Implicit Sequence Learning: Effects of Level of Structure, Adult Age, and Extended Practice

Participants

There were 48 participants, 24 young and 24 old, each of whom completed 10 1- to 1.5-hr sessions.¹ Half of those of each age were randomly assigned to each lag level. The young people were recruited from Georgetown University and the older participants from an advertisement in the Washington Post Health Section. No participant had been in similar studies, and all were paid for participating. Their characteristics are summarized in Table 1. The older participants were highly educated, reported high levels of health, and did not differ from the young on the Mini-Mental State Examination (Folstein, Folstein, & McHugh, 1975). Nonetheless, as is typical, the older participants performed more poorly on neuropsychological tests assessing memory and speed taken from the Wechsler Adult Intelligence Scale–III (Wechsler, 1997a) and the Wechsler Memory Scale–III (Wechsler, 1997b).

Design

The design was a 2 (age) × 2 (structure) × 2 (trial type) × 10 (session) mixed factorial, with age (young vs. old) and structure (lag-2 vs. lag-3) varying between subjects and trial type (Pattern vs. Random) and session (1–10) varying within subjects.

Procedure

Participants first completed a consent form, which had been approved by the Georgetown University Institutional Review Board, and a brief biographical questionnaire. They then completed 21 blocks of the ASRT task. At the end of the final ASRT block, they completed four blocks of the free-generation (described later). Finally, they completed a written end-of-session questionnaire containing general questions about their strategies but not referring to any regularity. Participants returned for nine additional sessions. There were no fewer than 45 min and no more than 2 days between sessions and no more than two sessions per day (except for 2 participants who had to complete three sessions in 1 day because of scheduling problems). After the free-generation task on Sessions 2 through 9, people completed one or more of the supplemental tasks listed in Table 1. Then, on the 10th and final session, after the free-generation task, participants were given a postexperimental interview described in J. H. Howard and Howard (1997, p. 637).

Do People Reveal Declarative Knowledge of the Regularity?

Questionnaire responses were similar across ages and pattern structure and were consistent with those from earlier ASRT studies. When asked whether they noticed “any regularity in the way the stimulus was moving on the screen,” 58% answered “yes,” and almost everyone reported that they had searched for a pattern at some time during the experiment. Nonetheless, no one believed they had found the pattern; some people concluded that none was there, and others concluded that there must have been a pattern that was too subtle for them to pick up.

When at the end of the last session we told them that there was a regularity and asked them to describe anything about it that they could, most offered vague hypotheses with little confidence. The two most accurate descriptions came from people who received the lag-2 pattern 1r2r4r3r. The better of these was from a 73-year-old, who reported that “3– 4 –1 is often, but can’t predict it.” This report is correct in that such triplets (and, in fact, any 3-x-1 sequence) are among the more frequent, structure-consistent triplets. The next most accurate report was from an 18-year-old, who said the targets “moved from left to right; each finger struck the same amount of keys.” This is substantially correct in that all the targets do occur equally often, and this sequence might be construed as moving generally from left to right, although, of course, random elements intervene and the 4r3 part of the sequence violates this regularity.

Otherwise, the guesses and descriptions offered were incorrect. The most frequent were that some positions occurred more often than others, that positions on one side occurred more often than those on the other, that certain pairs of items repeated more frequently, and that doubles or triples of the same items were common.

Although people were unable to describe the regularity, many did seem aware that they were learning something, perhaps unconsciously, and many reported that this learning involved their fingers or muscles and that it led to errors. For example, 33%, distributed approximately equally across the four groups, referred to “automatic” learning and responding, “fingers” or “muscles” taking over, “unconscious” learning, or “errors” caused by such learning. For example, a lag-2 20-year-old reported, “My fingers moved automatically. I thought I knew what was going to come next. I thought my mind was subconsciously telling my fingers where to move.” A lag-3 22-year-old reported, “Sometimes I got in a rhythm and it seemed that I was not consciously aware, but that it was automatic.” According to a lag-2 66-year-old, “I couldn’t do the patterns, but I think my brain is recognizing these sequences, the part I’m not conscious of. Like riding a bicycle. Sometimes I seemed to know which direction to go.” A lag-2 80-year-old said, “It threw me off because I would want to hit a certain key, but it tricked me a lot. It made me make mistakes.” A lag-2 67-year-old commented on the “muscle memory” he was using, comparing it with how dancers move habitually because of practice.

