Factor 2 scores and old/FL discrimination. Across the 72 participants in each condition, the correlation of Factor 2 scores with the old/FL difference in “old” judgments was .90 in the 1X-FL condition and .91 in the 3X-FL condition. In summary, Experiment 4 replicated the two-component outcome obtained in prior experiments. The key new finding was that old/FL discrimination in the 3X-FL condition showed a linear ROC that was well fit by a recollection-only model that constrained familiarity to be 0 as well as by a recollection-plus-familiarity model that estimated familiarity to be close to 0. In light of this finding, the strong correlations between Factor 2 scores and old/FL discrimination in the 3X-FL condition show that Factor 2—our hit-rate (and conjunction-effect) component—is strongly linked to context retrieval. Component Structure 35 General Discussion Two Components of Facial Recognition Memory The present research provides strong evidence that face recognition relies on two distinct processing components, one that is related to false recognition and another that is linked to correct recognition as well as a false memory phenomenon known as the conjunction effect. Experiments 1 through 4 and the Searcy et al. (1999) study differed from each other in several respects, including age range of participants, stimulus materials, list length, and experimental design. Yet, PCAs of all five data sets consistently returned a two-component outcome, with the two components accounting for at least 58% of the variance in the recognition performance. The first component (Factor 1) accounted for 36% to 55% of the variance across the data sets, and it was strongly related to false-alarm errors in response to conjunctions as well as new faces. Factor 1 loadings for these two lure-types were virtually identical in our reanalysis of the Searcy et al. (1999) data (_.79 and _.80, respectively), in Experiment 1 (_.85 and _.86, respectively), in Experiment 3 (_.85 and _.85, respectively), and in Experiment 4 (_.86 and _.80, respectively). Perhaps most strikingly, Factor 1 loadings in Experiment 2 were _.75, _.77 and _.77 for new faces, conjunctions in the onepresentation (1X-conjunction) condition, and conjunctions in the eight-presentations (8Xconjunction) condition, respectively. This invariance in loadings occurred despite large differences in average false-alarm rates among the three item-types (M’s = .18, .35, and .46, respectively, see Table 3). In Experiments 3 and 4, we included familiarized lures (FLs), which were falsely recognized approximately as often as conjunctions. Factor 1 loadings for FL false alarms were _.76 and _.52 in these two experiments. Although these Component Structure 36 loadings were somewhat weaker than the loadings for conjunctions and new faces, the data are largely in line with our conclusion that Factor 1 is linked to false alarms. Underscoring this conclusion, Factor 1 loadings for hits were extremely low in the Searcy et al. (1999) data (.01), in Experiment 1 (.02), with eight-times presented (8X-old) items in Experiment 2 (.14), and in Experiment 4 (_.26). The loadings for hits were somewhat stronger for the poorly learned, once-presented (1X-old) items in Experiment 2 (_.39), and in Experiment 3 in which the study list and recognition test were extremely long (_.42). Presumably, when learning is low and/or forgetting is high, Factor 1 might degenerate into a pure criterion factor, as the moderate loadings for hits in the same direction as false alarms suggest. The second component emerging from the PCAs was also substantial, accounting for 22% to 39% of the variance across the five data sets. Our suggestion that this component is related to hits was consistently supported: The loadings of hit rates on Factor 2 were always .85 or higher except in the case of the poorly learned oncepresented (1X-old) items in Experiment 2, for which the hit-rate loading was .47. Factor 2 was related not only to hits, but also to conjunction false alarms and newface false alarms, though in opposite directions. Although these false-alarm loadings were moderate in size, their sign reversals for the two lure-types produced an interesting effect: Factor 2 scores were consistently related to the conjunction effect. The Factor 2/conjunction-effect correlation was .65 in the Searcy et al. (1999) data, and .64, .49, .72 and .40 in Experiments 1 through 4, respectively. Though not as strong as the hit-rate loadings, these correlations were substantial and statistically reliable. Further, the Component Structure 37 relatively low correlation of .40 in Experiment 4 can be attributed to a floor effect on new-face false alarms. The Rapcsak et al. (1999) Model and Alternative Conceptions of the Two Components Our PCAs supported the existence of two separable components that we have referred to, for convenience, as a false-alarm-rate component and a hit-rate component. It is critical, however, to move beyond mere labeling to characterize these components in terms of theory. We began this research with the hypothesis that the two components emerging from our PCAs reflected the contributions of the frontal and occipito-temporal components of Rapcsak et al.’s (1999) neuropsychological model of face recognition memory (Figure 1). The hypothesis has faired well, as the PCA findings suggest a component that—like Rapcsak et al.’s (1999) “frontal executive” component—serves to reduce false alarms. These findings also suggest another component that — like Rapcsak et al.’s occipito-temporal “face recognition module” —works to increase hits. We find it impressive that our first component was related to false-alarm errors with three different types of lure (conjunctions, FLs and new faces). Because it is unlikely that all three types of lure were rejected through the same process, this outcome suggests that the component reflects control of false-alarm errors through a number of different processes, as does the frontal executive component in the Rapcsak et al. (1999) model. That Factor 1 approximated a criterion factor under conditions of low learning and high interference is also consistent with the model. According to the model, the setting