Event-related potentials during conscious and automatic memory retrieval

Transcription

1 Cognitive Brain Research 10 (2000) locate/ bres Research report Event-related potentials during conscious and automatic memory retrieval * Kimberly A. Kane, Terence W. Picton, Morris Moscovitch, Gordon Winocur Rotman Research Institute, Baycrest Centre for Geriatric Care, University of Toronto, 3560 Bathurst Street, North York, Ontario, Canada M6A 2E1 Accepted 1 March 2000 Abstract The effects of study-test lags of between 0 and 32 items on conscious (C) and automatic (A) memory processes in a running word-completion task were investigated with event-related potentials (ERPs). The process dissociation procedure (PDP) can distinguish between C and A contributions to memory by comparing performance when subjects respond with either an old item (inclusion) or a new item (exclusion). C can be estimated by subtracting the probability of an intrusion of an old item during the exclusion task (due to A without C) from the probability of correctly producing an old item during the inclusion task (due to C and/or A). The behavioral results showed that C was stronger when the test item followed the studied word in the next trial or after a lag of one stimulus. The strength of A did not vary with lag. The ERP waveforms contained a broad parietal positive wave between 300 and 800 ms. This parietal wave distinguished between correctly recalled old and new words. The early portion of this old new effect was significantly affected by lag. Subtracting waveforms to obtain a measure of C revealed an effect in the later portion of this wave, lateralized over the left hemisphere. A sustained frontal negativity occurred during all recordings and was larger during conscious retrieval. There was no consistent ERP effect related to automatic memory retrieval Elsevier Science B.V. All rights reserved. Theme: Neural basis of behavior Topic: Learning and memory: systems and functions Keywords: Memory; Event-related potential; Process dissociation procedure; Lag 1. Introduction refer to specific prior episodes. Nondeclarative memories are measured by implicit tasks (lexical decision, word- Memory can be dissociated into declarative and nonde- fragment completion, stem-completion and perceptual clarative memory (e.g., Refs. [13,67]). Declarative identification) that do not directly refer to past events. memories require conscious recollection, whereas nonde- Manipulations that affect performance on implicit and clarative memories, such as those contributing to priming, explicit memory tasks are usually associated with changes skills, habits and classical conditioning, do not [64]. in automatic and conscious memory processes, respective- Declarative memories can be further subdivided into ly. semantic and episodic memories [71]. Semantic memories Performance on implicit and explicit memory tasks can are based on facts that are available to everyone, whereas be dissociated in several ways. Massed repetition [11] and episodic memories concern events that are specific to each a deep level of processing at study [17,24,55,57,60,62,73] individual s own experiences. enhance performance on explicit memory tasks more than The different types of memories are evaluated using on implicit memory tasks. On the other hand, changing the modality of presentation [24,34,63] or the physical charac- teristics [24,56] of an item between study and test impairs performance on implicit memory tasks without affecting performance on explicit memory tasks. Age decreases different types of tests [62]. Declarative memories are measured by tasks of recall and recognition that explicitly *Corresponding author. Tel.: ; fax: address: picton@psych.utoronto.ca (T.W. Picton) / 00/ $ see front matter 2000 Elsevier Science B.V. All rights reserved. PII: S (00) performance on explicit memory tasks more than on implicit memory tasks (e.g., Refs. [39,40,43]). Amnesic

2 20 K.A. Kane et al. / Cognitive Brain Research 10 (2000) patients are impaired on explicit memory tasks even ty of responding correctly on the inclusion task minus the though their performance on implicit memory tasks re- probability of responding incorrectly on the exclusion task. mains intact (e.g., Refs. [18,26,43,47,72]). A pure estimate of the probability of automatic retrieval Although automatic and conscious memory processes (A) can then be obtained by inserting the value of C into are clearly dissociable, normal individuals use both pro- Eq. (2). cesses when performing either implicit or explicit memory In the previously described experiment comparing the tasks. On implicit tasks of memory, subjects may con- effects of anagrams and simple reading on later recall [27], sciously recall material even when not asked to do so. recall performance did not differ significantly between Likewise, memories may occur automatically and indepen- anagrams and reading on the inclusion task but did so dently of conscious recall processes during explicit tasks. when examined using an exclusion task. A PDP analysis of Sometimes automatic and conscious processes may coun- the results showed clearly that anagram solving increased teract each other and experimental manipulations designed the conscious contribution to recall whereas reading into affect one type of memory may not show any significant creased the automatic contribution. effects because the manipulation causes opposing effects The equations employed in the PDP assume that conon the other type of memory. Jacoby et al. [27] studied the scious and automatic processes are independent [28,25]. effects on later recall of constructing a word from an This assumption is supported by experimental dissociations anagram compared to simply reading the word. The deeper of these processes. Conscious retrieval is attenuated by level of processing required by anagrams should have dividing attention at study [27,29] and facilitated by facilitated episodic recall, but recall performance was the increasing the depth of processing at study [27]. Automatic same as when the words were read. Reading facilitated the retrieval is attenuated by changing the