Thus, people revealed virtually no ability to describe the regularities to which they had been exposed for 10 sessions for a total of 2,100 repetitions. Nonetheless, they did have some insight; the majority believed that some regularity was present and that it was influencing their performance in ways they could not control. We conclude, therefore, that the learning reported later is relatively pure implicit learning in that the participants revealed no declarative knowledge of the structures.

Data Reduction

The data from the 10 random trials that began each block were discarded. For the remaining trials (80 for lag-2 and 120 for lag-3), the median RT for correct trials was calculated separately for random and pattern trials for each block. Then the means of these medians were calculated across the 21 blocks in each session to yield the mean RT for each trial type for each person. Similar procedures were used for accuracy. An alpha level of .05 was used throughout, and significance tests were always two-tailed. Effect sizes reported are always for the age difference in the overall learning score, usually the trial-type effect, across all sessions combined.

Initial ASRT Analyses: Overall Accuracy and RTs

Figures 1 and 2, respectively, show RTs and accuracy across sessions for all four groups for pattern and random trials. The two most prominent features in Figure 1 are that RTs decrease over sessions and that older people respond more slowly than younger. The main effects of session, F(9, 396) = 198.08, MSE = 842.45, and age, F(1, 44) = 182.43, MSE = 32,144.87, were both significant. This change across sessions, indicating overall skill learning, is greater for the older than the younger groups, reflected in a significant Age × Session interaction, F(9, 396) = 21.27, MSE = 842.45.

More important for assessing pattern learning (although hard to see in Figure 1), random trials are slower overall than pattern trials, and this trial-type effect increases across sessions; the main effect of trial type, F(1, 44) = 200.52, MSE = 102.20, and the Trial Type × Session interaction, F(9, 396) = 259.20, MSE = 12.49, are both significant. Most important, the magnitude of this trial-type effect varies with age and with structure. For example, the Trial Type × Age × Structure, F(1, 44) = 7.14, MSE = 102.20, Trial Type × Session × Structure, F(9, 396) = 13.12, MSE = 12.49, and Trial Type × Session × Age, F (9, 396) = 2.47, MSE = 12.49, interactions are all significant. The nature of these interactions is examined in more detail later.

For accuracy in Figure 2, it is notable that for all groups accuracy is high and constant across sessions for the pattern trials but declines for the random trials. The main effects of session, F(9, 396) = 11.54, MSE =.0004, and trial type, F(1, 44) = 288.45, MSE =.001, and the Session × Trial Type interaction, F(9, 396) = 29.97, MSE = .0001, are all significant. Probabilistic regularities, unlike deterministic ones, usually yield such increasing errors with practice (Curran, 1997a; Feeney et al., 2002; D. V. ; J. H. Howard & ; Schvaneveldt & Gomez, 1998). Participants often report that their fingers seem to take over and lead them to make more “oops” errors. Of course, unbeknownst to the participants, these errors are occurring predominantly on random trials (even though these are slower than pattern trials) and hence reflect learning of the regularity. Thus, when probabilistic sequences are used, errors become as sensitive to learning as are RTs. Figure 2 shows that older adults are more accurate than younger people, but this age difference is primarily due to the random trials. The main effect of age, F(1, 44) = 68.38, MSE = .01, and the Age × Trial Type interaction, F(1, 44) = 88.91, MSE = .001, are significant. Finally, as with RTs, the magnitude of the accuracy trial type effect varies with both age and structure; the four-way interaction of Age × Structure × Trial Type × Session is significant, F(9, 396) = 2.83, MSE =.0001.