of recognition criteria is one way that the frontal executive component works to limit false alarms. In conditions of low learning and/or high interference, it may be the most important way. Component Structure 38 The Rapcsak et al. (1999) model also captures our finding that the hit-rate component was related not only to hits, but also to conjunction false alarms and FL false alarms in opposite directions. A core claim of the model is that the occipito-temporal component—the face recognition module—can trigger bottom-up retrieval of contextual information. Such information is likely to be helpful in accepting old test items, and also in rejecting FL items (which were encountered outside of the study-list context), but not in rejecting conjunctions (whose parents were encountered within the study-list context). Thus, a face recognition component that triggers bottom-up retrieval of contextual information provides an unforced explanation of why Factor 2 showed a strong positive relation to hits, a negative relation to FL false alarms, and a positive relation to conjunction false alarms. Three alternative characterizations of our two PCA components were considered in the course of this research. One characterization was that these two components reflect discrimination and criterion, as these concepts are defined in signal detection theory. However, contrary to this view, Experiment 1 revealed that each component was related to both aspects of performance. Another characterization was that one of our components reflected configural processing whereas the other reflected featural processing. Based on much evidence that inversion impairs configural processing more than featural processing (see Bartlett et al., 2003, for a review), this hypothesis predicts that one of our components should be weakened by inversion more than the other. However, contrary to this hypothesis, we found that neither component was exclusively sensitive to facial orientation, and that both components emerged from recognition of inverted faces. Still a third characterization that we considered was that our two PCA Component Structure 39 components reflected the familiarity-recollection distinction in dual-process theories of recognition memory (Jacoby, 1991, 1999), with Factor 2 reflecting context-free familiarity and Factor 1 reflecting recollection (specifically, the use of recollection to control false alarms). The hypothesis predicts that Factor 2 should show strong and same-direction loadings for (a) correct recognitions of study-list faces (hits), and (b) false recognitions of familiarized lures (FLs). Contrary to the prediction, Factor 2 produced strong and reversed-sign loadings for hits and FL false alarms, and Factor 2 scores were strongly correlated with measures of old/FL discrimination (r’s ≈ .90), even in conditions that minimized the contribution of familiarity to such discrimination (i.e., in the 3X-FL condition of Experiment 4). We did not directly test the notion that Factor 1 is a recollection-rejection component. However, recollection-rejection of facial conjunctions appears to be minimal, even in conditions designed to maximize its occurrence (Jones & Bartlett, 2009). Apart from the hypotheses that we initially considered, several additional conceptions might be applied to our data and deserve consideration. One of these is fuzzy trace theory (Brainerd & Reyna, 2005), an account that distinguishes verbatim representations of perceptual information from gist representations of meaning. Brainerd and Reyna (2005) proposed that gist representations for faces contain categorical information pertaining to race, gender and general body build. The authors did not specify the nature of “verbatim” representations of faces, but presumably they contain highly detailed information about facial features and configuration, as well as itemspecific episodic information. Accepting this characterization, fuzzy trace theory nicely handles our finding that Factor 1 (the verbatim component?) was linked to discrimination Component Structure 40 between old faces and three different types of lure (conjunctions, new faces, and familiarized lures). The theory is also in line with the finding that Factor 2 (the gist component?) was positively related to the false recognition of conjunction items because, presumably, conjunctions match well with gist information. Despite its strengths, however, the theory requires elaboration to explain why Factor 2 was negatively related to false recognition of FLs. In addition, a key claim of fuzzy trace theory is that fuzzy and verbatim traces work together to support correct recognitions. Thus, without elaboration, the theory does not explain why only one of our components showed strong loadings for hits. Another relevant theoretical conception distinguishes post-retrieval monitoring from the initial process—whatever its nature—that underlies the retrieval or activation of memories in response to a cue (Odegard & Lampinen, 2006; Schacter, 2001; Roediger & McDermott, 2000; Moscovitch, 1994). Post-retrieval monitoring is viewed as important for reducing false alarms through assessments of (a) the distinctiveness of the retrieved information (e.g., Dodson & Schacter, 2002), (b) information specifying its context or source (e.g., Roediger & McDermott, 2000), or (c) the goodness or completeness of match between a test cue and a memory trace (Reder, Wible & Martin, 1986). The idea is similar to the Rapcsak et al. (1999) conception that a frontal executive component performs such monitoring (along with other functions). Yet, like the dual-process and fuzzy-trace hypotheses, it does not explain why Factor 2 showed opposite-direction loadings for false recognitions of conjunctions and familiarized lures. Our application of the Rapcsak et al. (1999) model is preferred in this regard as it provides a simple, unforced explanation of why the hit-rate component should be positively correlated with Component Structure 41 false recognitions of conjunctions and yet negatively correlated with false recognitions of familiarized lures. A final hypothesis we consider is the inhibitory deficit theory of Hasher, Zacks, and their associates (Hasher & Zacks, 1988; Lustig, Hasher & Zacks, 2007). These researchers argue that inhibition reflects “. . . the ability to limit activation to information most relevant to one’s goals” (Lustig et al., 2007, p. 146). They distinguish inhibition from activation and argue that inhibition is the strongest source of individual differences in cognitive tasks. From the perspective of inhibitory deficit theory, it is tempting to link our falsealarm-rate component with inhibition and our hit-rate (and conjunction-effect) component with activation. If these connections are made, the theory nicely handles the finding that hits showed a strong positive loading on only one of our components. Second, since inhibition is a larger source of individual differences than activation, the theory predicts that Factor 1—our false-alarm-rate component—should not only exist but also account for more variance than does Factor 2 (our hit-rate component). None of the other aforementioned theories make this prediction, and yet it was supported in all five data sets considered in this report. These strengths notwithstanding, the theory must be elaborated if it is to explain why the hit-rate component was positively correlated with false recognitions of conjunctions and yet negatively correlated with false recognitions of familiarized lures, a finding neatly captured by our application of the Rapcsak et al. (1999). In arguing for the Rapcsak et al. model as a framework for our findings, we are not claiming that it is superior in other respects to the various alternatives we have Component Structure 42 considered. Our point is more narrow: The Rapcsak et al. model appears to handle individual differences in recognition of faces better than the alternatives. This is no mean achievement, but it is not everything. Consider our conclusion that neither of our factors reflects the process of familiarity posited by dual-process theories. While our data strongly support this conclusion, they do not rule out a familiarity process, and in fact the Rapcsak et al. model includes a familiarity component (resemblance). A reasonable idea, quite consistent with our data, is that familiarity is real and important, but does differ much across healthy individuals. Generality of the Two Components An issue attending face processing research is whether the findings and relevant models are specific to faces. In the Rapcsak et al. (1999) model, the characterization of the occipito-temporal component as a “face recognition module” implies it is specific to faces (see Fodor, 1983, for discussion of modules). Further, McKone and Peh (2006) have argued that general memory theories (which they call “memory-only” theories) do not work well with faces, as, for example, these conceptions do not explain why facial inversion impairs face recognition accuracy. In support of their argument, McKone nad Peh reported two studies using the conjunction paradigm, showing that inversion impaired discrimination between old faces and conjunctions, a finding we have also reported (Bartlett et al., 2003) and replicated here in Experiments 1 and 2. Although facial inversion has effects that are not handled by general memory theories, it is important to consider certain qualitative similarities in recognition memory for upright and inverted faces. While inversion reduced discrimination between old faces and conjunctions, it left the basic conjunction effect—the conjunction/new difference in Component Structure 43 false-alarm errors—largely unaffected (both here and in the McKone and Peh study). Moreover, the effects of presentation frequency on conjunction false alarms were impressively similar for upright and inverted faces in Experiment 2 (see Table 3). Finally, our success in replicating the dual-component outcome of our PCAs with inverted faces (Figure 10) is another instance of a qualitative similarity, and one that suggests that our dual components might generalize across stimulus domains. Further evidence that the dual-component outcome has impressive generality comes from an experiment by Rubin et al. (1999). Their young and senior participants received a study list of disyllabic words followed by a test including old and new words, together with conjunctions (each combining the first syllable of one study-list word with the second syllable of another) and “syllable” lures (each combining one old syllable with one new syllable). The seniors also took two neuropsychological test batteries, one for frontal function and the other for temporal lobe function. The upper graphs of Figure 15 compare the high and low performers on the frontal battery (upper-left graph) and the high and low performers on the temporal lobe battery (upper-right graph). Note that the frontal scores were related primarily to false-alarm errors whereas, by contrast, the temporal-lobe scores were related to hits as well as the conjunction/new difference in false alarms (false-alarm rates for syllable lures fell between those for conjunctions and entirely new items). For comparison, we have included data from our reanalysis of the Searcy et al. (1999) face recognition study in the lower graphs of Figure 15. These graphs compare participants with above and below average Factor 1 scores (lower-left graph) and above and below average Factor 2 scores (lower-right graph). These are the same data as shown in Figure 4, except that we have divided the participants into only two Component Structure 44 groups instead of six. Despite a difference between studies in the overall level of conjunction false alarms, Factor 1 is related to face recognition in the same way as frontal-test performance was related to word recognition performance in Rubin et al. (1999). Similarly, Factor 2 is related to face recognition in the same way as temporallobe-test performance was related to word recognition. A second relevant study by Stark and Squire (2003) compared patients who had suffered hippocampal damage with healthy controls in the conjunction paradigm. All participants performed five recognition memory tasks with five different types of stimuli: disyllabic words, compound words, word pairs, object pairs and face-house pairs. In each task, a study list was followed by a recognition test including old items, conjunctions, and entirely new items (it also included feature lures each composed of one old and one new part). As shown in Figure 16, the controls greatly exceeded the patients in hit rates to old items, and they also showed a stronger conjunction/new difference in false-alarm errors. 