modality of the automatic recall of the words, and this balanced the effects stimulus presentation between study and test [27]. Age of deeply processing anagrams words. decreases conscious retrieval but has little effect on Since both implicit and explicit tasks may involve automatic retrieval [23,29]. automatic and conscious processing, Jacoby and his col- The actual overlap between C and A is not known, and leagues (for review, see Ref. [23]) developed the process these processes may not be completely independent. One dissociation procedure (PDP) to separate what are assumed variant of the independence model is the redundancy to be independent contributions of each process. The PDP model [32] wherein C is considered to be a subset of A. A uses two tasks wherein automatic and conscious processes can exist in isolation but C only occurs together with A. either support or oppose each other. For example, subjects The combined probability of the two events, AC, is study a list of items and then perform either an inclusion or therefore equal to the probability of C. According to this an exclusion recall task. In the inclusion task, subjects logic, the two equations described previously reduce to the respond with an item studied previously ( OLD ). The probability of responding OLD on the inclusion task equal probability of responding correctly with an OLD item on to A and the probability of responding OLD on the the inclusion task may be due to conscious retrieval (C) exclusion task equal to A 2 C. Regardless of the model and/ or automatic retrieval when conscious retrieval fails to chosen, C can be obtained by subtracting the probability of occur, A(12 C): responding OLD in the exclusion task from the probability Probability of OLD in inclusion 5 C 1 A(1 2 C) (1) of responding OLD in the inclusion task. The difference between the two models lies in the calculation of A. The cerebral mechanisms of C and A are not clearly In the exclusion task, subjects respond with an item that understood. Event-related potentials (ERPs) may provide was not encountered during study ( NEW item). By doing some insight into when and where these processes occur in so, automatic and conscious processes are placed in the brain. ERPs have the advantage over other physiologiopposition. Conscious retrieval of the OLD item causes the cal measurements in that they can be separately analyzed subject to seek a NEW item. Automatic processes lead to on the basis of the subject s behavioral response and can an OLD (primed) word as a response. If conscious show the timing of the processes leading to the response. retrieval fails, these automatic processes will cause the ERP studies of memory [31,35,58] have demonstrated two subject to respond incorrectly with the OLD word. Theremain findings: a difference related to subsequent memory fore, the probability of responding OLD on the exclusion (D m) and the old/new effect. The Dm effect is measured at task is given by the probability of automatic retrieval study as an increased positive wave following a studied combined with the probability that conscious retrieval has stimulus that is subsequently recognized compared to a not occurred: studied stimulus that is not later recognized [16,49,50]. Probability of OLD in exclusion 5 A(1 2 C) (2) The Dm effect begins at approximately 300 ms and lasts for several hundred milliseconds. Eq. (2) can be subtracted from Eq. (1) to obtain a pure The ERP old/new effect is a late positive component estimate of the contribution of conscious processes (C). that is measured at test rather than study. This positivity is The probability of conscious retrieval equals the probabili- termed the old/ new effect because there is a greater

3 K.A. Kane et al. / Cognitive Brain Research 10 (2000) positivity for items correctly identified as old than for waveform for incorrect performance on the inclusion task items correctly identified as new (e.g. (neither C nor A) from the waveform for incorrect per- 1 [1,3,33,45,49,59,66,68,75 77]). Although many believe formance on the exclusion task (A). that it reflects conscious memory, the old/new effect can In this study, we applied the PDP technique to ERPs also occur in implicit memory tasks [7]. This finding may recorded during inclusion and exclusion tasks in order to be explained in several ways: (1) the ERP old/ new effect obtain electrophysiological indices of conscious and autois not measuring memory processes but some process that matic memory processes. Our hypothesis was that conis common to both implicit and explicit memory tasks; (2) scious recall would be associated with a late positive wave the old/ new effect observed in the implicit memory task in the ERP, and that this electrophysiological process may represent contamination by conscious processing; (3) would correlate with the usual old/ new effect. Our hypoththe ERP old/ new effect observed in explicit memory tasks eses concerning the ERP correlates of automatic retrieval may be the result of contamination from automatic pro- were much less specific: that there would be some change cessing; and (4) the late positivity may be composed of in the ERP occurring earlier than the correlates of conmore than one subcomponent. scious retrieval and that this would likely occur in more Behaviorally, an estimate of conscious memory pro- specifically visual areas of the brain. cesses can be obtained by subtracting the probability of In order to manipulate C, the time between study and responding with old (incorrect) items on the exclusion task test was varied. When short time intervals occur between from the probability of responding with old (correct) items study and test, memory relies on short-term memory rather on the inclusion task. Similarly, the ERP waveforms than a more permanent form of memory (e.g., Refs. corresponding to old item responses on the two tasks can [6,38]). Although memory performance decreases, and be subtracted to give an electrophysiological index of reaction time increases, as the number of intervening items conscious recall. This assumes that there are no other is increased between study and test, these changes