We now examine the nature of these interactions by focusing on the questions the current study aimed to answer. In particular, we consider first whether there are age differences in learning and whether both young and older people can learn lag-2 and lag-3 structure. We then ask at what point in training it is possible to detect pattern learning and age deficits therein. We next consider whether the types of errors people make are knowledge driven and equally so for both ages and examine exactly what is being learned. Finally, we ask whether learning is revealed in the conceptually driven free-generation task.

Are There Age Differences in Learning Lag-2 and Lag-3 Structure?

Figures 3 and 4 show the trial-type effect for RTs (RT on random trials minus that on pattern) and accuracy (accuracy on pattern trials minus that on random), respectively. In both cases, higher difference scores indicate greater learning, and the age differences in learning can be seen clearly.

The data from the lag-2 patterns replicate our earlier findings of age-related deficits (e.g., J. H. Howard & ) and indicate that the deficit remains even after 10 training sessions. The lag-2 groups revealed significant Age × Trial Type × Session interactions, F(9, 198) = 2.21, MSE = 15.06, and F(9, 198) = 7.57, MSE = .0001, for RT and accuracy, respectively. For the age difference in learning, r_effect = .50 for RT and .86 for accuracy.

There are also age-related deficits in learning lag-3 patterns, although they are smaller than for lag-2, likely because the lag-3 trial type effects are small even for the young group. When the lag-3 data are analyzed alone, the Trial Type × Age, F(1, 22) = 37.00, MSE = .0001, and Age × Trial Type × Session, F(1, 22) = 2.71, MSE = .0001, interactions are both significant for accuracy, although neither is significant for RT. For the age difference in learning, r_effect = .79 for accuracy and .12 for RT.

Because there was an overall age difference in accuracy, we also compared the groups using a learning measure that normalizes for overall accuracy. For each block, we divided the number of errors each person made on random trials by that person’s total number of errors. This measure revealed the same group differences as the nonnormalized accuracy data in Figure 4. For these proportional trial type effects, there are significant interactions of Age × Session, F(9, 396) = 2.78, MSE = .002, and Age × Structure, F(1,44) = 12.10, MSE = .009. For the age difference in learning, r_effect = .80 for lag-2, and .35 for lag-3.

Thus, we conclude that there are age differences in sensitivity to repeating regularities in the environment for both lag-2 and lag-3 structures and that these age effects can be detected by accuracy or RT trial-type effects (for pattern vs. random trials) and by normalized proportions of errors occurring on random trials.

Can Participants of Both Ages Learn Lag-2 and Lag-3 Structure?

To find out, we performed Trial Type × Session ANOVAs for each of the four groups separately. In keeping with the unambiguous evidence of pattern learning for the young and old lag-2 groups and for the young lag-3 group in Figures 3 and 4, there are significant Trial Type × Session interactions for all three of these groups for both accuracy and RT. The least significant of these comparisons was that for the young lag-3 group’s RT data, in which F(9, 99) = 2.90, MSE = 4.50. Whether older people can learn lag-3 patterns is less clear. Although Figures 3 and 4 show this group’s trial-type effect hovering slightly above zero throughout training for both measures, there is no noticeable increase beyond Session 1. In keeping with this observation, the older lag-3 group’s Trial Type × Session interaction was not significant for either measure ( p > .80), even though they do reveal a significant trial-type effect for both RT, F(1, 11) = 10.23, MSE = 53.38, and accuracy, F(1, 11) = 12.30, MSE = .0001. The quadruplet analyses described in a later section help to clarify why this small, constant trial-type effect is occurring for the older lag-3 group.