9 The pattern is similar to what we obtained when we divided the participants in Experiment 2 (or any other of our experiments) into high versus low Factor 2 groups (see lower right graph in Figure 16). We believe our analyses of these previously reported studies support the contention that the dual-component outcome of our PCAs reflect general frontal and medial temporal lobe (MTL) components that are not restricted to faces. In light of qualitative similarities between recognition of upright faces and recognition of other stimulus-types, it would appear ill advised to develop theories of face recognition without consideration of how general memory processes might be involved. Rather, we should attempt to identify and characterize those components of face Component Structure 45 recognition memory that are domain specific (used only with faces), or domain-related (relatively more important with faces than with other stimuli), and explore how these components interact with general memory processes. A proposed elaboration of the Rapcsak et al. (1999) model (see Figure 17) takes a step in this direction. A Proposed Modification of the Rapcsak et al. (1999) Model Our modification of the Rapcsak et al. model derives from recent evidence that recognition memory for upright faces is based on both configural and featural information (though the former appears to be more important in many situations, see Sergent, 1984; Cabeza & Kato, 2000; Bartlett & Searcy, 1993; Bartlett et al., 2003; Moscovitch, Winocur, & Behrmann, 1997; McKone et al., 2007), and that configural face processing involves (at least) two occipito-temporal regions (Schlitz & Rossion, 2006; Rossion, Caldara, Seghier, Schuller, Lazeyras & Mayer, 2003) that are distinguishable from the medial temporal lobe (MTL) structures that are known to be critical for recognition memory with a range of different stimuli, and that collectively support both resemblance (familiarity) and recollection (Squire, Stark & Clark, 2004; Eichenbaum, Yonelinas & Ranganath, 2007; Mayes, Montaldi & Migo, 2007). Controversy swirls as to whether the processing of upright faces is performed by a pre-wired module or is linked to expertise, and also as to whether resemblance and recollection involve separate or overlapping MTL structures. Nonetheless, there is sufficient consensus on these points to consider updating the Rapcsak et al. (1999) as shown in Figure 17. The core idea of this modified model is that face recognition involves (a) partbased processing that can support recognition of upright faces as well as objects and inverted faces, and (b) holistic/configural processing that can support recognition of Component Structure 46 upright faces much more than recognition of objects and inverted faces (see Bartlett et al., 2003; Tanaka & Farah, 2003; McKone et al., 2003). In a recognition test, part-based processing signals part-based resemblance, which is considerably higher for conjunctions (as they are composed of old parts) than for new faces. By contrast, holistic/configural processing signals configural resemblance, and this is assumed to be generally much higher for old faces (as they are exact copies of study-list faces) than for both conjunctions and new faces. Resemblance signals from both components are sent to the MTL memory component, which computes overall resemblance strength and also can trigger bottom-up retrieval of contextual information. We assume that both resemblance strength and bottom-up retrieval of context are greater for old faces than conjunctions (largely due to holistic/configural processing), and greater for conjunctions than new faces (largely due to part-based processing). The output of the MTL memory component—resemblance signals and contextual information—is monitored by the frontal executive component (FEC). The FEC can operate in either an automatic, bottom-up mode or in a controlled, top-down mode. In the bottom-up mode (signified by the upward-pointing arrows in Figure 17), the FEC derives a value of “memory strength” as a cumulative function of both bottom-up inputs from MTL (the upward-pointing arrows to the FEC in Figure 17). In the top-down mode (see downward pointing arrows), the FEC works to reduce false alarms by (a) checking the extent to which the resemblance signal reflects configural resemblance versus part-based resemblance (see arrow “A” in Figure 17), and (b) examining contextual/associative information for evidence that the various parts of a face were encountered together (i.e., in the same context, see arrow “B” in the figure). Component Structure 47 We propose that the hit-rate (and conjunction-effect) component emerging from these studies reflects primarily the amount of contextual/associative information retrieved in the bottom-up mode (though it may reflect resemblance information to some extent). We further propose that the false-alarm-rate component reflects the extent of top-down analysis of resemblance information and contextual/associative information in making recognition judgments. Participants who are strongest on this second component follow a decision strategy whereby “old” judgments are reserved for faces that: (a) produce strong configural resemblance relative to part-based resemblance, and (b) evoke retrieval of contextual/associative information specifying that various parts of a face were studied close together in time. By this modified model, one effect of facial inversion is to reduce memory strength for old faces relative to that for conjunctions and new faces. This will cause inversion to reduce hit rates if participants are functioning in the bottom-up mode. A second effect is to disable the top-down configural-match test which can aid lure rejection in the top-down mode. This will cause inversion to increase false alarms if participants are functioning in the top-down