Recollection and Familiarity Deficits in Amnesia ... and familiarity using the threshold model are presented in Table 1. According to those estimates, amnesics exhibited a deficit

Neuropsychology Copyright 1998 by the American Psychological Association, Inc. 1998, Vol. 12, No. 3, 323-339 0894-4105/98/$3.00

Recollection and Familiarity Deficits in Amnesia: Convergence of Remember-Know, Process Dissociation,

and Receiver Operating Characteristic Data

Andrew E Yonelinas, Neal E. A. Kroll, Ian Dobbins, Michele Lazzara, and Robert T. Knight

University of California, Davis

Previous studies using the process dissociation and the remember-know procedures led to conflicting conclusions regarding the effects of anterograde amnesia on recollection and familiarity. We argue that these apparent contradictions arose because different models were used to interpret the results and because differences in false-alarm rates between groups biased the estimates provided by those models. A reanalysis of those studies with a dual-process signal-detection model that incorporates response bias revealed that amnesia led to a pronounced reduction in recollection and smaller but consistent reduction in familiarity. To test the assumptions of the model and to further assess recognition deficits in amnesics, we examined receiver operating characteristics (ROCs) in amnesics and controls. The ROCs of the controls were curved and asymmetrical, whereas those of the amnesics were curved and symmetrical. The results supported the predictions of the model and indicated that amnesia was associated with deficits in both recollection and familiarity.

One of the most robust and theoretically important observations in cognitive neuropsychology is the dissociation between performance on explicit and implicit memory tasks in anterograde amnesic populations. Despite showing poor performance on explicit tests of memory such as free recall, many of these patients are unimpaired on implicit tests such as perceptual identification and category exemplar generation (for reviews, see Cohen & Squire, 1980; Richard- son-Klavehn & Bjork, 1988; Roediger, 1990; Schacter, Chiu, & Ochsner, 1993). These findings are generally interpreted as showing that amnesics exhibit a selective disruption of the' systems or processes that support conscious recollection but nonrecollective forms of memory are unaffected.

However, amnesics do not perform equally poorly on all explicit memory tests. Although earlier comparisons of recognition and recall in amnesics have led to conflicting results (e.g., Haist, Shimamura, & Squire, 1992; Hirst, Johnson, Phelps, & Volpe, 1988), Aggleton and Shaw (1996) reviewed 33 studies that examined amnesics' recognition memory and found that patients with selective damage to the hippocampus, or its diencephalic targets, had a relative

Andrew E Yonelinas, Neal E. A. Kroll, Ian Dobbins, and Michele Lazzara, Department of Psychology, University of Califor- nia, Davis; Robert T. Knight, Department of Neurology and Center for Neuroscience, University of California, Davis.

This research was supported by National Institutes of Neurologi- cal Disorders and Stroke Grants NS 17778, NS21135 and NS 17778. We sincerely thank all of the participants, and especially the patients, who participated in the study.

Correspondence concerning this article should be addressed to Andrew E Yonelinas, Department of Psychology, University of California, Davis, California 95616. Electronic mail may be sent to apyonelinas @ ucdavis.edu.

sparing of recognition performance compared with recall. Drawing on concepts developed within dual-process frame- works of recognition memory (e.g., Atldnson & Juola, 1974; Mandler, 1980; Jacoby & Dallas, 1981), Aggleton and Shaw suggested that damage to the hippocampal system disrupted the mechanisms that support conscious recollection of contextual information but spared familiarity-based pro: cesses. Thus, despite an inability to explicitly remember information specific to the context of item occurrence, these patients are able to discriminate between novel items and items that are perceived as more familiar because of recent experimental exposure.

Similarly, Huppert and Piercy (1978) found that amnesics had disproportionate difficulty discriminating between recently presented and frequently presented items relative to healthy controls. They argued that although healthy controls could base their discriminations on both recollection and familiarity, amnesics were restricted to making judgments based on familiarity and thus had difficulty discriminating between recently presented and frequently presented items. Based on these findings, anterograde amnesia might be characterized as a selective deficit in recollection sparing familiarity. Indeed, it has been speculated that the same processes that support normal performance on implicit tests in amnesics may be responsible for familiarity-based recognition judgments (e.g., Mandler, 1980).

However, several studies that used procedures designed to estimate recollection and familiarity have led to differing conclusions regarding the effects of anmesia on familiarity. One study using the process dissociation procedure (Verfael- lie & Treadwell, 1993) suggested that familiarity was preserved in amnesia. However, studies using the remember- know procedure lead to divergent conclusions. Whereas one remember-know study suggested that familiarity was dis-

323

324 YONELINAS, KROLL, DOBBINS, LAZZARA, AND KNIGHT

rupted in amnesia (Knowlton & Squire, 1995), another study suggested that familiarity was enhanced in amnesia (Schac- ter, Verfaetlie, & Pradere, 1996), and a third study suggested that familiarity was not consistently affected by amnesia (Schacter, Verfaellie, & Anes, 1997).

We review these studies and argue that the discrepancies between experiments arose because large differences in false-alarm rates between amnesics and controls biased the estimates of the memory processes and because different models were used to interpret the data in those experiments. Moreover, we describe a dual-process signal-detection model that explicitly incorporates response bias, and we show that when this model is used to reanalyze the data, the results of the previous studies are in good agreement. The model is then further tested by examining receiver operating characteristics (ROCs) for amnesics and controls in a recognition experiment, and the results of that experiment are used to further characterize the nature of the familiarity deficit seen in amnesia.

The Process Dissociation Procedure

Verfaellie and Treadwell (1993) examined recollection and familiarity in amnesics using the process dissociation procedure (Jacoby, 1991; Jacoby, Toth, & Yonelinas, 1993). 1 In Phase 1 of the study, participants read words aloud or solved anagrams (e.g., clsss = class). In Phase 2, they were presented with a different list of words and were asked to study these items in preparation for a subsequent memory test. For the recognition test, participants were presented with a mixture of previously encountered words and novel distractors and were given either inclusion or exclusion instructions. Under inclusion instructions, they were told to endorse (i.e., respond "yes") all items previously seen in the experiment and to reject the novel distracters. The assumption was that participants could correctly recognize an item from Phase 1 if it was recollected (R) or, in the absence of recollection (1 - R), if it was sufficiently familiar (F):

amnesics calculated using the process dissociation equations just presented. Examination of the estimates of the original process dissociation model shows that amnesics exhibited a sizable deficit in recollection for items solved as anagrams at study (recollection decreased by .32 or 97% 2) and only a minor reduction in recollection for items that were read at study (.01 or 8%). Moreover, amnesia was associated with a slight increase in familiarity for read items (.05 or 13%) and a slight decrease for the items solved as anagrams (.07 or 14%). Based on these results, Verfaellie and Treadwell (1993) concluded that amnesics were less likely than controls to base their recognition judgments on recollection, but they were equally likely to use familiarity for recognition judgments.

Roediger and McDermott (1994) observed that there were differences in the false-alarm rates between controls and amnesics and suggested that these differences may have contaminated the estimates of recollection and familiarity. They proposed that false-alarm rates might be incorporated into the process dissociation procedure by subtracting the false-alarm rates from the hit rates in the inclusion and exclusion conditions before using the process dissociation procedure to estimate recollection and familiarity. This method of incorporating false alarms is based on a threshold model that assumes that new items never exceed a memory threshold (i.e., they are never truly recognized), but new items may be accepted as old on the basis of a random guess. In this way, the effects of guessing can be removed by subtracting the false alarms from the hits. Estimates of recollection and familiarity using the threshold model are presented in Table 1. According to those estimates, amnesics exhibited a deficit in recollection for both read items (0.08 or 57%) and items solved as anagrams (0,39 or 100% 3 ) during the study phase and exhibited a smaller but consistent decrease in familiarity. Familiarity decreased by 0,04 (25%) for read items and 0.07 (3%) for anagram items. Roediger and McDermott contended that the large decrease in recollec-

P("yes" Iold)inc = R + (1 - R)E (~)

In contrast, participants in the exclusion condition were told to endorse only items from Phase 2 and to reject novel distracters and Phase 1 items. Thus, unlike the inclusion condition, participants were told to exclude Phase 1 items. In the exclusion condition, the assumption was that participants would incorrectly accept a Phase 1 item only if it was familiar (F) and not recollected (1 - R):

P("yes" [old)exc = (1 - R)E (2)

We will refer to Equations 1 and 2 as the original process dissociation model. Inclusion and exclusion scores were then used to estimate the contribution of recollection and familiarity to performance. That is, R = P("yes" Iold)inc - P("yes" Iold)exc, and F = P("yes" Iold)exc/(1 - R).

Table 1 presents the proportion of yes responses in the inclusion and exclusion conditions. Table 1 also presents estimates of recollection and familiarity for controls and

1 The process dissociation procedure has been used to examine the effects of amnesia in other memory tasks such as false fame and stem completion (e.g., Cermak, Verfaellie, Butler, & Jacoby, 1993; Cermak, Verfaellie, Sweeney, & Jacoby, 1992; Mayes, Van Eijk, & Isaac, 1995). However, these studies are not included in our review because the processes involved in these tasks may differ from those in recognition memory.

2 We report both the absolute differences in parameter estimates between the amnesics and controls as well as the percent change relative to the controls. Although the two indexes can differ considerably, the percent change is important in assessing the relative effects of amnesia on these two processes. Moreover, in the case of the dual-process model, it is not useful to directly compare the absolute parameter estimates of recollection and familiarity because the two parameters are measured on different scales (i.e., probabilities vs. d' values).