So far, then, we conclude that young people can learn both lag-2 and lag-3 structures, even when there is no lower level structure in the sequence. However, more exposure is needed for lag-3 than for lag-2 structure, and lag-3 learning is revealed primarily by accuracy. This finding that young people are able to learn lag-3 regularities is consistent with the work of Cleeremans and McClelland (1991) and Remillard and Clark (2001), but the sequences in these earlier studies had also contained lower level structure. Thus, the current finding demonstrates that young people can learn lag-3 structure even when no lower level structure is present. We find that older people are able to learn lag-2 regularities, but they show little, if any, learning of lag-3 structure when that is the lowest level of structure present.

How Early in Training Can Pattern Learning and Age Deficits Be Detected?

We examined the onset of learning by dividing each of the first two sessions into four epochs, with the first three epochs of each session containing five blocks (i.e., 50 repetitions of the pattern) and the last containing six blocks. The mean RT and accuracy for these first eight epochs for both age groups are shown in Figure 5 for the lag-2 structure and in Figure 6 for the lag-3 structure.

Figure 5 suggests, and ANOVA confirms, that lag-2 RTs show significant evidence of pattern learning but not an age difference therein; the Epoch × Trial Type interaction is significant, F(7, 154) = 5.63, MSE = 52.92, but no Trial Type × Age interactions approach significance. For the age effect on learning, r_effect = .10. The t tests indicate that both age groups first show a significant RT trial type effect on Epoch 5, t(11) = 2.80 and 5.48, for young and old, respectively.

The accuracy data for the lag-2 groups in Figure 5 indicate that both ages show learning, but the young do so earlier. There were interactions of Epoch × Trial Type, F(7, 154) = 7.49, MSE = 4.59E-4, and Age × Trial Type, F(1,22) = 32.14, MSE = .001, r_effect = .77. The young group first showed a significant trial type effect on Epoch 1, t(11) = 2.94, whereas the old group first did so on Epoch 4, t(11) = 2.37. Thus, both ages showed learning on accuracy before RTs. On accuracy, the young group revealed significant learning within the first 50 repetitions of the pattern, whereas the older group did not do so until between 150 and 210 repetitions. The age groups differ significantly from each other on this accuracy trial type effect as early as the second epoch, t(22) = 2.80 (i.e., within the first 100 pattern repetitions).

The lag-3 data in Figure 6 show smaller and more variable learning effects. RTs reveal no significant effects of trial type at all during the first two sessions. Accuracy does yield significant effects of trial type, F(1, 22) = 30.66, MSE = 2.69E–4, and Trial Type × Age, F(1, 22) = 7.92, MSE = 2.69E–4. For the age difference in learning, r_effect = .23 for RT and .50 for accuracy. When examined alone, the young group reveals a significant accuracy trial-type effect, F(1, 11) = 40.80, MSE = 2.31E–4, but the older group does not. Thus, there is evidence of lag-3 learning within the first two sessions for the young people but not the older people.

In summary, analyses of the first two sessions indicate that, on accuracy, age deficits in lag-2 learning are detectable within the first 100 repetitions of the pattern.

Are the Types of Errors Knowledge Driven and Equally So for Both Ages?

The accuracy analyses so far have focused on whether people make errors, but here we ask what kind of errors people make. One way of accounting for overall performance on this task is to assume that as people gain experience with the sequences, they build up a model of the regularity. This model primes items that are likely in context. As this model grows more accurate, it primes correct responses on pattern trials, but it will often prime incorrect responses on random trials, thus leading accuracy and RT on random trials to diverge from pattern trials. This interpretation suggests that when people do make errors on random trials, those errors should be structure consistent and increasingly so as they learn the regularity.

We examined all the random trials on which each person made an error and classified them as either structure consistent or structure inconsistent. For lag-2 patterns, a structure-consistent error results in a triplet that could occur on pattern trials, whereas a structure-inconsistent error does not. For example, if a participant whose pattern was 1r2r3r4r encountered the sequence 144, then incorrectly responding to the last item in the triplet with a 2 would be considered a structure-consistent error, whereas incorrectly responding with a 3 would be considered a structure-inconsistent error. For each person, we determined the percentage of all errors on random trials that were structure consistent. We did a similar count for the lag-3 pattern using quadruplets. The mean proportions for all four groups are shown in Figure 7.