are differences between the exclusion and inclusion tasks, a more rapid over the first few items than over remaining reasonable assumption for retrieval processes before the items [21,65]. Although the change in slope may be related time when a behavioral response is selected (for support of to the transfer of information from a short-term buffer to a this assumption, see Ref. [27]). In order to obtain a pure more permanent form of memory, there is little agreement measure of automatic retrieval from the probabilities of the in the literature about the nature or timing of this transfer behavioral responses, one must divide the probability of (e.g., Refs. [21,52,74]). The discrepancies within this incorrect (OLD) responses on the exclusion task by 1 2 C literature may be due to differences in task demands across (Eq. 2). This corrects for the different probabilities of the the different studies or due to a variable contamination of different processes over a block of trials. This is not time-sensitive conscious memory processes with time-inrequired for the ERPs since the electrophysiological analy- sensitive automatic memory processes. ses measure the presence or absence of the two processes This study used inclusion and exclusion tasks combined on single trials and since these single trials are then with a running word-completion task. The number of averaged to provide waveforms that are independent of the intervening items between study and test varied between 0 number of trials making up the average. Therefore, an and 32 items in the first experiment and between 0 and 8 index of automatic retrieval should be present in ERPs for items in the second experiment. The PDP was used to incorrect (OLD) responses on the exclusion task. Since the separate conscious memory processes from automatic redundancy model explicitly states that A, in the absence processes in both behavioral performance and ERPs. The of C, can be obtained from incorrect performance on the word-completion task was derived from a letter-insertion exclusion task, this electrophysiological approach is equal- task of Reingold [54]. Reingold s [54] task required ly valid for the independence model or the redundancy inserting a letter into a word to make a new word (i.e., model. The ERPs also allow us to remove processes turning cue into cure ). This task leads to PDP estimates occurring independently of memory retrieval (e.g., per- of conscious and automatic processes that are similar to ceptual processing, memory search) by subtracting the previous studies (e.g., Ref. [69]). Due to the large number of items required in the present task, the Reingold [54] task was modified by requiring the completion of words in 1 We should point out some ambiguities in the use of the word new, which a single letter was missing. For example, that the which can characterize either the test item, i.e., a word that was not studied previously, or the response on a word (or fragment) completion word cube can be made from the prompt cu ] e. task, i.e., a response that is not equivalent to a studied word. Responding with a new word can occur under several situations. A new word will be produced when the test-stimulus is new. However, correct performance 2. Methods during an exclusion task requires the subject to respond with a new word when he or she has correctly retrieved the old word on the basis of its fragment. This ambiguity could be obviated by using the words studied 2.1. Participants and unstudied, but this is only possible for explicit tasks and runs counter to the large literature on ERP old/ new effects. Sixteen young subjects (eight male) with a mean age of

4 22 K.A. Kane et al. / Cognitive Brain Research 10 (2000) years (range, years) participated in the first used in the sequences, were chosen randomly for each experiment. A further 10 subjects (five males) with an subject. average age of 25 years (range, years) participated In the second experiment, the same set of words was in the second experiment. The subjects in the first experi- used. The lists were re-organized so that the test items ment had an average of 19 (range, 14 23) years of followed the studied words by 0, 1, 2, 4 and 8 intervening education, while those in the second experiment had 17 items. This format allowed for complete and incomplete (range, 16 22). While many of the subjects in the first words to be more evenly dispersed than was the case in the experiment had frequently volunteered for previous ERP first experiment. studies (and all were participating as volunteers in the first experiment), those in the second experiment were largely 2.3. Behavioral measurements novices and were also paid for their participation ($10 an hour). All subjects were right-handed and all had normal or Button-press responses were collected and examined. corrected-to-normal vision. Subjects needing glasses wore Percent correct responses were calculated for each of the these during the experiment. stimulus types in each of the two experimental conditions. The reaction time was measured to the nearest millisecond. For each subject the modal reaction time was taken and 2.2. Experimental design these measurements were averaged across subjects. For the first experiment, six lists of 180 items were 2.4. ERP recordings created from a pool of four- to five-letter words. Each list consisted of 90 complete and 90 incomplete words (i.e., At the commencement of the experiment, subjects were words with one of the middle letters missing). The number fitted with an electrocap (Electro-Cap International) that of possible completions for the incomplete items was at contained 36 tin electrodes located according to the least two and ranged from two to nine, with a mean of International System (as modified by the American three and with only 3% exceeding five. Eighty of these EEG Society [5]). The electrodes were Fp1, Fp2, AF3, incomplete items had a completion that corresponded to AF4, AFz, Fz, F3, F4, F7, F8, FC1, FC2, FC6, FC5, C3, one of the complete words from the list (studied items), C4, Cz, T7, T8, CP1, CP2, CP5, CP6, Pz, P3, P4, T5, T6, while the remaining 10 items did not (nonstudied items). PO3, PO4, Oz, O1, O2, Cb1, Cb2 and Iz). Three additional Likewise, there were 10 complete words (unrepeated) that electrodes were placed on the left and right mastoids and did not correspond to any of the subsequent incomplete on the back of the neck at a distance of 10% of the items. nasion inion distance below the Iz electrode. Five elec- The presentation of complete and incomplete words was trodes were placed on the face to monitor eye movements. random with the constraint that test items followed their These electrodes were placed at the nasion, the suborbital respective complete word by 0, 2, 4, 6, 8, 12, 20 or 32 ridge just inferior to the eye (IO1, IO2) and at the outer intervening items (equiprobable study-test lags). Complete canthus of the eye (LO1, LO2). The skin beneath the words were therefore more frequent at the beginning of the electrodes was gently abraded with the blunt end of a list. Three different presentation sequences were created syringe. Inter-electrode impedances, measured at 20 Hz, that complied with the study-test-lag constraints. Subjects were kept below 5 kv. Electrode AFz was used as ground performed three inclusion and three exclusion tasks, each and Cz was used as reference. with a different list of items. An ordered example of list Signals were amplified with a Neuroscan SynAmp at a items and response requirements is provided in Table 1. gain of 2500 and filtered between 0.05 and Hz (23 The order of the tasks and sequences, and the actual words db points). The data were digitized with an A 2 D conversion rate of 250 Hz and saved off-line for further analysis. A separate set of ocular calibration-signals was Table 1 obtained, consisting of averaged eye movements in each of a Stimuli four directions and averaged eye blinks. A principal List Type Condition Inclusion Exclusion component analysis (PCA) of the ocular recording was item response response performed in order to determine a set of components that Cane Study Repeated Un/ pleasant Un/ pleasant best represented each subject s eye movements. Between Chick Study Repeated Un/ pleasant Un/ pleasant two and four components were used depending on their Ch ck Test Studied (lag 0) Chick Check ] b contribution to the EOG recording (.1%) and their Ca e Test Studied (lag 2) Cane Care ] b b Cu e Test Nonstudied Cute Cute relation to the EOG patterns. The scalp projections of these ] Clove Study Unrepeated Un/ pleasant Un/ pleasant components were then subtracted from experimental ERPs a This table presents an example of list items, list order and correct in order to minimize ocular contamination [9,41]. responses on the inclusion and exclusion task. The instructions for the two tasks and an example of a b Other responses are also valid. correct response were presented prior to the presentation of

5 K.A. Kane et al. / Cognitive Brain Research 10 (2000) each list. If a complete word (e.g., CUBE) occurred, two levels. Post-hoc pairwise comparisons were conducted subjects decided whether it represented something pleasant using Tukey s honestly significant differences. Statistical or unpleasant. They pressed the spacebar once they had significance was set at P,0.05. In behavioral analyses, decided and typed in the letter p for pleasant or u for repeated-measure ANOVAs considered task (inclusion unpleasant. As well, they were asked to remember these versus exclusion) or process (C versus A) by lag (lags words because most, but not all, would be presented later 0 32 or lags 0 8). in the list in an incomplete form. If an incomplete word For the behavioral measurements, the main hypothesis (e.g., CU ] E) occurred, the subjects selected a letter to was that while behavioral measures of automatic processcomplete the word. In the inclusion task they were told to ing would be insensitive to the study-test lag, conscious choose a letter that would form a word seen previously on processing would decrease with increasing lag. A secthe list ( B ) or, failing this, to complete the item with the ondary hypothesis was that there would be a distinct first word that came to mind. In the exclusion task, they step-change in the strength of conscious processing at the were told to choose a letter that completes the item with a time when memory retrieval involved long-term rather word not seen previously on the list ( R or T ). Once than short-term memory. they had selected a letter, they pressed the spacebar and Since complete ANOVAs of the ERP measurements then typed in the letter. Items remained on the screen for 8 were statistically unwieldy, we restricted our analyses to s, or until a letter was typed in. Subjects were asked to the waveforms at the electrodes where the measurements minimize blinking and to look at the keyboard only after were largest, and the corresponding contralateral electrode. the spacebar was pressed. Brief practice sessions (22 For each portion of the waveform being considered, items) occurred prior to the start of the experiment. separate repeated-measure ANOVAs at each peak were The ERP data were averaged over epochs beginning 100 conducted for both laterality (left versus right) by task by ms prior to stimulus onset and ending 4 s after stimulus lag and for laterality by task by condition (studied items onset. Averaging at each electrode was conducted for versus nonstudied items; correct inclusion versus incorrect combinations of stimulus, task, response accuracy and lag. exclusion; incorrect exclusion versus incorrect inclusion). The resultant waveforms were measured relative to the Due to the small number of behavioral errors, the ammean value in the 100 ms before the stimulus. As well, in plitudes of incorrect-trial waveforms were collapsed across Experiment 1, averaging was conducted for ranges extend- lag. Because of this, the ERP analysis of C (correct ing from 2.9 s prior to the spacebar response to 1.2 s inclusion minus incorrect exclusion) and A (incorrect post-response. These waveforms were measured relative to exclusion compared to incorrect inclusion) could not be the value at 0 ms (i.e., the time when the spacebar was directly examined for any lag