3 The threshold model led to a negative estimate of recollection for the anagram items for the amnesics, which represents a 111.4% decrease in recollection relative to controls. However, we treated this as a 100% decrease on the assumption that the actual probability of recollection was 0 and the negative value was spurious.

RECOGNITION IN AMNESIA

Table 1 inclusion and Exclusion Data for Controls (Con)and Amnesics (Amn) From VerfaeUie and Treadwell (1993), Along With Estimates of Recollection and Familiarity

Read i t e m s Anagram items New items

Condition or parameter Con Amn Con Anm Con Amn

Proportion of yes responses Inclusion .45 .49 .67 .45 Exclusion .33 .38 .34 ~44

Estimates from the original process dissociation model

Recollection .12 .11 .33 .01 Familiarity .38 .43 .51 .44

Estimates from the threshold model Recollection .14 .06 .35 -.04 Familiarity .16 .12 .23 .16

Estimates from the dual-process signal- detection model

Recollection .14 .06 .35 .00 Familiarity (d') 0.59 0.37 0.94 0.40

.17 .32

.19 .27

Note. The original process dissociation model indicated that amnesia led to a deficit in recollection and no consistent change in familiarity, the threshold model indicated a deficit in recollection and a slight reduction in familiarity, and the dual-process signal-detection model indicated a deficit in both recollection and familiarity.

325

tion was more in keeping with known explicit memory deficits in amnesics and that familiarity was left relatively intact.

In response, Verfaellie (1994) argued that a more appropriate method of incorporating response bias into the estimates was to use a model in which familiarity was assumed to reflect a signal-detection process of the type that underlies standard d ' tables (see Jacoby et al., 1993; Yonelinas, 1994). In such a model, false alarms arise because the familiarity of some of the new items exceeds the participants' response criterion. In this way, familiarity is measured as the differ- ence between the average familiarity of the old and new items (i.e., d'). The model is based on the original process dissociation equations except that familiarity is measured in terms of d ' and the model incorporates performance on old as well as new items (see Appendix A for a more detailed description of the model). With healthy participants, the model has been shown to provide process dissociation procedure estimates that are not contaminated by response bias (Yonelinas & Jacoby, 1996; Yonelinas, Regehr, & Jacoby, 1995).

Table 1 shows the parameter estimates based on the inclusion and exclusion scores using the dual-process signal- detection model. According to these estimates, amnesics suffered a deficit in recollection for read (0.08 or 57%) and anagram items (0.35 or 100%) as well as a deficit in familiarity for read (0.22 or 37%) and anagram items (0.54 or 57%). Note that because familiarity is assumed to reflect a signal-detection process it is measured in terms of d ' rather than as a probability. Verfaellie (1994) conducted a similar analysis on individual participants using a simplified version of the model and found a similar pattern.

In summary, for the process dissociation procedure, all three of the models indicated that amnesics exhibited a deficit in recollection. However, the effect on familiarity was less clear. The original process dissociation model showed

that amnesia did not consistently effect familiarity, the threshold method indicated a slight decrease in familiarity, and the dual-process signal-detection model showed a more noticeable decrease in familiarity. We will return to these results after discussing the results from three remember- know experiments.

The Remember -Know Procedure

The process dissociation procedure measures recollection as a basis of control. If the participants can recollect an item, they can either include or exclude it when instructed to do so. However, recollection should also serve as a basis for subjective reports. For example, in the remember-know procedure developed by Tulving (1985), participants are asked to respond that they "remember" when they can consciously recollect having encountered an item before and to respond that they "know" the item was presented before if it is familiar in the absence of recollection. Thus, the remember-know procedure also provides a way of examining the contribution of recollection and familiarity to memory performance. Three studies have examined recognition memory in anmesics using the remember-know procedure. In one study, Knowlton and Squire (1995) presented a list of words to study to a group of amnesics and a group of controls and tested recognition memory after a 10-min delay. Participants were instructed to respond "remember" (R) if they could recollect specific details regarding the item's occurrence during study and to respond "know" (K) if they recognized the item was in the study list on the basis that it was familiar but it did not evoke specific recollections. (For a review of the remember-know procedure, see Rajaram & Roediger, 1997.) Table 2 shows the probability of remember and know responses for amnesics and controls. With these response proportions as indices of recollection and familiarity, it would appear that amnesics displayed a deficit in


Table 2 The Proportion of Remember and Know Responses for Controls (Con) and Amnesics (Amn ) Reported by Knowlton and Squire (1995), Along With Estimates of Recollection and Familiarity

Studied items New items

Condition or parameter Con Amn Con Amn

Proportion of remember and know responses a Remember (recollection) .46 .17 Know (familiarity) .21 .16

Estimates from the threshold remember-know model Recollection .44 .08 Familiarity .17 .06

Estimates from the dual-process signal-detection model

Recollection .45 .09 Familiarity (d') 1.47 0.35

.02 .09

.04 .10

Note. All models showed that recollection decreased with amnesia. The original remember-know model indicated that amnesia led to a slight deficit in familiarity (i.e., know responses), the threshold model indicated that amnesia led to a slightly larger decrease in familiarity, and the dual-process signal-detection model indicated a more noticeable decrease in familiarity. aEstimates from the original remember-know model.

recollection (i.e., the probability of an R response decreased by 0.29 or 63%) and a smaller impairment in familiarity (the probability of a K response decreased by 0.05 or 23%).

However, as with the Verfaellie and Treadwell (1993) study, there were sizable differences in the false-alarm rates between the amnesics and controls that complicated the interpretation of the results. An examination of Table 2 shows that the false-alarm rates for amnesics were more than twice those of the controls. One method of incorporating response bias that is commonly used in remember-know studies is to subtract the proportion of incorrect remember and know responses from the respective proportions of correct responses (e.g., Gardiner & Java, 1991). As with the hit-minus-false-alarm-rate method used with the process dissociation procedure, this method relies on a threshold model. The estimates based on the threshold model (see Table 2) indicate that there were decrements in recollection and familiarity for amnesics relative to controls. Amnesia was associated with a deficit in recollection of 0.36 (82%) and decrease in familiarity of 0.11 (65%). Note that Knowl- ton and Squire (1995) used the hit rates and the false-alarm rates to calculate d ' values for both recollection and familiarity and reached similar conclusions.

The dual-process signal-detection model that was used with the process dissociation procedure can also be used to incorporate response bias in the remember-know procedure (e.g., Jacoby, Yonelinas, & Jennings, 1997; Yonelinas & Jacoby, 1995; Yonelinas, Dobbins, Szymanski, Dhaliwal, & King, 1997). Recollection is measured as a simple probability. However, because familiarity is assumed to reflect a signal-detection process, it must be measured in terms of d ' . The method of estimating recollection and familiarity using the model is described in Appendix A.

The estimates of recollection and familiarity using the dual-process signal-detection model are given in Table 2. According to this model, amnesics exhibited deficits in both recollection and familiarity. Estimates of recollection were

0.36 (80%) lower for amnesics than for controls, and estimates of familiarity, measured as d ' values, were 1.12 (76%) lower.

In summary, an examination of the Knowlton and Squire (1995) results showed that in the original remember-know model amnesia was associated with a large deficit in recollection and a modest deficit in familiarity. Similarly, according to both the threshold model and the dual-process signal-detection model, recollection and familiarity were reduced in amnesics relative to controls. Although all three models are in general agreement in terms of the current data set, conflicting results have been reported in other remember- know studies.

Schacter et al. (1996) also examined recognition memory in amnesics and controls using the remember-know procedure. They used a procedure introduced by Deese (1959) and later developed by Roediger and McDermott (1994) in which the test list contained studied items, new items, and new items that were associatively related to studied items. They found that participants often falsely recognized the associated items. However, recognition performance for the studied and new items is of most interest in the current context. In Roediger and McDermott's experiment, participants studied a list of words then either tried to recall the words or performed an arithmetic distracter task. After several such study lists, participants were given a recognition test and were asked to make remember-know judgments.

The proportion of remember and know responses to studied and new items for amnesics and controls is presented in Table 3. Based directly on the proportion of remember and know responses, the original remember-know model indicates that amnesia led to a decrease in recollection (the proportion of remember responses decreased by 0.41 or 58% and 0.46 or 65% for the recall and the arithmetic conditions, respectively) but led to an increase in familiarity (the proportion of know responses increased by 0.10 or 71% and 0.09 or 75% for recall and the arithmetic conditions,

RECOGNITION IN AMNESIA 327

Table 3 The Proportion of Remember and Know Responses for Controls (Con) and Amnesics (Amn) From Schacter, Verfaellie, and Pradere (1996), Along With Estimates of Recollection and Familiarity

Studied items

Free reca l l Arithmetic task New items

Condition or parameter Con Amn Con Amn Con Amn

Proportion of remember and know responses a

Remember (recollection) .71 .30 .71 .25 Know (familiarity) .14 .24 .12 .21

Estimates from the threshold remember-know model

Recollection .65 .16 .65 .11 Familiarity .02 .04 .00 .01

Estimates from the dual-process signal- detection model


.06 .14

.12 .20

Note. All models showed that recollection decreased for the amnesics relative to the controls. The original remember-know model indicated that familiarity increased, the threshold model indicated that familiarity increased slightly, and the dual-process signal-detection model indicated that familiarity decreased for amnesics relative to controls. aEstimates from the original remember-know model.

respectively). The paradoxical finding that amnesia led to an increase in familiarity is discussed later. However, as in the two previous studies, the amnesics produced approximately twice as many false alarms as the controls, indicating that it was necessary to examine the data using a model that incorporated response bias.