Figure 7 suggests, and ANOVA confirms, that a greater proportion of the errors are structure consistent for the lag-2 than for the lag-3 patterns, F(1, 44) = 126.49, MSE = .01, and for the younger than older people, F(1, 44) = 108.73, MSE = .01. For the age difference in learning, r_effect = .89 for lag-2 and .75 for lag-3. In addition, the proportion of structure-consistent errors increases significantly over sessions for the young people for both lag-2 and lag-3 patterns but does not increase for the older groups. This Age × Session interaction is significant, F(9, 396) = 2.42, MSE = .003. Subsequent one-way ANOVAs for each of the four groups revealed significant session effects for the young lag-2, F(9, 99) = 2.23, MSE = .004, and young lag-3 groups, F(9, 99) = 3.22, MSE = .001, but not for the two older groups ( p > .15).

The young people were different from chance (i.e., .25)² on every session in both the lag-2 and lag-3 groups, ts(11) = 2.79–13.17. For the older lag-2 group, this proportion was different from chance ( p > .05) on all sessions except Sessions 3 and 7, but the older lag-3 group differed from chance only on Sessions 5 and 8.

These findings indicate a notable difference in the kinds of errors, with the younger people making knowledge-driven errors for both lag-2 and lag-3 patterns and increasingly so as they gain experience. The older people make fewer knowledge-driven errors, and their proportion does not increase with experience.

What Is Being Learned and Are There Age Differences Therein?

As experimenters, we know that there is an alternating regularity that makes certain trials completely predictable, but people who show trial-type effects are not necessarily learning this. In fact, no one reported awareness of such a structure.

Is Learning Revealed in the Free-Generation Task?

People performed the free-generation task at the end of each session. To find out whether the sequences people produced showed the influence of the regularity, we divided each lag-2 person’s output into a series of triplets by stepping a three-item window through the responses, one response at a time. We then sorted each of these triplets into one of four categories, corresponding to those in the triplet analyses just discussed. Repetitions (R) were repetitions of the same three positions (e.g., 333). Spans (S) were triplets starting and ending with the same position (e.g., 313) that were not repetitions. Inconsistent (I) triplets were inconsistent with the person’s pattern (but were neither R nor S triplets). Consistent (C) triplets were those that were consistent with the person’s pattern. We normalized the total count for each triplet type by the number of unique triplets, thus obtaining the mean generation rate per triplet for each category. We analyzed the lag-3 data in the same way, except using quadruplets instead of triplets.

The ANOVAs reported later are based on all 10 sessions. However, these data were variable and changed little over sessions, so Figure 10 only shows the mean generation rates for Sessions 1 and 10. For completeness, all four triplet (quadruplet) types are shown, but the critical comparison is that between the C and I triplets, which are counterbalanced across participants, in that a given triplet is consistent for some participants and inconsistent for others. So, unlike R and S types (which might reflect preexisting tendencies), if C and I types yield different generation rates, this must be due to the specific experience people had with their assigned regularity. Therefore, it is noteworthy that C and I triplets (and quadruplets) in Figure 10 do not differ from each other for either age group or lag. ANOVAs revealed no significant main effects or interactions of triplet or quadruplet type (C vs. I). Thus, when the uninterpretable R and S types are eliminated from consideration, no group produced structure-consistent sequences during the free-generation task, and thus we find no evidence that people can use their pattern knowledge in the absence of stimuli in this conceptually driven task.

Age Deficits in Learning or Only in Performance?

There are several indications that the trial-type deficits observed reflect age deficits in learning and are not just performance differences caused by the older groups’ lower overall error rates. First, in J. H. Howard and , we matched a young and an old group on overall accuracy in the ASRT task. When we did so, we still found significant age deficits in trial-type effects. Second, in the current study, we found that normalized error scores, consisting of the proportion of a person’s total errors that occurred on random trials, also revealed age deficits. Third, in the current study, when we examined the kinds of errors that people made, we found that older people’s errors were less likely than young people’s errors to be structure consistent. Thus, we conclude that older people have learned less about the regularity than younger.