mode. We believe that these assumptions may explain why inversion had its variable effects, decreasing hits (and decreasing Factor 2 scores) in the proximal condition of Experiment 2b, and increasing false alarms (and reducing Factor 1 scores) in the nonproximal condition of Experiment 2b, in Experiment 2a, and in Experiment 1. By this account, participants in the proximal condition relied primarily on the bottom-up mode, possibly because the top-down contextual/associative test was useless for rejecting conjunctions in that condition (in which conjunctions’ parents had been studied close together). Component Structure 48 Independence of the Two Components After perusing our data, colleagues have questioned whether it is plausible that a medial temporal lobe component and a frontal executive component would be truly orthogonal as our PCAs suggest. Is not it likely that two such components would interact with each other, and therefore would be correlated? Our first response to this question is that two such components almost certainly interact, but this does not mean the corresponding individual differences are correlated. An analogy might be drawn to how a modern car’s engine interacts with its automatic transmission. An increase or decrease in RPMs will cause a gear change, and, likewise, a change of gears can affect RPMs. However, this does not mean that there must a correlation between the engine horsepower and number of gears across a set of cars. Our second response is based on a hierarchical factor analysis that we performed on the data from all five data sets. Hierarchical factor analysis (Wherry, 1984) derives oblique factors by identifying clusters of the stronger loadings emerging from a PCA, and rotating the axes in the factor space to go through these clusters. It then computes the correlations among these oblique factors and the original, primary, factors. Finally, it derives a general factor by performing a PCA on the correlation matrix of oblique factors and primary factors. Hence, the technique provides a quantitative assessment of the orthogonality of the factors, and also assesses the importance of any higher-order factor that might explain variance not accounted for by the primary factors. In the case of Experiment 4, the two oblique factors were essentially uncorrelated with each other (r = –.06), and each was almost perfectly correlated (r = .97) with one of the primary factors. Further, the general factor had only very weak loadings (.28 or less) Component Structure 49 on the dependent measures (hits to old items and false alarms to conjunction, new, and FL faces). In sum, the analysis did not identify oblique factors that differed from the primary (orthogonal) factors, and it failed to identify a general factor with substantial loadings on our dependent measures. Essentially the same outcome was obtained from the data of Searcy et al. (1999), and Experiments 1 and 2. In Experiment 3, there was some suggestion of oblique factors that were slightly correlated with each other (r = .36), and that were less than perfectly correlated with the primary factors (r’s = .80). Further, the general factor had same-direction loadings of .45 to .66 on the four dependent measures. This is in accordance with our prior conclusion that, due to the lengthy study list and recognition test, Factor 1 approximated a criterion factor in that experiment. In general, the hierarchical factor analyses did not suggest any substantive modification of the conclusions we drew from the original PCAs. Broader Implications We close by discussing three implications of the dual-component view of face recognition supported in this article. A first implication is that a model emerging from neuropsychological studies of brain-injured patients was supported by individual difference research. This implication is important because many types of brain damage might produce deficits that are dissociable from each other while having no bearing on those processes that vary independently among healthy adults. The present data stand as evidence that components like those in the Rapcsak et al. (1999) model vary independently among healthy adults, and research should examine why this is the case. A second implication concerns an interesting commonality between the Rapcsak et al. (1999) model we have favored in this article and inhibitory deficit theory (e.g., Component Structure 50 Lustig et al., 2007). Both of these models propose a component of cognition that does not represent a single process, but rather a set of correlated processes that serve a common function. The frontal executive system of Rapcsak et al. (1999) has the function of controlling false alarms through criterion setting, source monitoring and strategic search (we have proposed configural-match testing as well). In the inhibitory deficit theory, inhibition works to limit activation through controlling the focus of attention, deleting irrelevant information from working memory, and suppressing inappropriate responses. The delineated processes are presumably dissociable, but nonetheless they are correlated with each other. Rapcsak et al. (1999) characterize their frontal component as a “system,” and by some definitions, it qualifies as such. Tulving (1985), for example,, characterized a system as “a set of correlated processes” (p. 386). By other definitions, it probably does not (Schacter, Wagner & Buckner, 2000). In any case, however, there appears to be a place in contemporary theory for constructs that pertain to sets of correlated processes that serve a common function, whether they qualify as systems or not (see also Friedman, Miyake, Young, DeFries, Corley, & Hewitt, in press). A third implication is more practical than others, and more specific to the problem of face recognition: Dimensions of difference among healthy persons—such as those identified by our Factors 1 and 2—suggest targets for training and other interventions to improve face recognition. False recognitions of faces are a major problem in eyewitness testimony, and so it is important to examine whether our first component generalizes to the lineup task, and, if so, to determine whether regimens might be devised