When the threshold model was applied to remember and know responses (see Table 3), it indicated that there was a large decrease in recollection (recollection decreased by 0.49 or 75% and 0.54 or 83% for the recall and the arithmetic conditions, respectively) and a slight increase in familiarity (familiarity increased by 0.02 or 100% and 0.014 for the recall and arithmetic conditions, respectively).

One of the difficulties that both of the aforementioned models have in interpreting the results of this last study is exacerbated by the high recollection (R) hit rates. Because participants were instructed to make a K response only when an item was familiar and not recollected, recollection mathematically constrains the proportion of K responses. Thus, as recollection hit rates become high, as they did for controls in this experiment, familiarity scores become more constrained and, as a result, the familiarity scores of the amnesics will look high in comparison. Thus, the proportion of K responses was greater for the amnesics than the controls because the amnesics' K responses were much less constrained by high levels of recollection.

The fact that K responses are mathematically constrained by recollection leads to another seemingly paradoxical result in the experiment. That is, as Schacter et al. (1996) pointed out, participants were almost as likely to make a K response to new items as to old items, and this conflicts with results from similar remember-know studies. This finding might be interpreted as showing that there were no familiarity effects on recognition performance. However, it seems more likely that these results were due to the mathematical constraints of

the remember-know procedure. That is, a high proportion of R responses to old items would be expected to lead to an artifactuaUy low proportion of K responses to those items. In contrast, because participants made very few R responses to new items, the proportion of K responses to these items were much less constrained. Thus, a similar number of K responses associated with old and new items does not mean that those items are equally familiar. Because the remember and know responses are mutually exclusive (i.e., participants must make one response or the other), whenever the proportion of remember responses is greater than 0, the K responses will be mathematically constrained by recollection and thus cannot be used as an index of familiarity.

In contrast, the dual-process signal-detection model al- lows for the independent contribution of recollection and familiarity, and, as a result, the estimates of familiarity it provides are not limited by the proportion of remember responses. Estimates using the dual-process model (see Item C in Table 3) show that anmesia was associated with a large deficit in recollection (0.50 or 73% and 0.56 or 81% for the recall and arithmetic conditions, respectively) and also a large deficit in familiarity (0.75 or 69% and 0.74 or 82% for the recall and arithmetic conditions, respectively). The estimates of familiarity were well above 0, suggesting that familiarity did contribute to performance.

In summary, the analysis of the data reported by Schacter et al. (1996) showed that amnesia was associated with a decrease in recollection, regardless of the method used to incorporate response bias. However, the effect on familiarity depended on the model used to interpret the data. According to the estimates from the original remember-know model,

4 Because the estimate of familiarity was 0 for controls, the change measured as a percentage is undefined.


familiarity increases with amnesia, which seems unlikely. The threshold model also suggested that amnesia was associated with an increase in familiarity, albeit a much smaller increase. Finally, the dual-process signal-detection model showed that amnesics exhibited a sizable deficit in familiarity.

In a more recent study, Schacter et al. (1997) examined the occurrence of false recognition for semantically and perceptually related items. As in their previous study, they collected remember and know responses from amnesics and controls, thus their data can be used to examine the effects of amnesia on recollection and familiarity. In two experiments, participants studied lists that consisted of words that were either highly conceptually similar (i.e., high associates) or highly perceptually similar (e.g., fate, fade, fame, face). After a brief distracter task, participants were given a recognition test and were asked to make remember-know judgments. Schacter et al. examined performance on items that were in the study lists (old items), unrelated words that were not in the study lists (new items), and items that were either conceptually or perceptually related to studied items (false recognition items). As with the previous study, we will only examine performance on the old and new items. Note that Schacter et al. examined performance in two separate experiments; because the two experiments led to similar results and the second experiment was designed to remove several potential confounds that were present in the first experiment, we will focus only on Experiment 2.

The proportion of remember and know responses studied and new items for amnesics and controls (Experiment 2; Schacter et al., 1997) is presented in Table 4. Based directly on the proportion of remember and know responses, the original remember-know model indicated that amnesia led to a decrease in recollection (the proportion of remember responses decreased by 0.29 or 58% and 0.29 or 60% for the semantic and perceptual lists, respectively) but had no

consistent effect on familiarity (the proportion of know responses increased by 0.03 or 12% and decreased by 0.03 or 9% for the conceptual and :perceptual lists, respectively). However, as in the two previous studies, the amnesics produced approximately twice as many false alarms as the controls indicating that it was necessary to examine the data using a model that incorporated response bias. Note that unlike the previous study where the high level of remember responses pushed the proportion of K responses to baseline levels, the K responses in the current study were above baseline.

When the threshold model was applied to remember and know responses (see Table 4), it indicated that there was a large decrease in recollection (recollection decreased by 0.40 or 80% and 0.31 or 64% for the conceptual and perceptual lists, respectivdy), a slight increase in familiarity for the conceptual lists (0.03 or 21%), and a slight decrease in familiarity for the perceptual lists (0.07 or 26%).

Estimates using the dual-process model (see Item C in Table 4) show that amnesia was associated with a large deficit in recollection (0.39 or 78% and 0.31 or 64% for the conceptual and perceptual lists, respectively) and a large deficit in familiarity (0.44 or 36% and 0.92 or 50% for the conceptual and perceptual lists, respectively).

In sunurmry, the analysis of the data reported by Schacter et al. (1997) showed that amnesia was associated with a decrease in recollection, regardless of the method used to incorporate response bias. However, as with earlier studies, the effect on familiarity depended on the model used to interpret the data. According to the estimates from the original remember-know model and the threshold model, amnesia led to a slight increase in familiarity in one condition and a slight decrease in another. In contrast, the dual-process signal-detection model suggested that amnesics exhibited a sizable deficit in familiarity in both conditions.

Table 4 The Proportion of Remember and Know Responses for Controls (Con) and Amnesics (Amn) From Schacter, Verfaellie, and Anes (1997), Along With Estimates of Recollection and Familiarity

Conceptual Perceptual

Old items New items Old items New items

Condition or parameter Con Amn Con Amn Con Amn Con Amn

Proportion of remember and know responsesa Remember (recollection) .50 .21 Know (familiarity) .25 .28

Estimates from the threshold remember-know model Recollection .50 .10 Familiarity .14 .17

Estimates from the dual-process signal-detection model

Recollection .50 .11 Familiarity (d') 1.23 0.79

.00 .11 .48 .19 .00 .02

.11 .11 .35 .32 .08 .12

.48 .17

.27 .20

.48 .17 1.84 0.92

Note. All models showed that recollection decreased for the amnesics relative to the controls, The original remember-know model and the threshold model indicated that amnesia was associated with an increased in familiarity in one condition and decreased in familiarity in another. The dual-process signal-detection model indicated that familiarity decreased for amnesics relative to controls in both conditions. aEstimates from the original remember-know model.


Interpreting the Results o f ~ e Previous Experiments

The process dissociation and remember-know studies showed that recollection was impaired in amnesics, regardless of the model used to incorporate response bias. How- ever, the effects of amnesia on familiarity were found to depend critically on the response bias model that was used. The original process dissociation model suggested that familiarity was not consistently disrupted in amnesics (Ver- faellie & Treadwell, 1993). The original remember-know procedure showed that amnesia led to a decrease in familiarity (i.e., knowing) in one study (Knowlton & Squire, 1995), an increase in familiarity in another (Schacter et al., 1996), and no consistent effect in another (Schacter et al., 1997).

In ~ four of these studies, there were large differences in false alarms between the participant groups, with the amnesics often producing false-alarm rates that were more than twice those of the controls. Such differences in false-alarm rates necessitate incorporating response bias into any interpretation of the results. The original process dissociation and remember-know models do not incorporate response bias and, therefore, cannot be expected to provide reliable estimates of recollection and familiarity when differences in false-alarm rates arise between groups. The threshold model does provide a convenient way of incorporating response bias (i.e., simply subtracting false alarms from hits) but fails to resolve the inconsistencies observed between studies. When the threshold model was applied to the process dissociation data, it indicated that amnesia was associated with a slight decrease in familiarity. However, when ~ threshold model was applied to the remember- know data, it showed that amnesia was associated with a large decrease in familiarity in one study, a slight increase in familiarity in the other study, and inconsistent effects in another study.

In contrast, the dual-process signal-detection model revealed that the apparent inconsistencies observed in previous studies of amnesia were due in part to differences in response bias between amnesics and controls and that amnesia had consistent effects on recollection and familiarity across all conditions in all experiments. Regardless of the data set, the model indicated that anmesia was associated with a large deficit in both recollection and familiarity, with the deficit in familiarity tending to be less than that for recollection. On average, amnesia was associated with a decrease in recollection estimates of approximately 77% and a decrease in familiarity estimates of 58%.

The ability of the dual-process signal-detection model to reveal a relatively consistent pattern of results despite sizable differences in response bias provides support for the model. However, such consistency in itself does not prove the model is correct nor does it prove that familiarity is disrupted in amnesia. To further test the validity of the model and to assess the previous conclusions, we explicitly measured the model's account of memory and response bias for amnesics and controls in an experiment that generated ROC curves for both groups.