Are Age Deficits Due to Implicit or Explicit Learning?

There is continuing debate as to how to differentiate empirically between implicit and explicit learning and on how to assess awareness and the presence of declarative knowledge (Shanks & St. John, 1994). It is agreed that tasks are rarely process pure. This is important for age comparisons because age deficits in explicit learning are well established (Prull et al., 2000). Thus, it might be argued that the age deficits observed here reflect age deficits in explicit learning.

Evidence suggests, however, that learning in the ASRT task is largely implicit. First, people’s verbal reports indicate that they are uniformly unable to describe the regularity, even though, if they did discover it, the alternating regularity should be easy to state in words. In addition, it is not that people are unwilling to guess. Even though their verbal reports rarely describe anything accurate about the pattern (i.e., they rarely name any triplets that were particularly frequent), they often do describe regularities that are incorrect (i.e., that repetitions of two or three items in a row are frequent).

Second, the free-generation data in the current experiment indicate that when instructed to do so, people do not generate sequences that contain details of the regularity they had encountered, even though their RT and accuracy during the ASRT task itself indicated that they were sensitive to these regularities. For example, the triplet analyses revealed that lag-2 groups of both ages responded more quickly and accurately to structure-consistent (P) events than to structure-inconsistent (RI) events during the ASRT task. Nonetheless, during the free-generation task, people were no more likely to produce pattern-consistent than pattern-inconsistent triplets. This same dissociation occurred for the lag-3 young people; their performance distinguished between the P and I quadruplets during the ASRT, but not the free-generation task.

Third, in other ASRT studies, we have included sensitive recognition and sorting tests that failed to reveal any ability to discriminate between structure-consistent and structure-inconsistent sequences (Japikse, Howard, & Howard, 2001; Negash, Howard, Japikse, & Howard, 2003). We conclude that young people learn lag-2 and lag-3 structure implicitly and that the age deficits observed are not due to explicit learning deficits.

Implications for Theory

The current findings suggest that, without awareness, people build up a mental model of the stochastic structure of the sequence (in this case, the frequencies with which triplets or quadruplets occur), perhaps by connectionist principles (Cleeremans & Jimenez, 1998). Early on in experience with a given task or environment, responding is based primarily on data-driven processes: People simply respond to the events they encounter. However, with experience, their mental model becomes richer, and knowledge-driven processes play an increasingly important role in responding. Although this enables more efficient processing of predictable events, it can also lead to increasing errors for unexpected events, which occur in probabilistic sequences such as those used here (and often in daily life; Schvaneveldt & Gomez, 1998). Thus, the smaller trial-type effects and fewer knowledge-based errors we observe for older people suggest that healthy aging is accompanied by an impaired ability to acquire a mental model of sequential structure.

What cognitive mechanisms underlie this deficit? We have appealed to Salthouse’s simultaneity mechanism of cognitive aging (Salthouse, 1996); because of cognitive slowing, older people have less information available simultaneously than do younger people and so they develop an impoverished mental model (D. V. ; J. H. Howard & ). This view attributes the age impairment to a capacity limitation: Old people are unable to keep enough of the recent past activated simultaneously to learn how it relates to present and future events.

That older people are poorer than young people at learning lag-2 structure and show little learning of lag-3 structure is consistent with this interpretation. However, it is possible that aging brings not only a reduction in capacity but also a reduced ability to form associations among activated items. This latter binding-deficit view builds on the finding that elderly people have difficulty associating two or more items or features in memory (Mitchell, Johnson, Raye, Mather, & D’Esposito, 2000; Naveh-Benjamin, 2000). Age deficits in sequence learning may reflect a more general binding impairment because sequence learning entails binding information across time.