to help lowFactor-1 eyewitnesses avoid false identifications. Indeed, our PCA results may aid in understanding the variable effects of sequential presentation found in the lineup task. Component Structure 51 Author Note James C. Bartlett, Kalyan K. Shastri, and Hervé Abdi, School of Behavioral and Brain Sciences, University of Texas at Dallas. Marsha Neville-Smith, School of Occupational Therapy, Texas Woman’s University. The authors thank Catherine Truxillo for help in constructing the stimuli used in Experiments 1 and 2. Please send correspondence to James C. Bartlett, School of Behavioral and Brain Sciences, The University of Texas at Dallas, Campus Mailbox GR41, Richardson, TX 75080-3021 ([email protected]). Component Structure 52 Footnotes 1 The conjunction effect is well established with a variety of stimuli (see Jones & Jacoby, 2001; Kroll, Knight, Metcalfe, Wolf, & Tulving, 1996; Reinitz, Lammers & Cochran, 1992; Rubin, Van Petten, Glisky, & Newberg, 1999; Stark & Squire, 2003), but the effect with faces poses a puzzle because it conflicts with evidence that faces are processed configurally as wholes. McKone and Peh (2006) used simplified faces designed to maximize the role of configural-holistic processing when the faces were upright, and yet they found that the difference between hit rates and conjunction error rates was moderate at best (about .14). Moreover, when the faces were inverted to reduce configural-holistic processing, the conjunction effect (i.e., the conjunction-new difference in false-alarm rates) was not changed (it was about .20 regardless of orientation). The puzzle of the facial conjunction effect remains unsolved. 2 In this and all subsequent PCAs, we assigned signs to loadings such that false alarms in response to new faces load negatively on the factors. 3 Factor scores reflect the extent to which a participant’s data reflect the pattern of the loadings on a factor. Hence, participants with higher Factor 1 scores show lower false-alarm rates, while those with higher Factor 2 scores show higher hit rates. 4 Presentation frequency had these same effects in Experiment 1 as well, although the pattern there was slightly different (repetition increased hits and reduced new-face false alarms, without affecting conjunction false alarms, see Figure 5). We suspect that participants in the three-presentations condition of Experiment 1 set higher recognition criteria than those in the one-presentation condition, masking an effect of repetition on conjunction false alarms. Component Structure 53 5 It is now well established that recollection can be used to reject lures as well as accept targets (Lampinen, Odegard & Neuschatz, 2004). Given the near-0 loadings of hit rates on Factor 1, it would presumably reflect a “recollection rejection” component rather than a general recollection component. 6 The study and test lists were lengthened because we wished this study to be procedurally identical to a parallel event-related potential study in which many observations per participant and condition were required. 7 The correlation of Factor 2 scores with d's for old/FL discrimination was +.87. 8 The corresponding correlations of Factor 2 scores with d's for old/FL discrimination were .92 and .87, respectively. 9 The hit rates in the Stark and Squire (2003) experiments were estimated from false-alarm rates and d' values reported in the article. Component Structure 54 ReferencesAnderson, N. D., & Craik, F. I. M. (2000). Memory in the aging brain. In E. Tulving & F. I. M. Craik (Eds.), The Oxford Handbook of Memory (pp. 411-422). New York:Oxford University Press.Bartlett, J. C. (1993). Limits on losses in face recognition. In J. Cerella, J. Rybash, W.Hoyer, & M. Commons (Eds.), Adult Information Processing: Limits on loss (pp.351-379). New York: Academic Press.Bartlett, J. C., & Memon, A. (2007). Eyewitness memory in young and older adults. InR. C. L. Lindsay, D. F. Ross, J. D. Read, & M. P. Toglia (Eds.), Handbook ofEyewitness Psychology, Volume II: Memory for People (pp. 309-338). New York:Psychology Press.Bartlett, J. C., & Searcy, J. H. (1993). Inversion and configuration of faces. CognitivePsychology, 25, 281-316.Bartlett, J. C., Searcy, J. H., & Abdi, H. (2003). What are the routes to face recognition?In M. A. Peterson & G. Rhodes (Eds.), Perception of Faces, Objects, and Scenes:Analytic and Holistic Processes (pp. 21-52), London: Oxford University Press.Bors, D. A., & MacLeod, C. M. (1996). Individual differences in memory. In E. L. Bjork& R. A. Bjork (Eds.), Handbook of Perception and Cognition, Volume 10:Memory (pp. 411-441). San Diego, CA: Academic Press.Brainerd, C. J., & Reyna, V. F. (2005). The Science of False Memory. New York: OxfordUniversity Press.Cabeza, R., & Kato, T. (2000). Features are also important: Contributions of featural andconfigural processing to face recognition. Psychological Science, 11, 429-433. Component Structure 55 Charman, S. D., & Wells, G. L. (2007). Applied lineup theory. In R. C. L. Lindsay, D. F.Ross, J. D. Read, & M. P. Toglia (Eds.), Handbook of Eyewitness Psychology, Volume II: Memory for people (pp. 219-254). New York: Psychology Press.Damasio, A. R., Tranel, D., & Damasio, H. (1990). Face agnosia and the neuralsubstrates of memory. Annual Review of Neuroscience, 13, 89-109.De Renzi, E., Perani, D., Carlesimo, G. A., Silveri, M. C., & Fazio, F. (1994).Prosopagnosia can be associated with damage confined to the right hemisphere:An MRI and PET study and a review of the literature. Neuropsychologia, 32,893-902.Diamond, R., & Carey, S. (1986). Why faces are and are not special: An effect ofexpertise. Journal of Experimental Psychology: General, 115, 107-117.Diana, R. A., Reder, L. M., Arndt, J., & Park, H. (2006). Models of recognition: A reviewof arguments in favor of a dual-process account. Psychonomic Bulletin & Review,13, 1-21.Dodson, C. S., & Schacter, D. L. (2002). When false recognition meets metacognition:The distinctiveness heuristic. Journal of Memory and Language, 46, 782-803.Eichenbaum, H., Yonelinas, A. P., & Ranganath, C. (2007). The medial temporal lobeand recognition memory. Annual Review of Neuroscience, 30, 123-152.Farah, M. J. (1991). Patterns of co-occurrence among the associative agnosias:Implications for visual object recognition. Cognitive Neuropsychology, 8, 1-19.Fodor, J. (1983). The Modularity of Mind: An Essay on Faculty Psychology. Cambridge,MA: MIT Press. Component Structure 56 Freire, A., Lee, K., & Symons, L. A. (2000). The face-inversion effect as a deficit in theencoding of configural information: Direct evidence. Perception, 29, 159-170. Friedman, N. P., Myake, A., Young, S. E., DeFries, J. C., Corley, R. P., & Hewitt, J. K.