An ROC Analysis of Recognit ion in Amnesia

To further test the dual-process signal-detection model and to examine the recognition deficits in amnesia in more detail, we conducted a recognition memory experiment in which amnesics and healthy controls rated the confidence of their recognition judgments. Confidence-based ROCs were plotted to determine if the dual-process model could accurately account for memory and response bias. Moreover, the recognition memory experiment provided a method, different from the process dissociation and remember-know procedures, to assess the degree to which familiarity was disrupted in amnesics. We begin by describing the experiment and how the ROC analysis was conducted. We then describe the qualitative predictions of the model and how the model was assessed. Finally, we describe how the confidence responses were used to derive estimates of recollection and familiarity.

ROCs

ROC analysis has been used for over 20 years to test whether memory models provide accurate accounts of memory and response bias (e.g., Murdock, 1974; Ratcliff, Sheu, & Gronlund, 1992). ROC analysis involves plotting the relationship between hits and false alarms as the response criterion is varied. If a model can accurately account for this function, then it should provide estimates of memory that are not confounded by differences in response bias.

In the current experiment, participants were given a recognition memory test for previously studied words. This required them to rate the confidence of their recognition responses from 6 (certain it was studied) to 1 (certain it was not studied). ROCs were generated by plotting performance (hits vs. false alarms) as a function of response confidence. The first point of the function was determined by adopting a strict scoring criterion that only accepted the most confi- dently recognized items (i.e., responses of 6) as hits and false alarms. The next point reflected a slightly less strict criterion, and included both 5s and 6s as hits and false alarms. In this way, separate points were plotted for each level of response confidence.

Assessing the Dual-Process Signal-Detection Model

If the model is correct, then the probability of a hit is equal to the probability that an old item is recollected (R) plus the probability that it is not recollected (1 - R) but its familiarity exceeds the response criterion (Fold):

P("old"lold) = R + (1 -R)(F~d). (3)

Note that this is identical to the inclusion equation presented earlier. The probability of a false alarm (i.e., incorrectly accepting a new item as old) is equal to the probability that its familiarity exceeds the response criterion (Fnew):

P("old" I new) = (Fne w). (4)


If familiarity is well described as a signal-detection process, then the probability that old and new items will be accepted on the basis of familiarity (Fold and Fnew) will be a function of how much more familiar studied items are relative to new items (i.e., d') and response criterion (i.e., c; see Appendix A).

Given a set of values for recollection and familiarity, the model predicts an ROC. Figure 1 presents two ROCs generated by the model. The upper curve represents performance when the probability of recollection is .40 and familiarity, measured in terms ofd ' , is 0.60. The lower curve represents predicted performance when recollection is equal to 0 and familiarity is 0.60.

Examination of Figure 1 shows that when recollection does not contribute to performance (the lower curve), the predicted ROC is symmetrical along the diagonal. However, when recollection does contribute to performance (i.e., the upper curve), the ROC will be asymmetrical because recollection increases the hit rate and effectively pushes the ROC up. Under most experimental conditions, the ROCs of healthy participants are asymmetrical like the upper function in Figure 1 and the dual-process model fits their ROCs quite well, presumably because healthy participants can make their recognition judgments based on both recollection and familiarity. It is important to note, however, that we do not know if the model accounts for the ROCs of amnesics. To our knowledge, recognition memory ROCs of anmesics have not previously been published.

Examining the ROCs of amnesics is interesting because the dual-process model makes some unique predictions

1.0

0.8

"o 0.6 0

> , 0.4 a

0.2

0

. 4 0 , d ' = 0 . 6 0 _

= .00, d' = 0.60 " I ..... t i I

0 0.2 0.4 0.6 0.8 1.0

P( 'yes ' lnew)

Figure 1. Two receiver operating characteristics (ROCs) generated by the dual-process signal-detection model. The lower curve shows the symmetrical ROC produced by the familiarity process. The upper curve shows the asymmetrical ROC produced when both recollection and familiarity contribute to performance.

about the nature of the ROCs that should be observed with those patients. If the model is correct and the amnesics' recognition judgments are based primarily on familiarity, as authors of numerous previous studies have suggested (e.g., Aggleton & Shaw, 1996; Huppert & Piercy, 1978; Verfaellie & Treadwell, 1993), then their ROCs should be unlike those of the controls in that they should be symmetrical along the diagonal (e.g., like the lower curve in Figure 1).

However, there is one important caveat to this prediction. Although the previous analysis showed that amnesics did exhibit a profound deficit in recollection, it was not always completely eliminated (see Tables 1--4). If amnesics could recollect some small proportion of test items, then their ROCs should be slightly asymmetrical. To ensure that we would be able to observe familiarity-based ROCs, we contrasted performance for items encoded under shallow- and deep-encoding conditions. Shallow processing (e.g., count the number of syllables in each word) compared to deep processing (e.g., judge whether the word is abstract or concrete) should be less likely to support later recollection. Thus, we expected amnesics' ROCs to be more symmetrical than controls' ROCs and amnesics' shallow encoding ROC to be the most symmetrical.

In addition to testing the general predictions of the model, we also fit the model equations to the observed ROCs for both control and amnesic participants to determine the accuracy of the model's account of recognition performance. If the model does fit the data, then it should provide an accurate way of incorporating response bias into the estimation procedure. This would justify the use of the model in the previous experiments and lend support to the conclusions drawn there.

Examining the Effects of Amnesia on Recollection and Familiarity

The second goal of the experiment was to use the confidence judgments along with the dual-process model to derive new estimates of recollection and familiarity to determine if amnesics do exhibit deficits in familiarity and recollection. This was done by fitting the dual-process equations to the observed ROC to determine the values of R and d' that provided the best account of the data. This, of course, relies on the success of the model in fitting the observed ROC. If the model does not fit the data well, then the estimates it yields are not valid. If, however, the model does fit the ROCs, then the derived parameter values of R and d' will serve as estimates of recollection and familiarity. Based on the previous findings that both recollection and familiarity were disrupted in amnesics (i.e., in the process dissociation and remember-know experiments), we expected amnesia to be associated with a decrease in both R and d'.

In summary, in the current experiment, we examine ROCs in amnesics and controls for words encoded under deep- and shallow-processing conditions. The dual-process model predicts that the ROCs for the controls should be asymmetrical and the ROCs for the amnesics should be more symmetrical, particularly for the shallow-encoding condition. Moreover,


the model equations should provide a good fit for the functions from both groups. Finally, if the model does fit the ROCs, then it should provide unbiased estimates for recollection and familiari ty in the process dissociation and the remember -know procedures and the ROCs can be used to estimate recollection and familiarity.

We examined recognition memory in stroke patients with unilateral damage to the left medial temporal lobe (the patients are described in more detail later). These patients were judged to be suitable for the current experiment because they were intellectually well preserved, therefore they should not have difficulty making confidence judgments. Moreover, because patients with unilateral medial temporal damage are generally not as severely amnesic as patients with bilateral damage, examination of these unilateral patients should provide a particularly strong test of the claim that amnesics have a deficit in familiarity. An important l imitation to using patients with left medial temporal damage due to stroke is that their damage may include the splenium of the corpus callosum and thus these patients often exhibit reading problems. To circumvent these difficulties in the current study, we presented the stimuli auditorally.

M e t h o d

Par t i c ipan t s

We tested 3 amnesic patients. Table 5 presents the Wechsler Adult Intelligence Scale Revised (WAIS-R; Wechsler, 1981) and the Wechsler Memory Scale--Revised (WMS-R; Wechsler, 1987) scores for the patients and shows that they exhibited normal levels of intelligence despite severe long-term memory deficits for verbal material. Of most relevance for the current study, the patients' performance on the delayed verbal paired associates test (i.e., WMS-R delayed) was approximately 3 SD below normal, indicating that the patients were severely amnesic. The standardized scores were -4.92, -2.64, and -2.64, for patients A.L., W.M., and E.M., respectively. Aside from these memory problems, the patients were capable of understanding complicated instructions and of carrying on intelligent conversations with the examiner.

The patients had lesions to the left medial temporal lobe because of embolus or atherosclerotic occlusion of the posterior cerebral artery (see Figure 2 for lesion reconstructions of the patients). The effects from unilateral posterior cerebral infarction and the resulting persistent acute anterograde amnesia have been reviewed elsewhere (von Cramon, Hebel, & Schuri, 1988; DeRenzi, Zambo- lin, & Crisi, 1987; Ott & Shaver, 1993; Kroll, Knight, Metcalfe, Wolf, & Tulving, 1996). The patients' lesions centered in the posterior hippocampal region, including hippocampus proper, dentate gyrus, subiculum, parahippocampal gyri, entorhinal cortex, and fornix (see Eichenbaum, Otto, & Cohen, 1994, for a parcella-

tion of anatomical subregions of the medial temporal lobe). Variable amounts of damage to adjacent fusiform, lingual, and calcarine gyri existed in individual patients. This form of infarction results in severe unilateral mammillary body atrophy (Efanov, Knight, & Amaral, 1997). The patients all had variable degrees of homonymous field defects due to calcarine damage. Two of the patients also suffered damage to the splenium of the corpus callosum resulting in some degree of alexia without agraphia. One patient (A.L.) could read only a few letters at a time and the other patient (E.M.) could only read a syllable at a time. Three healthy participants from the local community served as controls. They were matched to the amnesics in terms of sex, age, and years of education.