effect. pressed). The ERP data were examined at multiple levels. First, we looked at the ERPs to the incomplete words occurring 2.5. ERP measurements along the two main stimulus-manipulations in the experiment: old versus new, and lag. For these analyses we Several peaks were identified in the ERP waveforms. At collapsed the data across the inclusion and exclusion occipital electrodes, a N200 was defined as the maximum conditions. Second, we addressed C and A by applying the negativity between 180 and 220 ms. The next major PDP subtractions for the inclusion/ exclusion task manipufeature of the ERP waveform was a broad parietal positive lation. As well as performing the main PDP subtraction wave. In the waveforms from the first experiment, two across the tasks (correct on the inclusion minus incorrect successive positive peaks could be identified at electrode on the exclusion), we also performed a similar analysis PO4, the latencies of which fell within ranges of within the exclusion condition (correct response versus ms and ms. The first of these peaks occurred just incorrect response), thereby eliminating the possibility of before a left frontal (F7) negative wave. The amplitude contamination by task differences. Although this is not in values at each of the other electrodes were based on these keeping with the formal PDP analysis, it should also latencies. Since these parietal peaks were difficult to provide a measurement of C since the correct response distinguish in the waveforms from the second experiment, most commonly (although not always) occurs because the we measured the mean amplitude over the two latency subject consciously remembers the original word prior to ranges in both experiments. In addition, slow-wave activity choosing a new word. was analyzed between 800 and 1100 ms at both frontal and We hypothesized that an old/ new effect would be occipital sites. observed as a late parietal positivity, thereby replicating previous research. The index of conscious memory retriev Statistics al processes, obtained by the PDP, was expected to parallel this effect. Automatic processes were hypothesized to Mean results are reported with their respective standard occur prior to conscious memory retrieval and to be deviations. Greenhouse Geisser corrections were incorpo- located over posterior regions of the scalp. Lag was rated into any analysis involving a factor with more than expected to affect the amplitude of conscious memory

6 24 K.A. Kane et al. / Cognitive Brain Research 10 (2000) Table 2 Behavioral measurements from Experiment 1 Study-test lag a Task (% old): Inc (0.06) (0.36) (0.08) (0.11) (0.08) (0.07) (0.08) (0.08) Exc (0.09) (0.09) (0.09) (0.10) (0.10) (0.14) (0.15) (0.16) Process (% strength): C (0.13) (0.16) (0.16) (0.20) (0.13) (0.18) (0.22) (0.23) A (0.38) (0.25) (0.25) (0.28) (0.30) (0.23) (0.19) (0.16) a The upper part of the table presents the proportion responses with an OLD item on inclusion and exclusion tasks. The lower section of the table shows the estimated strengths of the C and A processes. Measurements of standard deviation are presented in brackets. processes in the ERPs in a manner similar to the be- baseline values were not significantly different: 92619% havioral results. That is, the initial lags would result in in the inclusion task and 88610% in the exclusion task. greater parietal positivities than later lags. Finally, based The results for the second experiment were similar to those on the PDP assumption of processing equivalence between of the first. For the inclusion and exclusion tasks, the the inclusion and exclusion tasks, no significant task performance on lag 0 was 9865 and 9764%, respectively. differences were expected until after conscious memory Subjects again responded faster in the inclusion than the retrieval. exclusion task (689 vs ms; t(9)511.76, P,0.001) and the completions of nonstudied items in the inclusion and exclusion tasks were 8763 and 9167%. Subjects in 3. Results the first experiment found the task more difficult than did the subjects in the second experiment. While the subjects in 3.1. Behavioral the first experiment felt that they were guessing on a majority of incomplete items, the subjects of the second In order to determine whether subjects understood the experiment reported that their memories for the words task, performance in the lag 0 condition was considered. were quite strong, even though the subjects were less Although not all subjects exhibited perfect performance in experienced with memory experiments. this condition (probably due to motor errors rather than The data for the PDP calculations are shown in Tables 2 memory errors), subjects did not respond differently with and 3. When considering the effect of lag on OLD item the different instructions in the inclusion task (9567%) responses in the inclusion and exclusion task, there was a and the exclusion task (9669%). Subjects responded faster significant effect of task (F(1,15)5330.7, P,0.001) and a (678 vs ms; t(15)513.4, P,0.001) in the inclusion significant task3lag interaction (F(7,105)515.3, P, task than in the exclusion task, the latter requiring the 0.001). Responses with a studied item significantly degeneration of a new word. Baseline performance for creased between lags 0 and 2 in the inclusion task (t(15)5 responding in the inclusion and exclusion tasks was 3.5, P,0.01) and significantly increased in the exclusion determined by the completions of nonstudied items. The task (t(15)54.3, P,0.001; see Table 2). The results were Table 3 Behavioral measurements from Experiment 2 Study-test lag a Task (% old): Inc 0.98 (0.05) 0.97 (0.05) 0.88 (0.04) 0.90 (0.05) 0.87 (0.06) Exc 0.02 (0.14) 0.03 (0.15) 0.12 (0.16) 0.10 (0.14) 0.13 (0.15) Process (% strength): C 0.95 (0.07) 0.94 (0.09) 0.75 (0.09) 0.74 (0.14) 0.74 (0.18) A 0.49 (0.33) 0.45 (0.22) 0.47 (0.20) 0.39 (0.26) 0.52 (0.23) a The upper part of the table presents the proportion responses with an OLD item on inclusion and exclusion tasks. The lower section of the table shows the estimated strengths of the C and A processes. Measurements of standard deviation are presented in brackets.