(in press). Individual differences in executive functions are almost entirelygenetic in origin. Journal of Experimental Psychology: General.Hannigan, S. L., & Reinitz, M. T. (2000). Influences of temporal factors on memoryconjunction errors. Applied Cognitive Psychology, 14, 309–321.Hasher, L., & Zacks, R. T. (1988). Working memory, comprehension and aging: Areview and a new view. In G. H. Bower (Ed.), The Psychology of Learning andMotivation: Advances in research and theory, Volume 22 (pp. 193-225). NewYork: Academic Press.Jacoby, L. L. (1991). A process dissociation framework: Separating automatic fromintentional uses of memory. Journal of Memory and Language, 30, 513–541.Jacoby, L. L. (1999). Ironic effects of repetition: Measuring age-related differences inmemory. Journal of Experimental Psychology: Learning, Memory, andCognition, 25, 3-22.Jennings, J. M., & Jacoby, L. L. (1997). An opposition procedure for detecting age-related deficits in recollection: Telling effects of repetition. Psychology andAging, 12, 352–361.Jolliffe, I. T. (2002). Principal Component Analysis. New York: Springer-Verlag.Jones, T. C. & Bartlett, J. C. (2009). When false recognition is out of control: The caseof facial conjunctions. Memory & Cognition, 37, 143-157.Jones, T. C., Bartlett, J. C., & Wade, K. A. (2006). Nonverbal conjunction errors in Component Structure 57 recognition memory: Support for familiarity but not for feature bundling. Journalof Memory and Language, 55, 138-155. Jones, T. C., & Jacoby, L. L. (2001). Feature and conjunction errors in recognitionmemory: Evidence for dual-process theory. Journal of Memory and Language,45, 82–102.Jones, T. C., Jacoby, L. L., & Gellis, L. A. (2001). Cross-modal feature and conjunctionerrors in recognition memory. Journal of Memory and Language, 44, 131–152.Kassin, S.M., Tubb, V. A., Hosch, H. M., & Memon, A. (2001). On the “generalacceptance” of eyewitness testimony research: A new survey of experts.American Psychologist, 56, 405-416.Kelley, C. M., & Jacoby, L. L. (2000). Recollection and familiarity: Process-dissociation.In E. Tulving & F. I. M. Craik (Eds.), The Oxford Handbook of Memory (pp. 215-228). Oxford: Oxford University Press.Kroll, N. E. A., Knight, R. T., Metcalfe, J., Wolf, E. S., & Tulving, E. (1996). Cohesionfailure as a source of memory illusions. Journal of Memory and Language, 35,176-196.Lampinen, J. M., Odegard, T. N., & Neuschatz, J. S. (2004). Robust recollection rejectionin the memory conjunction paradigm. Journal of Experimental Psychology:Learning, Memory, and Cognition, 30, 332-342.Lindsay, R. C. L., & Wells, G. L. (1985). Improving eyewitness identifications fromlineups: Simultaneous versus sequential lineup presentation. Journal of AppliedPsychology, 70, 556-564.Lustig, C., Hasher, L., & Zacks, R. T. (2007). Inhibitory deficit theory: Recent Component Structure 58 developments in a “new view.” In D. S. Gorfein & C. M. MacLeod (Eds.), ThePlace of Inhibition in Cognition (pp. 145-162). Washington, DC: American Psychological Association.Macmillan, N. A., & Creelman, C. D. (2005). Detection theory: A user’s guide (2edition). Mahwah, NJ: Erlbaum.Mandler, G. (1980). Recognizing: The judgment of previous occurrence. PsychologicalReview, 87, 252–271.Mayes, A. R., Montaldi, D., & Migo, E. (2007). Associative memory and the medialtemporal lobes. Trends in Cognitive Sciences, 11, 126-135.McKone, E., Kanwisher, N., & Duchaine, B. C. (2007). Can generic expertise explainspecial processing for faces? Trends in Cognitive Sciences, 11, 8-15.McKone, E. Martini, P. & Nakayama, K. (2003). Isolating holistic processing in faces(and perhaps other objects). In M. A. Peterson & G. Rhodes (Eds.), Perception ofFaces, Objects, and Scenes: Analytic and Holistic Processes (pp. 92-119).London: Oxford University Press.McKone, E., & Peh, Y. X. (2006). Memory conjunction errors for faces are consistentwith configural processing. Psychonomic Bulletin & Review, 13, 106-111.Melton, A. W. (1967). Individual differences and theoretical process variables: Generalcomments on the conference. In R. N. Gagné (Ed.), Learning and IndividualDifferences (pp. 238-252). Columbus, OH: Merrill.Memon, A., & Bartlett, J. C. (2002). The effects of verbalisation on face recognition. Applied Cognitive Psychology, 16, 635-650. Component Structure 59 Milner, B. (1968). Visual recognition and recall after right temporal-lobe excision inman. Neuropsychologia, 6, 191-209. Moscovitch, M. (1994). Memory and working with memory: Evaluation of a componentprocess model and comparisons with other models. In D. L. Schacter & E.Tulving (Eds.), Memory Systems, 1994 (pp. 269-310). Cambridge, MA: MITPress.Moscovitch, M., Winocur, G., & Behrmann, M. (1997). What is special about facerecognition? Nineteen experiments on a person with visual object agnosia anddyslexia but normal face recognition. Journal of Cognitive Neuroscience, 9, 555-604.Odegard, T. N., & Lampinen, J. M. (2006). Memory editing: Knowledge, criteria andalignment. Memory, 14, 777-787.Penrod, S. D., & Bornstein, B. H. (2007). Generalizing eyewitness reliability research. InR. C. L. Lindsay, D. F. Ross, J. D. Read, & M. P. Toglia (Eds.), Handbook ofEyewitness Psychology, Volume II: Memory for People (pp. 529-556). New York:Psychology Press.Rakover, S. S. (2002). Featural vs. configurational information in faces: A conceptualand empirical analysis. British Journal of Psychology, 93, 1-30.Rapcsak, S. Z. (2003). Face memory and its disorders. Current Neurology andNeuroscience reports, 3, 394-501.Rapcsak, S. Z., Nielsen, L., Littrell, L. D., Glisky, E. L., Kazniak, A. W., & Laguna, J. F.