M a t e r i a l s

We selected 480 words from the Toronto word pool. The words were randomly divided into two sets to serve in Sessions 1 and 2. Each set was randomly divided into three blocks. Block 1 served as the first study list, Block 2 served as the second study list, and Block 3 served as nonstudied lure items. The test list consisted of a random mixture of the items from the three blocks.

D e s i g n a n d Procedure

Amnesics and age-matched controls were tested in two 1-h sessions. In the first session, participants heard one list of words that they processed under deep-encoding conditions (i.e., make an abstract-concrete judgment about each word) followed by a separate list that they processed under shallow-encoding conditions (i.e., count the number of syllables in each word). The order of the encoding conditions was reversed for the second session. Partici- pants then listened to a test list and made recognition memory judgments using a 6-point confidence scale from certain it was new (1) to certain it was old (6). The study and test phases were self-paced. The participants responded verbally and the experi- menter recorded their responses. Participants were instructed to spread their responses across the whole range from 1 through 6 and were reminded of this after they made their first few responses. Pilot studies with healthy controls suggested that these instructions were necessary to avoid having participants use only high- confidence or low-confidence responses. Failure to use the entire range of response confidence would lead to ROCs in which the points were closely clustered together, thus making the assessment of the function difficult.

Resu l t s and D i s c u s s i o n

The R O C s

The average ROCs for deep- and shallow-encoding conditions for amnesics and controls are presented in the left

Table 5 Characteristics o f Amnesic Patients

WAIS-R WMS-R Age at Age at Dominant

Patient Sex test (yr) lesion (yr) hand Verbal/Performance/Full Verbal/Visual/General/Attention/Delayed

A.L. M 62 58 Right 96/88/92 73/81/81/85/69 W.M. M 71 56 Right 118/112/117 71/92175/88/63 E.M. M 78 76 Left 113/109/112 82/92/81/93/69

Note. WAIS-R = Wechsler Adult Intelligence Scale--Revised; WMS-R = Wechsler Memory Scale--Revised; M = male.


Figure 2. Computer reconstructions of computerized tomography or magnetic resonance brain scans for 3 patients (identified by initials) with hippocampal lesions resulting from infarctions of the posterior cerebral artery. Dark areas represent site of lesion on transverse sections. The lateral view illustrates the level and orientation of each section from the most ventral section (1) to the most dorsal section (7). A coronal cut is not provided because the amygdala was not damaged in this group, and a lateral view is not provided because there was no damage to the anterior temporal lobe.

panel of Figure 3, and the z-transformed ROCs (z-ROCs) are presented in the right panel. The number of responses in each confidence category are presented in Appendix B. An examination of the ROCs showed that the controls performed better than the amnesics in both the deep- and shallow-encoding conditions (i.e., the ROCs of the amnesics fell below those of the controls). Moreover, the controls

exhibited better performance for the deep than shallow conditions. The amnesics' ROCs for shallow encoding fell slightly below that of the deep condition on the left side of the ROC. However, the functions overlapped, indicating that anmesics did not show a consistent benefit for deep- encoding over shallow-encoding conditions.

It is important to note that the ROCs for the amnesics

ROC 1o

0.8

.~0 .6

~" 0.4

0.2

0 0 0.2 0.4 0.6 0.8 1.0

P('yez'lnew)

2.5

1.5

0.5 |

-o.s

- I .5

-2.5

-2.5

Z-ROC

/

r / / I I 1 I

-1.5 -0.5 0.5 1.5 2.5

Z(~es'Inew)

Figure 3. The recognition memory receiver operating characteristics (ROCs) for amnesics and controls for deep- and shallow-processing conditions plotted in probability space (left panel) and z space (fight panel). The amnesics' ROCs are symmetrical (i.e., the slopes of their z ROCs are close to 1). The control participants' ROCs are asymmetrical (i.e., the slopes of their z ROCs are less than 1).


were symmetrical in contrast to the asymmetrical ROCs of the controls. The symmetry of the functions was assessed in two ways. First, the ROCs were plotted in z space. A perfectly symmetrical ROC will have a slope of 1.0 in z space. Linear regressions of the z ROCs showed that the amnesics' z ROCs had slopes very close to 1.0--their slopes for the deep and shallow z ROCs were 0.90 and 1.06, respectively. In contrast, the slopes of the controls' z ROCs were considerably less than t.0---the deep- and shallow- encoding z ROC slopes were 0.66 and 0.71, respectively. Second, the log likelihood estimation method of Ogilvie and Creelman (1968) produced similar slope estimates and showed that the slope of the amnesics' z ROCs did not differ significantly from 1, but the slope of the controls' z ROCs did differ significantly from 1.

We conducted several subsequent analyses to assess whether the amnesics and controls made use of the entire response confidence scale and to determine whether averaging or practice effects influenced the observed ROCs. An examination of the number of items in each response confidence category (see Appendix B) showed that the amnesics and controls used the entire range of response confidence and neither group perseverated on any one response category. Thus, the instructions to participants to use the entire range of responses was effective for both amnesics and controls. To assess whether practice had any effects on the nature of the observed ROCs, we plotted performance separately for the first and second sessions. The same pattern of ROCs was observed across sessions, suggesting that practice did not critically influence the ROCs and that averaging over sessions did not influence the shape of the ROCs. To assess within-session practice effects, we plotted ROCs for the first and second half of the test session. An examination of those ROCs showed that although performance was slightly better for the first half of the session than the second half for both groups, the shapes of the ROCs did not change. We further assessed the effects of averaging by examining the ROCs for each participant. The ROCs for the individuals within each group were remark- ably similar, and the average ROCs were representative of the individual participants.

The ROC symmetry for the amnesics and asymmetry for the controls is in agreement with the predictions of the dual-process signal-detection model. Although the slope of the deep-encoding condition z ROC for the amnesics was slightly less than that of the shallow-encoding condition, neither deviated significantly from 1. This indicates that amnesia virtually eliminated recollection for both deep-and shallow-encoding conditions. Note, however, that the slope of the deep-encoding z ROC was slightly less than 1, leaving open the possibility that recollection may not have been entirely eliminated in this condition.

Could the symmetry observed in the amnesics' ROCs have reflected the fact that their performance was lower than the controls? As performance approaches chance, the observed ROC will approach the positive diagonal and the ROCs will be forced to become more symmetrical. Thus, low levels of performance may be associated with more

symmetrical ROCs. However, an examination of the exist- ing ROC literature indicated that the performance of the amnesics was high enough to avoid such floor effects. For example, Ratcliff, McKoon, and TindaU (1994) found asymmetrical ROCs for healthy participants even when performance was very low. That is, asymmetrical ROCs were observed under conditions where the intercept of the z ROC was close to 0.5, which is comparable to the intercepts of the amnesics' z ROCs in the current experiment (the intercepts for the amnesics' z ROCs were slightly greater then 0.5).

To further assess the possibility that the low level of performance observed for the anmesics influenced the shape of their ROCs, we tested a second control group under conditions where performance was as low as that seen with the amnesics. Six undergraduates were tested using the same procedure, except that the study words were read at a rate of 1 s per word and each participant was tested in only one session. Data were collapsed across the deep and shallow conditions, because at the rapid rate of presentation participants were often unable to accurately perform the processing tasks. The ROC data from the amnesics and the controls are presented in Figure 4. Examination of Figure 4 shows that the two groups performed approximately equally well (the ROCs overlapped). The controls exhibited an asymmetrical ROC, with a z ROC slope of 0.55 compared to the amnesics' z ROC slope of 0.97. Thus, even when the overall level of performance was similar for the two groups, the ROCs of the controls still exhibited an asymmetry that was not seen with the amnesics.

Fitting the ROCs

We fit the dual-process signal-detection model to the deep- and shallow-encoding ROCs for the amnesics and controls by using a search algorithm that reduced the sum of the squared errors between the predicted function and the observed ROC points (see Appendix A for a full description of the curve-fitting procedure). Figure 5 shows the fit of the model to the observed ROC data. An examination of Figure 5 shows that the model provided a good fit for the ROCs of the amnesics and controls.

An examination of the ROCs also shows why the threshold models did not provide a reliable way of incorporating response bias in the previous studies. The threshold models predict a linear relationship between hits and false- alarm rates (see Murdock, 1974), resulting in a linear ROC, unlike the curved ROCs observed for both amnesics and controls. Because the models do not provide an accurate account of memory and response bias, they conflate estimates of memory with response bias. Similar criticisms have been raised about threshold models before (e.g., Murdock, 1974; Kinchla, 1994).

Estimating Recollection and Familiarity

Estimates of recollection and familiarity for the amnesic and control participants are presented in Table 6. The estimates are based on the parameter values obtained when


1.0

0.8

--~'0.6 o . . . . .

~o.4 Q .

0.2

0

ROC Z-ROC 2.5

i I I I

0.2 0.4 0.6 0.8 1.0 P('yes'lnew)

1.5

~'0.5

v

N

-1.5

-2.5 -2.5

j -1.5 -0.5 0.5

Z('yes'lnew)

I

1.5 2.5

Figure 4. Average recognition receiver operating characteristics (ROCs) for control participants who were tested under fast encoding conditions plotted with the ROCs for amnesics tested under the slower encoding conditions. Overall performance of the two groups is similar, but the ROC for the controls is asymmetrical in contrast to the symmetrical ROC of the amnesics.

the dual-process equations were fit to the observed ROC in Figure 5. An examination of Table 6 shows, as expected, that the amnesics exhibited a deficit in recollection for both deep and shallow conditions. The deficits in the recollection estimates for amnesics, compared to that for the original controls, were 0.27 (100%) and 0.30 (68.2%) for shallow and deep conditions, respectively. There were also deficits in their estimates of familiarity of 0.10 (16.4%) and 0.48 (52.7%) for shallow and deep conditions, respectively.