7 K.A. Kane et al. / Cognitive Brain Research 10 (2000) similar in the second experiment (Table 3) except that the waveforms. In addition, there was a left frontal negativity main differentiation occurred between interval 1 and 2. peaking around 400 ms. Towards the end of the tracings The results for lags 0 and 1 were not significantly different. there was a sustained frontal negativity recorded maximal- Conscious (C) and automatic (A) contributions to mem- ly over the left frontal scalp. ory were determined for each study-test lag, according to the PDP equations. In the first experiment, C was greater Old/new differences than A across all lags (F(1,15)511.27, P,0.01). In The main difference between studied (old) and nonaddition, there was a significant process3lag interaction studied (new) items was a greater amplitude in the later (F(7,105)517.8, P,0.001) due to a decrease in C from portion of the parietal positive wave (P600) for studied lags 0 to 2 (t(15)54.3, P,0.001) without any significant words (see Fig. 2). The P600 peak measured over elecchange in A (t(15).0.05; see Table 2). The pattern of trodes PO3 and PO4 was significantly larger for the old results did not change when perfect performance was words (F(1,15)58.3, P,0.01). The P600 mean measureremoved to eliminate ceiling effects. That is, the only ments at P3 and P4 showed a main effect of the old/ new significant difference between successive lag periods was variable (F(1,15)512.7, P,0.005). Although the wavein C at study-test lags of 0 and 2 (t(12)53.7, P,0.05). form at this time was significantly larger over the right Similar results were obtained in the second experiment, hemisphere (F(1,15)56.1, P,0.05), there was no signifiwith A showing no significant effect of lag and C showing cant interaction indicating that the effect was asymmetria main effect of lag as the result of a significant decrease cal. Similar effects were also found for the P400 mean between lags 1 and 2 (F(4,32)57.1, P,0.05; t(9)52.5, measurement, indicating that the old new effect lasted P,0.05). C values were similar for lags 0 and 1. Fig. 1 through the duration of the positive wave. Finally, there plots the estimates of A and C obtained in both experi- was a greater frontal sustained negativity for old items than ments. for new items when measured at Fz between 800 and 1100 ms (t(15)52.9, P,0.01). The effects in the second 3.2. Event-related potentials experiment were less clear. Although the effects in the mean waveforms were similar, the only significant statisti- The ERP waveforms for old and new items are shown in cal finding was an electrode by task by old/ new effect Fig. 2 and the topographies of the various peaks for old (F(1,9)55.7, P,0.05) for the early P400 mean measureitems are shown in Fig. 3. A small posterior positive wave ment. peaked at approximately 100 ms and was largest over electrodes Cb1 and Cb2. This was followed by a negative Process dissociations: C wave peaking at about 200 ms that was maximal over the The difference between the ERPs for correct responding posterior temporal scalp and a long-lasting parietal positive in the inclusion task and incorrect responding in the wave peaking at around 550 ms. In many of the subjects exclusion task provides an electrophysiological index of C. this consisted of two peaks (at 400 and 600 ms) and a The main difference was a broad parietal positivity peaksuggestion of this division remains in the grand-mean ing later than the P550 ms wave (see Fig. 4). The difference related to the C wave was clearly larger over the left hemisphere and its onset paralleled only the later portion of the old new effect. As a result of the laterality of C, the difference occurred statistically as a significant electrode by condition effect (F(1,15)56.4, P,0.05). The C effects were similar (although a little smaller) when the statistics were calculated within the exclusion condition. The C effect was also present as a late frontal sustained negativity (t(15)52.4, P,0.05). The results in the second experiment were similar on the grand mean waveforms although noisy because of the small number of errors. Fig. 5 shows the difference waveforms for the two experiments. Fig. 1. Behavioral estimates of C and A. The estimated strengths of conscious and automatic memory processes are plotted against lag for both experiments Process dissociations: A The difference between the waveforms for incorrect responding (with the old word) on the exclusion task and incorrect responding (with a new word) on the inclusion task should contain an index specific to A. This difference waveform showed a sustained right-sided positivity over the posterior scalp (see Fig. 6). Significant electrode by condition effects were found for both P400 (F(1,15)55.9,

8 26 K.A. Kane et al. / Cognitive Brain Research 10 (2000) Fig. 2. Old/ new differences. Grand average waveforms of correct completions to studied (old) and nonstudied (new) test items on the inclusion task. The ERPs are plotted as they were recorded from the scalp with the nose at the top of the figure and the left on the left. The ERPs at the different electrode locations are plotted according to an azimuthal equidistant projection, adapted so that the left-right distance of the middle of the ERP from the vertex (Cz) is 1.3 times an equivalent anterior posterior distance. There was a greater parietal positivity from 400 to 800 ms for old items. Asterisks indicate significant differences. P,0.05) and P600 (F(1,15)55.3, P,0.05) mean mea- measurement at PO3 and PO4 (F(7,105)53.1, P,0.005). surements at P3 and P4 as a result of the A difference The only pairwise difference between lags was observed as being largest over the right hemisphere. This finding was a greater parietal positivity for the lag 0 condition than for not found in the second experiment, which gave opposite all other lags (t(15)52.6, P,0.05). This effect was differences for the sustained potentials, i.e., the posterior observed with inverse polarity, over left frontal sites. When sustained potential was more negative for the incorrect the negative waves peaking near 400 ms were measured at exclusion trials (see Fig. 5). There was also no evidence electrodes F7 and F8, a significant main effect of electrode for an influence of A on the late frontal sustained