(2001). Face memory impairments in patients with frontal lobe damage.Neurology, 57, 1168-1175. Component Structure 60 Rapcsak, S. Z., Reminger, S. L., Glisky, E. L., Kazniak, A. W., & Comer, J. F. (1999).Neuropsychological mechanisms of false facial recognition following frontal lobe damage. Cognitive Neuropsychology, 16, 267-292.Reder, L. M., Wible, C., & Martin, J. (1986). Differential memory changes with age:Exact retrieval versus plausible inference. Journal of Experimental Psychology:Learning, Memory, and Cognition, 12, 72-81.Reinitz, M. T., Lammers, W. J., & Cochran, B. P. (1992). Memory conjunction errors:Miscombination of stored stimulus features can produce illusions of memory.Memory & Cognition, 20, 1–11.Reinitz, M. T., Morrisey, J., & Demb, J. B. (1994). Role of attention in face encoding.Journal of Experimental Psychology: Learning, Memory, and Cognition, 20,161–168.Roediger, H. L., & McDermott, K. B. (2000). Distortions of memory. In E. Tulving & F.I. M. Craik (Eds.), The Oxford Handbook of Memory (pp. 149-162). New York:Oxford University Press.Rossion, B. Caldara, R., Seghier, M., Schuller, A-M., Lazeyras, F., & Mayer, E. (2003).A network of occipito-temporal face-sensitive areas besides the right middlefusiform gyrus is necessary for normal face processing. Brain, 126, 2381-2395.Rubin, S. R., Van Petten, C., Glisky, E. L., & Newberg, W. M. (1999). Memoryconjunction errors in younger and older adults: Event-related potential andneuropsychological evidence. Cognitive Neuropsychology, 16, 459–488.Schacter, D. L. (2001). The Seven Sins of Memory: How the mind forgets and remembers.New York: Houghton Mifflin. Component Structure 61 Schacter, D. L., Wagner, A. D., & Buckner, R. L. (2000). Memory systems of 1999. InE. Tulving & F. I. M. Craik (Eds.), The Oxford Handbook of Memory (pp. 627643). New York: Oxford University Press.Schiltz, C., & Rossion, B. (2006). Faces are represented holistically in the humanoccipito-temporal cortex. Neuroimage, 32, 1385-1394.Searcy, J. H., & Bartlett, J. C. (1996). Inversion and processing of component and spatial-relational information in faces. Journal of Experimental Psychology: HumanPerception and Performance, 22, 904-915.Searcy, J. H., Bartlett, J. C., & Memon, A. (1999). Age differences in accuracy andchoosing in eyewitness identification and face recognition. Memory andCognition, 27, 538-552.Sergent, J. (1984). An investigation into component and configural processes underlyingface perception. British Journal of Psychology, 75, 221-242.Sergent, J., & Signoret, J. L. (1992). Varieties of functional deficits in prosopagnosia.Cerebral Cortex, 2, 375-388.Squire, L. R., Stark, C. E. L., & Clark, R. E. (2004). The medial temporal lobe. AnnualReview of Neuroscience, 27, 279-306.Stark, C. E. L., & Squire, L. R. (2003). Hippocampal damage equally impairs memoryfor single items and memory for conjunctions. Hippocampus, 13, 281-292.Steblay, N. (1997). Social influence in eyewitness recall: A meta-analytic review oflineup instruction effects. Law and Human Behavior, 21, 283-298.Steblay, N., Dysart, J., Fulero, S., & Lindsay, R. C. L. (2001). Eyewitness accuracy ratesin sequential and simultaneous lineup presentations: A meta-analytic comparison. Component Structure 62 Law and Human Behavior, 25, 459-473.Tanaka, J. W., & Farah, M. J. (1993). Parts and wholes in face recognition. Quarterly Journal of Experimental Psychology, 46A, 225-245.Tanaka, J. W., & Farah, M. J. (2003). The holistic representation of faces. In M. A.Peterson & G. Rhodes (Eds.), Perception of Faces, Objects, and Scenes: Analyticand Holistic Processes (pp. 53-74). Oxford: Oxford University Press.Tovée, M. J., & Cohen-Tovée, E. M. (1993). The neural substrates of face processingmodels: A review. Cognitive Neuropsychology, 10, 505-528.Tulving, E. (1985). How many memory systems are there? American Psychologist, 40,385-398.Underwood, B. J. (1975). Individual differences as a crucible in theory construction.American Psychologist, 30, 128-134.Underwood, B. J., Kapelak, S. M., & Malmi, R. A. (1976). Integration of discrete verbalunits in recognition memory. Journal of Experimental Psychology: HumanLearning and Memory, 2, 293-300.Wherry, R. J. (1984). Contributions to correlational analysis. NY: Academic Press.Yonelinas, A. P. (1997). Recognition memory ROCs for item and associativeinformation: The contribution of recollection and familiarity. Memory &Cognition, 25, 747-763.Yonelinas, A. P. (1999). The contribution of recollection and familiarity to recognitionand source-memory judgments: A formal dual-process model and an analysis ofreceiver operating characteristics. Journal of Experimental Psychology:Learning, Memory, and Cognition, 25, 1415-1434. Component Structure 63 Yonelinas, A. P. (2002). The nature of recollection and familiarity: A review of 30 yearsof research. Journal of Memory and Language, 46, 441-517. Yonelinas, A. P. & Parks, C. M. (2006). Receiver operating characteristics (ROCs) inrecognition memory: A Review. Psychological Bulletin, 133, 800-832. Component Structure 64 Table 1Correlations among Proportions of “Old” Judgments for Old, Conjunction, and New Faces, and Loadings of Each Measure on Two (Unrotated) Factors Identified by thePrincipal Component Analysis of Experiment 1 Correlations (r’s) among Measures and with Factor Loadings Measure/PCA FactorOldConjunctionNew Conjunction+.21New-.22+.46 Factor 1+.02-.85-.86 Factor 2+.96+.35-.32 Component Structure 65 Table 2.Pearson Product-Moment Correlations between Factors 1 and 2 of the Principal Component Analysis and d' and C scores for Old/New Discrimination and forOld/Conjunction Discrimination across the 192 Participants in Experiment 1. Old/New DiscriminationOld/Conjunction Discrimination PCA Factord'Cd'C Factor 1+.61+.73+.65+.48Factor 2+.77-.42+.44-.86 Note. All correlations are reliable (p < .001) Component Structure 66 Table 3.Means (M) and Standard Deviations (SD) of Proportions of “Old” Judgments for New Faces, Conjunctions of Faces Studied Once (1X-Conj.) and Eight Times (8X-Conj.), andOld Faces Studied Once (1X-Old) and Eight Times (8X-Old) in the Upright and InvertedConditions of Experiment 2a (E2a) and the Non-proximal and Proximal Conditions ofExperiment 2b (E2b-NProx, E2b-Prox), along with Upright-Inverted and Proximal-Non-proximal Differences