In summary, the results of the current study are in agreement with the predictions of the dual-process model. The ROCs of healthy participants were curvilinear and asymmetrical. In contrast, the ROCs for the amnesics were curvilinear and symmetrical. The ROC for the deep- encoding conditions was slightly less symmetrical than that for the shallow conditions for the amnesics, however, neither ROC deviated significantly from symmetry. In contrast, even when the task for controls was made more difficult so that their performance was as low as that of the amnesics, the healthy participants still exhibited an asymmetrical ROC. The dual-process model was found to provide a good fit for the observed ROCs, demonstrating that it provided an accurate way of incorporating response bias. Thus, the results support the use of the model in examining performance of amnesics' and controls' recognition memory and support the conclusions of the reanalysis of the process dissociation and remember-know studies. Finally, fitting the model to the ROC data showed that amnesia was associated with a decrease in the estimates of recollection and familiarity.

Conclusions

Previous studies of amnesia were in agreement in showing that amnesics exhibit a deficit in recollection. However, these studies were not in agreement with respect to the fate of familiarity. Results from an experiment using the process dissociation procedure (Verfaellie & Treadwell, 1993) suggested that familiarity was not consistently disrupted in

amnesia. However, on the basis of the results from remember- know studies, familiarity was disrupted in amnesia (Knowl- ton & Squire, 1995), was enhanced in amnesia (Schacter et al., 1996), or was inconsistently affected by amnesia (Schac- ter et al., 1997).

Interpretation of the results of all of these studies was complicated by the fact that amnesics exhibited inflated false-alarm rates compared to controls. Thus, it was necessary to develop a model that explicitly incorporated response bias to interpret the data. A reexamination of the data with a threshold model failed to resolve the inconsistencies in the literature. In contrast, the dual-process signal-detection model indicated that amnesia was associated with a decrease in recollection and a smaller, but consistent, decrease in familiarity.

We tested the dual-process model by examining ROCs in amnesic and control participants. The ROCs of healthy participants were asymmetrical, and the ROCs of the amnesics were symmetrical, showing a dissociation between the two groups. The results are in agreement with the dual-process model in finding that controls were able to use both familiarity and recollection to make recognition judgments, but amnesics appeared to be limited primarily to familiarity-based judgments. These results converge with numerous other studies finding that amnesics rely primarily on familiarity (e.g., Aggleton & Shaw, 1996; Huppert & Piercy, 1978).

The dual-process model fit the observed ROC data quite well, showing that it provided an accurate account of memory and response bias. These results support the use of the model when recognition performance is examined, and, thereby, justify our use of the model to reexamine the previous studies.

Finally, the parameters derived from the ROC analysis indicated that both recollection and familiarity were disrupted in amnesia. Thus, the ROC analysis, the process dissociation procedure, and the remember-know procedure


converged in showing that amnesics exhibit a reduction in recollection and familiarity.

Although recollection and familiarity were both disrupted in amnesics, recollection appeared to be most critically influenced by amnesia. That is, estimates of recollection often approached 0 in the amnesics (see Tables 1-4 and 6), whereas estimates of familiarity never did so. Nonetheless, we emphasize that there was a decrease in familiarity in every condition in every experiment, which indicates that a disruption of familiarity is a pervasive characteristic of amnesia.

Cortical Structures Important for Recollection and Familiarity

The four previous studies included a mixture of patients with different etiologies, including patients with medial temporal lobe damage and Korsakoff patients. Korsakoff patients are known to have widespread cortical damage including frontal regions i n addition to the mammillary bodies and the dorsomedial nucleus of the thalamus. The deficits seen in those experiments may have reflected frontal damage. However, Knowlton and Squire (1995) examined the performance of Korsakoff patients separately from other amnesics and found a similar pattern for both groups. Moreover, Schacter et al. (1997) examined performance of the Korsakoffpatients and found that they did not differ from the non-Korsakoff patients. Thus, the frontal atrophy associated with Korsakoff patients is not necessary to exhibit

1.0

0.8

_ 06 o'0.4

0 . 2

0 ,

--~- Amneslc/Sh, allow

0 0.2 0.4 0.6 0.8 1.0 P('yes'lnew)

Figure 5. The recognition receiver operating characteristics (ROCs) of the amnesics and control participants for deep- and shallow-encoding conditions fit to the dual-process signal- detection model. The model provided an accurate account of the observed ROCs for the amnesics and controls•

Table 6 Estimates for Recollection and Familiarity for the Controls (Con) and Amnesics (Amn) Based on the Average Receiver Operating Characteristics (ROCs) for Deep- and Shallow-Encoding Conditions

Shallow Deep

Estimate Con Amn Con Amn


Note. The estimates of recollection and familiarity were lower for the amnesics than for the controls.

deficits in recollection and familiarity. The stroke patients examined in the current experiment had unilateral damage to the medial temporal lobes, but 2 patients had splenial damage leading to difficulties with reading but not with comprehension or writing. Although it is conceivable that this additional damage may have contributed to the patients' visual memory deficits, aural presentation of the stimuli should have circumvented any simple visual deficits. In any case, taken together with the results from the previous studies, the results show that medial temporal regions play an important role in both recollection and familiarity.

The Role of Medial Temporal Regions in Familiarity Judgments

Numerous previous studies have shown that the medial temporal lobes play a critical role in conscious recollection, however, it is much less clear in what way they support familiarity-based recognition judgments. The preservation of some familiarity-based recognition in amnesia, even when recollection approached 0, suggests that some processes supporting familiarity are not disrupted by medial temporal lobe damage. The results suggest that familiarity is not a unitary process. Jacoby has suggested that participants may make use of perceptual as well as conceptual fluency as a basis for recognition judgments (see Jacoby, 1991; Jacoby & Dallas, 1981). In addition, familiarity judgments may in part reflect a novelty detection process. That is, items that are perceived as particularly novel at test may be rejected as new (e.g., the item seems novel, thus it probably was not in the study list). It is possible that the familiarity deficit seen in amnesics is related to a deficit in their ability to detect novelty. In support of this claim, Knight (1996) has found that amnesics with posterior hippocampal damage (including the 3 patients reported here) exhibit reduced orienting responses to novel or unusual items. If healthy participants can make use of this orienting response as a basis for recognition judgments and amnesics are unable to, then we would expect the estimates of familiarity to be lower for the amnesics than the controls. This could occur even if perceptual and conceptual fluency were preserved in these patients. A deficit in novelty detection would also explain the increased false-alarm rate that is typically seen with amnesics.


Functionally Distinct Areas Within the Medial Temporal Lobe

One intriguing possibility is that recollection and familiarity may be dependent on the hippocampus and parahippocampal region, respectively. The belief that the hippocampus is important for conscious recollection is supported by the finding that lesions restricted to field CA 1 of the hippocampus are sufficient to cause severe deficits on memory tasks such as free recall that require recollection (e.g., Zola- Morgan, Squire, & Amaral, 1986). The belief that the hippocampus is less important for familiarity than recollection is supported by the observation that amnesics with lesions restricted to the hippocampus-fornix-mammillary body system exhibit relatively preserved forced-choice recognition memory performance despite their severe deficits in free recall (Aggleton & Shaw, 1996). The hippocampal patients' ability to perform well on the forced-choice recognition test may reflect the fact that the task relies heavily on familiarity.

The idea that familiarity may be subserved by the parahippocampal region is in general agreement with the observation that the current patients, who all: had para.hippocampal lesions, exhibited deficits in familiarity. Further support for this idea is gained from the results of lesion and single-cell recording studies in animals showing that the parahippocampal region is important for discriminating between familiar and novel stimuli (for a review, see Mishkin & Murray, 1994).

Finally, a recent functional magnetic resonance imaging (MRI) study with humans (Gabrieli, Brewer, Desmond, & Glover, 1997) showed that studied items were associated with bilateral anterior hippocampal activation (i.e., subiculum) relative to items that were new to the experiment and that novel items, compared with repeated items, were associated with bilateral posterior medial temporal activation focused in the parahippocampal cortex. The authors suggested that the parahippocampal activation may have reflected familiarity-based processes in the sense that the region was more active for the novel than for the familiar items; the hippocampal activation may have reflected recollection, because this region was more active for repeated than for new items. Future studies contrasting patients with damage to these two regions should be useful in determining how these structures are related to recollection and familiarity.

Levels-of-Processing Effects

The current results showed that deep, compared to shallow, processing led to an increase in recognition memory performance in controls but had a much reduced effect on amnesics. Although the amnesics manifested a slightly higher ROC after deep encoding than after shallow encoding, the functions overlapped, Previous studies examining the effects of levels of processing on amnesics' recognition have generally found that amnesics do exhibit a levels-of- processing effect when recognition is tested shortly after

study and the level of performance is high~ However, because performance falls with longer delays, the magnitude of the levels-of-processing effects decreases, and effects are sometimes absent (e.g., Cermak & Reale, 1978; Graf, Squire, & Mandler, 1984; Mayes, Meudett, & Neary, 1980). Whether K s reflects the practical problem of detecting a levels-of-processing effect when overall performance is quite low or whether amnesics show a disproportionately fast loss of semantically encoded information is not clear.