negativi- (F(1,15)535.0, P,0.001) and an electrode by lag interty. action (F(7,105)52.9, P,0.05) occurred. Again, this difference was due to a greater amplitude in the lag Lag effect condition than in all other lag conditions at F7 (t(15)52.8, Fig. 7 illustrates the main effects of lag and shows the P,0.01). However, the left frontal effects did not show up differences between the lag 0 condition and the average as significant on the mean measurements, although there ERPs from the later lags in the first experiment. There is a was a very significant asymmetry (F(1,15)552.6, P, larger right parietal positive wave peaking at around ). There was no significant main effect of lag ms and lasting for several hundred milliseconds beyond the observed in the frontal sustained negativity between 800 peak. The electrode by task by lag analysis showed a and 1100 ms at Fz. In the second experiment, the effects of significant interaction between electrode and lag lag were less clearly defined. There was an increased (F(7,105)54.1, P,0.005) for the P380 mean measure- posterior positivity for lags 0 and 1 than for later lags but ment at P3 and P4 and a similar effect for the peak this occurred later and had a slightly more anterior scalp-

9 K.A. Kane et al. / Cognitive Brain Research 10 (2000) Fig. 3. Scalp topographies. The different distributions of the ERPs are shown at latencies corresponding to the main peaks in the scalp-recorded waveforms. The continuous lines represent contours at negative voltages and the dashed lines represent contours at positive voltages. The scalp is viewed from above with the nose at the top of the figure. The projection is azimuthal equidistant with the outer edge of the circle showing the scalp down to 208 below the equator of the idealized sphere, i.e., down to about the level of the mastoids. distribution than in the first experiment (showing up understood. Furthermore, because the percentage of remaximally at the CP1 and CP2 electrodes there was a sponses to nonstudied items did not differ between the two significant main effect of lag, F(4,36)52.9, P,0.05). tasks, response criteria were assumed similar. Subjects took longer to respond in the exclusion task than in the Response-locked ERPs inclusion task. This increase in response time is likely due Although there were no significant task differences to the additional requirement of generating a new word in observed prior to 680 ms, the likelihood of contamination the exclusion task. from responses in the inclusion task past this latency (mean The different role of forgetting, as seen by changes in modal reaction time in experiment ms) was too behavioral performance across lag, indicates that two high to compare waveforms. Therefore, average different memory mechanisms are occurring. The first waveforms for studied words were response-locked to the mechanism, with a fast rate of decay, lasts for only a few spacebar response, in order to see what might have been items and likely corresponds to short-term memory prorelated to word generation. This analysis revealed a greater cesses. With a greater number of intervening items, profrontal positivity prior to the response and a greater cesses related to more stable memories occur. Based on posterior negativity for the exclusion task than for the Experiments 1 and 2, the transition between these two inclusion task. However, these waves likely correspond to forms of memory is represented by changes in conscious residual differences in the sensory evoked potentials that memory performance between lags of 1 and 2 intervening occur with different latencies relative to the response. The items, thereby representing more than just immediate ERP differences were no longer apparent when only those repetition. No other changes in performance were seen for responses with similar reaction times in the two tasks up to 32 intervening items or with automatic memory ( ms) were considered. processes of any lag. Toth et al. [70], who investigated the effect of lags 0 8, found a difference of conscious processes between lags 4. Discussion 0 2 and 2 8. Although these results are in accordance with the present results, another study investigating the 4.1. Behavioral results effect of lag, using lags of 0, 3 and 12 on a yes/ no recognition task, found no difference between conscious The absence of any difference between the two tasks on processes at lags 0 and 3, but did find a difference between the percentage of correct responses in the lag 0 condition lags 0 and 12 [30]. The less taxing task demands of the indicates that the instructions for both tasks were equally yes/ no recognition task may have allowed for more items

10 28 K.A. Kane et al. / Cognitive Brain Research 10 (2000) Fig. 4. Conscious memory processes. An ERP index of C is represented by the difference obtained by subtracting ERPs on trials in which subjects incorrectly responded with an OLD item in the exclusion task (A) from the ERPs on trials in which they correctly responded with an OLD item in the inclusion task (C 1 A). Asterisks indicate significant differences. to remain in a short-term buffer than in the word comple- performance of young adults thereafter. Jennings and tion tasks used in both the present study and that of Toth et Jacoby [30] looked at lags of 4, 12, 24 and 48 and found al. [70]. Although not looking at conscious and automatic no significant effects of lag for either conscious or memory processes separately, Nielsen-Bohlman and automatic processing. Although this study did not involve Knight [46] used a recognition task to investigate lags of 0, a condition of immediate repetition, the results agree with 2 6 and They found significant differences be- those found in the present study in that no study-test lag tween lags 0 and 11 79, but not between lags 0 and 2 6. effects are observed between 2 and 32 intervening items. Again, these results may be explained by the idea that It is not clear what determines the point at which the more items can be retained within a short-term buffer due short-term memory is superseded by another longer-term to simple task demands. In one study, performance in a lag memory. In our experiments, there is no way to disen- 0 condition did differ from lags 2 6 in a recognition task tangle time from the number of items held in short-term [12]. However, because lags 2 6 were not considered memory. In the absence of rehearsal, both time and separately, it is not certain whether any particular lag is number of items probably play a role in the memory responsible for this difference. In a yes/ no recognition transition in that items not only decay from short-term task, processes such as memory search are not required, memory with time but also, as a result of a capacity thereby allowing more items to be retained and/ or at- limitation, can be bumped from the store by new incoming tended to in a short-term buffer. information. Regardless of the exact point of transition from a short- The behavioral results therefore clearly demonstrate that term buffer, lag appears to have little effect on the automatic memory processes are much less affected by lag