The binding hypothesis goes beyond simultaneity in assuming that, even when older people activate two items simultaneously (which is necessary to learn lag-1 structure), they are less able than young to form an association between them. Of course, most evidence for such binding deficits comes from explicit measures of learning and memory. Whether there are age deficits in learning new associations on implicit tests of memory is still unclear (e.g., Light, Prull, LaVoie, & Healy, 2000). The fact that classical eyeblink conditioning shows age-related deficits (e.g., Woodruff-Pak & Jaeger, 1998) suggests that at least one kind of implicit learning is age sensitive, even though it requires binding only two events: the conditioned stimulus and the unconditioned stimulus.

The fact that most early studies of SRT learning revealed age constancy does not necessarily argue against a binding deficit. To our knowledge, all of the studies that have revealed age constancy in SRT learning used structures that contained some 0th-order regularity, that is, individual events differed in frequency (e.g., Cherry & Stadler, 1995; Frensch & Miner, 1994; D. V. ; Salthouse et al., 1999). Hence, the lack of age deficits could have been due to age equivalence in learning the frequency of occurrence of individual events (which presumably does not require binding) rather than in learning first- (lag-1) and higher order structure.

Both the simultaneity and the binding explanations are consistent with theories that attribute many aspects of cognitive aging to a general age-related deficit in neuromodulation (Braver et al., 2001; Li, 2002). At the biological level, age impairments are thought to reflect changes in the dopamine system as manifest in impaired dorsolateral prefrontal cortical function. These neuromodulation deficits result in healthy elderly suffering “cognitive control impairments resulting from a failure to properly represent and maintain context information” (Braver et al., 2001). One characteristic of neural networks that simulate such neuromodulation deficits is that such “aged” networks act like the elderly participants in the current study in revealing deficits in asymptotic learning even after extended practice (Li, 2002).

The neuromodulation theory can potentially integrate the simultaneity and binding accounts into a single conceptual framework that ties the behavioral findings to their neural mechanisms. In this sense, the neuromodulation framework, like the simultaneity and binding views, points to frontal lobe impairments as being central to cognitive aging. Therefore, it is important that imaging and patient studies indicate that the prefrontal cortex is involved in learning higher order sequence structure (Gomez Beldarrain, Grafman, Ruiz De Velasco, Pascual-Leone, & Garcia-Monco, 2002; Sakai et al., 1998). Of course, this does not exclude the possibility that other structures (e.g., medial temporal lobe) also contribute to age deficits; there is some evidence from amnesia patients and imaging studies for medial temporal lobe involvement in the learning of higher order sequences (Curran, 1997b; Schendan, Searl, Melrose, & Stern, 2003). Although the age deficits in the current experiment are consistent with the neuromodulation/context-deficit perspective, additional work is required to distinguish or integrate the simultaneity, binding, and neuromodulation-deficit views of age-related sequence learning impairment.

Practical Implications

In the case of implicit learning of sequences, the aged person’s glass is both half full and half empty. From the half-full perspective, the ability to learn lag-2 structure implicitly continues at least into the decade of the 80s, even though older people are often not aware that they are learning. This ability enables people to remain sensitive to subtle regularities in their environment and to thus adapt to a changing world. Such implicit learning is spared compared with more strategic kinds of learning. For example, we have found that when given intentional instructions to discover and describe the pattern in lag-2 ASRT sequences, some young people were able to do so, but no elderly people could, even though both groups showed significant implicit learning as revealed by trial-type effects (D. V. ).

On the half-empty side, higher order sequential structure is thought to be important for learning language (e.g., Conway & Christiansen, 2001) and skills such as using a software package. Thus, the age-related deficits in learning higher order sequential structure observed here probably make it more difficult for older people to acquire new skills and less likely that they will become sensitive to subtle regularities in their environment, even after extended exposure.

Establishing what levels and kinds of implicit learning are impaired and what factors affect such learning should be useful for designing environments and educational programs that take advantage of the older person’s spared capacities and compensate for those that have declined.