The levels-of-processing effect seen in the control participants is in agreement with recent studies using the process dissociation procedure with healthy participants. For example, Toth (1996) showed that deeper levels of processing lead to an increase in both recollection and familiarity, and Iacoby (i991) found that words solved as anagrams led to higher estimates of familiarity and recollection than those simply read aloud at study. However, in the current experiment, levels of processing had no discernible effect on the familiarity component of the amnesics' recognition performance. What little effect it had was on their recollection component. If recollection processes in amnesics are particularly short lived, as they appear to be, then this would explain the short duration of their levels-of-processing effects. Of course, this does not explain why the amnesics do not show a levels-of-processing effect: in their familiarity component. Given the very small levels-of-processing effect in the amnesic group, however, it may be unwise to draw strong conclusions about the effects of levels of processing on the separate recognition processes in this group.

Model Assumptions and Measurement Tools

The dual-process model that we examined in the current study is based on the assumption that recollection and familiarity represent two independent memory processes. Such an assumption can of course be questioned and, if future studies find that the two processes are dependent, it may be appropriate to modify the model equations, How- ever, the conclusion that amnesia is associated with a decrease in recollection and familiarity holds even i f the two processes are positively related. For example, if recollection increases the likelihood of familiarity, or vice versa, then the independence assumption would lead to an underestimation of the familiarity scores. Because recollection was always greater for the controls we would tend to underestimate their familiarity scores to a greater extent than for the amnesics. Thus, the decrease in familiarity that we observed would underestimate the true decrease, Alternatively, if the two processes were negatively related, that is, recollection leads to a decrease in familiarity, then we would overestimate amnesics' deficit in familiarity. However, we know of no evidence that would suggest that these memory processes are operating in such a manner. The observations that the predictions of the independence-based: m ~ l were verified and the model provided an accurate account for the observed ROCs suggest that the independence assumption is reasonable.

The assumption that familiarity behaves like a signal-


detection process was supported by the observation that the amnesics' ROCs were curvilinear and symmetrical. The alternative assumption, which underlies the hits-minus-false- alarms method of incorporating false-alarm rates, relies on a threshold model. Although this method is extremely simple to use, the model predicts linear ROCs that were contra- dicted by the observed ROCs, showing that the method was not appropriate for either amnesics or controls. The inaccura- cies of the threshold model led the hits-minus-false-alarm- rate method of incorporating response bias to inconsistent sets of conclusions, and we believe that this method should be avoided.

An important assumption that was not directly assessed in the current study is the assumption that recollection is a threshold process and thus can be measured as a simple probability. That is, participants are assumed to either recollect qualitative information about a prior event or fail to do so. Of course participants can retrieve many different things about a study event; however, if the model is correct, then there should be conditions where recollection falls completely and participants cannot recollect anything about a previous event. The observation that the model provided a good fit for the controls' ROCs based on this assumption suggests that it is not unreasonable. Moreover, other studies have provided more direct support for the threshold assumption. If recollection is a threshold process, then it should be possible to observe relatively linear ROCs under conditions where performance relies primarily on recollection. Such ROCs have been reported for recognition of source (Yonelinas, 1998) and recognition of associative information (Yonelinas, 1997). Interestingly, amnesics are found to perform extremely poorly on tests that require the recollection of source information (e.g., Hirst, 1982; Mayes, Meud- ell, & Picketing, 1985) and associative information (e.g., Kroll et al., 1996; Reinitz, Verfaellie, & Milberg, 1996). The severe recollection deficits that the amnesics exhibited in the current study would explain why these patients do so poorly on tests of source memory and associative recognition.

In furthering our understanding of human memory, it is important that we develop a collection of measurement procedures that converge on a stable set of findings. To accomplish this, it is necessary to develop quantitative models that accurately account for memory and response bias. This is particularly important when patient and control populations show differences in false-alarm rates, such as those seen between amnesics and controls. We believe that the dual-process signal-detection model along with the process dissociation, remember-know, and ROC procedures represent a powerful set of tools for examining the processes underlying memory performance in healthy and memory- impaired patients.

References

Aggleton, J. P., & Shaw, C. (1996). Amnesia and recognition memory: A re-analysis of psychometric data. Neuropsychologia, 34(1), 51-62.

Atkinson, R. C., & Juola, J. E (1974). Search and decision processes in recognition memory. In R. C. Atkinson, R. D. Luce, D. H. Krantz, & P. Suppes (Eds.), Contemporary developments in mathematical psychology: L Learning, memory and thinking (pp. 243-293). San Francisco: Freeman.

Cermak, L. S., & Reale, L. (1978). Depth of processing and retention of words by alcoholic Korsakoff patients. Journal of Experimental Psychology: Human Learning and Memory, 4(2), 165-174.

Cermak, L. S., Verfaellie, M., Butler, T., & Jacoby, L. L. (1993). Attributions of familiarity in amnesia: Evidence from a fame judgment task. Neuropsychology, 7, 510-518.

Cermak, L. S., Verfaellie, M., Sweeney, M., & Jacoby, L. L. (1992). Fluency versus conscious recollection in the word completion performance of amnesic patients. Brain and Cognition, 20, 367-377.

Cohen, N. J., & Squire, L. R. (1980). Preserved learning and retention of pattern-analyzing skill in amnesia: Dissociation of knowing how and knowing that. Science, 210, 207-209.

Deese, J. (1959). On the prediction of occurrence of particular verbal intrusions in immediate recall. Journal of Experimental Psychology, 58, 17-22.

DeRenzi, E., Zambolin, A., & Crisi, G. (1987). The pattern of neuropsychological impairments associated with left posterior cerebral artery infarcts. Brain, 110, 1099-1116.

Efanov, M., Knight, R. T., & Amaral, D. (1997). Effects of hippocampal lesions on the medial mammilary nucleus in humans and non-human primates. Society for Neuroscience, 23, 774.

Eichenbaum, H., Otto, T., & Cohen, N. (1994). Two functional components of the hippocampal memory system. Behavioral and Brain Sciences, 17, 449-518.

Gabrieli, D. E., Brewer, J. B., Desmond, J. E., & Glover, G. H. (1997). Separate neural bases of two fundamental memory processes in human medial temporal lobe. Science, 276, 264- 266.

Gardiner, J. M., & Java, R. I. (1991). Forgetting in recognition memory with and without recollective experience. Memory and Cognition, 19(6), 617-623.

Graf, E, Squire, L. R., & Mandler, G. (1984). The information that anmesic patients do not forget. Journal of Experimental Psychol- ogy: Learning, Memory, and Cognition, 10(1), 164-178.

Haist, E, Shimamura, A. E, & Squire, L. R. (1992). On the relationship between recall and recognition memory. Journal of Experimental Psychology: Learning, Memory, and Cognition, •8(4), 691-702.

Hirst, W. (1982). The amnesic syndrome: Descriptions and explana- tions. Psychological Bulletin, 91, 435-460.

Hirst, W., Johnson, M. K., Phelps, E. A., & Volpe, B. T. (1988). More on recognition and recall in anmesics. Journal of Experi- mental Psychology: Learning, Memory, and Cognition, 14(4), 758-762.

Huppert, E A., & Piercy; M. (!978). The role of trace strength in recency and frequency judgments by amnesic and control subjects. Quarterly Journal of Experimental Psychology, 30, 347-354.

Jacoby, L. L. ( 1991). A process dissociation framework: Separating automatic from intentional uses of memory. Journal of Memory and Language, 30(5), 513-541.

Jacoby, L. L., & Dallas, M. (1981). On the relationship between autobiographical memory and perceptual learning. Journal of Experimental Psychology: General, 110(3), 306-340.


Jacoby, L. L., Toth, J. P., & Yonelinas, A. E (1993). Separating conscious and unconscious influences of memory: Measuring recollection. Journal of Experimental Psychology: General, 123, 216-219.

Jacoby, L. L., Yonelinas, A. E, & Jennings, J. M. (1997). The relation between conscious and unconscious (automatic) influences: A declaration of independence. In J. D. Cohen & J. W. Schooler (Eds.), Scientific approaches to consciousness (pp. 13-48). Hillsdale, NJ: Erlbaum.

Kinchla, R. A. (1994). Comments on Batchelder and Riefer's multinomial model for source monitoring. Psychological Re- view, 101(1), 166-171.

Knight, R. T. (1996). Contribution of human hippocampal region to novelty detection. Nature, 383, 256-259.

Knowlton, B. J., & Squire, L. R. (1995). Remembering and knowing: Two different expressions of declarative memory. Journal of Experimental Psychology: Learning, Memory, and Cognition, 2•(3), 699-710.

Kroll, N. E. A., Knight, R. T., Metcalfe, J., Wolf, E. S., & Tulving, E. (1996). Cohesion failure as a source of memory illusions. Journal of Memory and Language, 35(2), 176-196.

Mandler, G. (1980). Recognizing: The judgment of previous occurrence. Psychological Review, 87(3), 252-271.

Mayes, A. R., Meudell, P., & Neary, D. (1980). Do amnesics adopt inefficient encoding strategies with faces and random shapes? Neuropsychologia, 18(4-5), 527-540.

Mayes, A. R., Meudell, E R., & Pickering, A. (1985). Is organic anmesia caused by a selective deficit in remembering contextual information? Cortex, 21, 167-202.

Mayes, A. R., Van Eijk, R., & Isaac, C. L. (1995). Assessment of familiarity and recollection in the false fame paradigm using a modified process dissociation procedure. Journal of the Interna- tional Neuropsychological Society, 1, 469-482.