11 K.A. Kane et al. / Cognitive Brain Research 10 (2000) Fig. 5. Conscious and automatic memory processes. This figure shows the difference waveforms representing C and A processes for the two experiments. Within each half of the figure, the left waveforms are from the left scalp (FC1 and P3) and the right waveforms from the right (FC2 and P4). C is associated with a positive wave (P600) maximally recorded from the left parietal region and a negative slow wave (NSW) from the left frontocentral regions. A showed a right parietal positivity in the first experiment but this did not replicate in the second. those when the subject responded correctly on the exclu- sion task) do not participate in the calculations. This is of no concern when assessing response probabilities, since the probability of the unused responses in part determines the probability of those used in the calculations. However, discarded trials are problematic when recording ERPs where better signal-to-noise ratios occur when more re- sponses are averaged together. Second, subtracting one ERP from another also increases the level of noise in the than conscious memory processes. Furthermore, the effects of lag on conscious memory processes indicate two different storage modes, one lasting either for a few seconds or for a few items and the other lasting longer. Can the ERP results suggest physiological bases for the difference between the automatic and conscious retrieval and for the two types of storage accessed by conscious retrieval? 4.2. Event-related potentials recording. If difference-waveforms are calculated between two equally noisy recordings, the noise level in the Some caveats resultant waveform is increased by a factor of 1.41 over Applying PDP to studying the ERPs is beguiling in theory but difficult in practice. The hope is that separate ERP components might be dissected out by the algebra of the PDP. The main problem is that some of the conditions may not provide sufficient trials to allow for reliable ERP measurements. This was especially true for the incorrect trials in the second experiment. Two further problems affect the noisiness of the ERP recordings. First, PDP derives its measurements of A and C from only selected responses. For example, C is based on the difference between correct responses on the inclusion task and incorrect responses on the exclusion task. Many trials (e.g., those used in the subtraction. If one of the signals is much noisier than the other (e.g., if it was based on fewer trials), the noise in the difference waveform approaches the noise level of the noisier signal (although it still remains larger). This is very problematic for experiments using PDP since crucial responses (e.g., incorrect responses on the inclusion task in the second experiment) occur infrequently. The incidence of these trials cannot be significantly increased without altering the strategies used by the subjects for memory and attention. For example, subjects level of attention to accuracy could be decreased, but this would make it impossible to know whether incorrect responses

12 30 K.A. Kane et al. / Cognitive Brain Research 10 (2000) Fig. 6. Automatic memory processes. An ERP index of A is represented by the difference between grand average waveforms of incorrect trials on the inclusion (neither C nor A) and exclusion task (A), collapsed across lags. Asterisks indicate significant differences. were related to lapses in remembering the words or in monitoring the responses. In addition, particular characteristics of the subjects themselves can decrease the signal-to-noise ratio, especially when few trials are included in the overall average. Differences in the subjects for the two experiments may account for the smaller waveforms and greater amount of noise observed in the recordings of the second experiment. For instance, the subjects in the first experiment were older and more highly educated, were known to the experimenter and were present as volunteers. On the other hand, the subjects in the second experiment were not known to the experimenter and received payment for their participation, regardless of their performance. In addition, the subjects in the second experiment were less experienced as ERP subjects. Therefore, the subjects in the second experiment may have been less motivated in having clean recording (i.e., by sitting still, taking breaks when needed and restricting eye movements to specified times) and may have been less invested in the task since they were unknown to the experimenter. Trial-to-trial latency variability of the waves in the ERP may have affected our recordings. Clearly, there may have been some variability in the timing of the P600 wave in relation to the certainty with which the subject was consciously aware that the stimulus had occurred before. This would have been close to the trial-to-trial variability in the reaction time. There could also have been some variability in the latency of the earlier P380 wave. The fact that the P380 wave was less prominent in the second experiment suggests that it might have occurred later with the different experimental context, for example, when it was worthwhile to check short-term memory more closely. There are several techniques available to compensate for latency variability. Most are based on a simple waveform and usually involve a single peak. Most require that the waveform be detectable, at least to some extent, in the single-trial recording. Neither of these criteria was met in