Mishkin, M., & Murray, E. A. (1994). Stimulus recognition. Current Opinion in Neurobiology, 4, 200-206.

Murdock, B. B. (1974). Human memory: Theory and data. Potomac, MD: Erlbaum.

Ogilvie, J. C., & Creelman, C. D. (1968). Maximum likelihood estimation of receiver operating characteristic curve parameters. Journal of Mathematical Psychology, 5, 377-391.

Ott, B. R., & Shaver, J. L. (1993). Unilateral amnesic stroke. Stroke, 24, 1033-1042.

Rajaram, S., & Roediger, H. L. (1997). Remembering and knowing as states of consciousness in recognition. In J. D. Cohen & J. W. Schooler (Eds.), Scientific approaches to consciousness (pp. 213-240). Hillsdale, NJ: Erlbaum.

Ratcliff, R., McKoon, G., & Tindall, M. (1994). Empirical general- ity of data from recognition memory receiver-operating characteristic functions and implications for the global memory models. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20(4), 763-785.

Ratcliff, R., Sheu, C. E, & Gronlund, S. D. (1992). Testing global memory models using ROC curves. Psychological Review, 3, 518-535.

Reinitz, M. T., Verfaellie, M., & Milberg, W. E (1996). Memory conjunction errors in normal and amnesic subjects. Journal of Memory and Language, 35, 286-299.

Richardson-Klavehn, A., & Bjork, R. A. (1988). Measures of memory. Annual Review of Psychology, 39, 475-543.

Roediger, H. L. (1990). Implicit memory: Retention without remembering. American Psychologist, 45(9), 1043-1056.

Roediger, H. L., & McDermott, K. B. (1994). Creating false memories--remembering words not presented in lists. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 803-814.

Schacter, D. L., Chiu, C. Y. E, & Ochsner, K. N. (1993). Implicit memory--a selective review. Annual Review of Neuroscience, 16, 159-182.

Schacter, D. L., Verfaellie, M., & Aries, M. D. (1997). Illusory memories in amnesic patients: Conceptual and perceptual false recognition. Neuropsychology, 11, 331-342.

Schacter, D. L., Veffaellie, M., & Pradere, D. (1996). Neuropsychol- ogy of memory illusions--false recall and recognition in amnesic patients. Journal of Memory and Language, 35, 319-334.

Toth, J. E (1996). Conceptual automaticity in recognition memory: Levels-of-processing effects on familiarity. Canadian Journal of Experimental Psychology, 50(1), 123-138.

Tulving, E. (1985). Memory and consciousness. Canadian Psychol- ogy, 26(1), 1-12.

Veffaellie, M. (1994). A re-examination of recognition memory in amnesia: Reply to Roediger and McDermott. Neuropsychology, 8(2), 289-292.

Verfaellie, M., & Treadwell, J. R. (1993). Status of recognition memory in amnesia. Neuropsychology, 7(1), 5-13.

Von Cramon, D. Y., Hebel, N., & Schuri, U. (1988). Verbal learning and memory in unilateral posterior cerebral infarction. Brain, 111, 1061-1077.

Wechsler, D. (1981). WAIS-R manual. New York: The Psychologi- cal Corporation.

Wechsler, D. (1987). Wechsler Memory Scale--Revised [Manual]. San Antonio, TX: The Psychological Corporation.

Yonelinas, A. E (1994). Receiver-operating characteristics in recognition on memory: Evidence for a dual-process model. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20(6), 1341-1354.

Yonelinas, A. E (1997). Recognition memory ROCs for item and associative information: The contribution of recollection and familiarity. Memory and Cognition, 25, 747-763.

Yonelinas, A. E (1998). The contribution of recollection and familiarity to recognition and source memory: A formal model and an ROC analysis. Unpublished manuscript.

Yonelinas, A. E, Dobbins, I., Szymanski, M. D., Dhaliwal, H. S., & King, L. (1997). Signal detection, threshold, and dual-process models of recognition memory: ROCs and conscious recollection. Consciousness and Cognition, 5, 418-441.

Yonelinas, A. E, & Jacoby, L. L. (1995). The relation between remembering and knowing as bases for recognition: Effects of size congruency. Journal of Memory and Language, 34(5), 622-643.

Yonelinas, A. E, & Jacoby, L. L. (1996). Response bias and the process-dissociation procedure. Journal of Experimental Psychol- ogy: General, 125(4), 422-434.

Yonelinas, A. E, Regehr, G., & Jacoby, L. L. (1995). Incorporating response bias in a dual-process theory of memory. Journal of Memory and Language, 34(6), 821-835.

Zola-Morgan, S., Squire, L. R., & Amaral, D. G. (1986). Human amnesia and the medical temporal region: Enduring memory impairment following a bilateral lesion limited to field CA1 of the hippocampus. Neuroscience, 10, 2950-2967.


Appendix A

Estimating Recollection and Familiarity in Recognition Memory With the Dual-Process Signal-Detection Model

Using the Process Dissocia t ion Procedure

Familiarity is assumed to reflect a Gaussian equal variance signal-detection process such that the probability that an item is accepted on the basis of familiarity is a function of sensitivity (d') and response criterion (c). Sensitivity reflects the degree to which the old items are more familiar than the new items (i.e., the distance between the means of the two distributions is d') . A participant's response criterion reflects that participant's bias. That is, a low criterion results in a bias toward yes responses and a high criterion results in a bias toward no responses. The probability that an old item is accepted on the basis of familiarity is equal to qb ( d ' 1 2 - c); the proportion of the old item distribution exceeds the response criterion (c). Substituting this equation into the process dissociation equations:

P("yes" Iold)inc = R + (1 - R)~(d ' /2 - Ci,c)

and

P("yes" [old)ex ~ = (1 - R)~(d ' /2 - Cex~).

Note that there are separate c terms for the inclusion and exclusion conditions to accommodate different response criteria in the two conditions.

The probability that a new item will be accepted on the basis of familiarity is equal to @ ( - d ' 2 - c); the proportion of the new item distribution exceeds the response criterion. Thus, the probability of accepting new items under inclusion and exclusion conditions can be represented as:

P("yes" Inew)i~c = d p ( - d ' 1 2 - Cinc)

and

P("yes" IneW)exc = d P ( - d ' / 2 - C~x¢).

Given the inclusion and exclusion scores for old and new items, it is possible to solve the four equations to determine R, d ' , cite, and C~xc. However, given the nature of the normal distributions underlying signal-detection theory, a simple algebraic solution is not possible and a search algorithm must be used. For example, the solver available in Excel can be used to find the best fitting parameters for these equations by reducing the sum of squared errors between the predicted and observed data. This algorithm, as well as a compiled Pascal program described in Yonelinas, Regehr, and Jacoby (1995), are available on request.

Us ing the R e m e m b e r - K n o w Procedure

True recollection is estimated by subtracting the proportion of false remember responses from the proportion of true remember responses then dividing by the opportunity to observe a true remember response, R = (Rold -- Rnew)/(1 - Rnew). Simply subtracting false alarms from hits often leads to very similar estimates. However, in keeping with the model used with the process dissociation procedure, we use the high-threshold model. The high-threshold model assumes that, although new items may seem familiar, as in signal-detection theory, participants cannot truly

recollect new items (however, see Yonelinas, 1997, for a discussion of possible exceptions).

To calculate familiarity, it is necessary to determine the probability of correctly accepting an old item on the basis of familiarity and of incorrectly accepting a new item on the basis of familiarity. Because participants are instructed to respond " K " if the item is familiar and not recollected, the probability of responding K to an old item will be equal to the probability that the old item is familiar and that it did not receive a remember response, Kold = Ford(1 -- Rold). By rearranging this equation, Fold can be calculated a Kold(1 - Rold)- Similarly, the probability that the familiarity of a new item will exceed the response criterion, Fnew will be equal to Knew/(1 - Rnew). Given Fola and Fnew, standard d ' tables can be used to determine d ' .

Us ing Rece ive r Operat ing Characteris t ics

Recognition performance is assumed to be described by the same equations used for the aforementioned inclusion equations:

P("yes" Iold)i = R + (1 - R)qb(d'/2 - ci)

and

P("yes"[new) i= c b ( - d ' / 2 - ci).

These equations represent performance at 1 point on the receiver operating characteristic (ROC). Thus, an ROC with 5 points will have a set of 10 equations. Assuming that memory (R and d ' ) remains constant across the ROC and that only ci varies, then the set of equations can be solved to derive estimates of R and d ' . The solver in Excel can be used to find the best fitting parameters for these equations by reducing the sum of squared errors between the predicted and observed data. Tiffs algorithm is available on request. Note that the estimation procedure has been found to provide estimates that are similar to those using a log-likelihood estimation method (Yonelinas, 1998).

Appendix B

Counts per Confidence Category

Response category

Conditions 1 2 3 4 5 6

Alnneslcs New 33 67 " 87 97 109 87 Deep 13 35 47 80 109 196 Shallow 8 33 60 90 123 166

Controls New 164 103. 136 36 25 16 Deep 33 35 62 50 66 234 Shallow 69 42 99 63 . 60 147

R e c e i v e d Apri l 2 r, 1997 Rev i s ion rece ived Sep tember 15, 1997

A c c e p t e d Sep tember 30, 1997 •

Recollection and Familiarity Deficits in Amnesia ... and familiarity using the threshold model are presented in Table 1. According to those estimates, amnesics exhibited a deficit

Documents