Aging and Recollection in the Accuracy of Judgments of Learning

Participants

Seventy-six young adults (42 male; mean age = 19.6) and 78 older adults (25 male; mean age = 71.4) participated¹. Young adults were undergraduates at the Georgia Institute of Technology and participated for course credit. Older adults were from the Atlanta community and were paid $10 per hour. Participants from the two age groups were randomly assigned to the recognition and cued-recall test conditions.

Materials

Stimuli consisted of 120 five-letter nouns and adjectives of moderate frequency (e.g., glass, clown) that were divided into three sets of 40. Each participant studied words from two of the sets with the remaining set acting as unstudied items during test. Word sets were counterbalanced across participants such that each word served equally often as studied and unstudied. The recognition and cued-recall tests used the same words and word sets. For cued recall, word-stem cues were created by replacing the final two letters from each word with underlines. Each stem could be completed by at least two words but only one word from the stimulus set. Testing was done on Windows-based PCs with 11 keys across the top of the keyboard (“~” to “0”) re-labeled “0” to “100” in intervals of 10.

Procedure

At study, participants were told they would see a list of words that they should remember for a later memory test, and that they “may want to associate each word with something that is personally meaningful” or “generate a mental image.” They were also asked to rate how likely they were to remember each word on the later memory test. Words were presented one at a time in the center of the screen with the JOL scale (“0 … 100”) appearing at the bottom of the screen 3 seconds after the word was presented. Participants were told a rating of 0 meant they were “absolutely certain” they would not remember the word, 100 meant they were “absolutely certain” they would remember the word, and 10 to 90 indicated intermediate levels of certainty. Participants were encouraged to use the entire scale. Words remained visible until the JOL rating was entered. One second after the rating, the next word was presented.

Following study, participants completed two distractor tasks, resulting in a study-test delay of 10 to 12 minutes. For recognition, participants saw words one at a time in the center of the screen and were asked to rate the quality of their memory for each using “Recollect”, “Familiar”, or “No Memory”. Definitions of these responses were similar to those used in Remember/Know studies (e.g., Gardiner, 1988). Briefly, participants were told to respond Recollect when they could clearly remember specific details associated with studying a word; Familiar when they felt the word was from the study list but could not remember specific details; and No Memory when the word neither felt familiar, nor could they remember any details about its earlier presentation. The response options were presented at the bottom of the screen [“Recollect (R), Familiar (F), No Memory (N)”] and participants responded using the “R”, “F”, and “N” keys.

For cued recall, participants were presented with word-stems and told to complete each with a word from the study list. If they could not remember a study word, they were to complete the stem with the first word that came to mind. Immediately after a word was entered, the R/F/N response options, defined as in the recognition condition, appeared at the bottom of the screen.

Memory Performance

Table 1 contains the proportion of R, F, and N responses given by young and older adults on the recognition test. Table 2 contains the corresponding data for cued recall, presented as both Observed and Rescaled Proportions. For the Observed Proportions (top panel), the R, F, and N responses are expressed as a proportion of the total number of target words (80 studied, 40 unstudied); thus, the observed R, F, and N proportions sum to the total proportion of stems completed with target words. For the Rescaled Proportions (bottom panel), the R, F, and N responses are expressed as a proportion of the number of test stems completed with target words; thus, the adjusted R, F, and N proportions sum to 1.0, as in recognition. Presenting the cued-recall data as adjusted proportions allows a better comparison across conditions (tests or age groups) that differ in the total number of target words for which a participant can make an R/F/N judgment (see Hamilton & Rajaram, 2003; Toth, 2005).

For recognition, older adults were less likely than the young to correctly classify studied words as old (R+F: .83 vs. .91; t(75) = 2.85, p < .01) but more likely to incorrectly classify unstudied words as old (.22 vs. .14; t(75) = 2.95, p < .01). There were no age differences in R and F responses to studied items, but older adults were more likely than the young to make false R responses to unstudied items (.08 vs. .02; t(75) = 4.95, p < .01); the age difference in F responses to new items was not reliable. Overall then, there was an age-related decrease in recognition accuracy, driven primarily by an age-related increase in false recollection.

For cued recall, target output was higher for young, compared to older, adults for stems corresponding to studied (.60 vs. .46; t(75) = 5.78, p < .001) and unstudied items (.32 vs. .27; t(75) = 2.35, p < .05). Given these differences, all subsequent analyses were performed on the rescaled proportions. The rescaled data showed older adults to be less likely than the young to correctly classify recalled studied words as old (R+K: .95 vs. .90; t(75) = 2.50, p < .05), a difference driven mainly by older adults making fewer correct R responses (.68 vs. .82; t(75) = 3.76, p < .01). And, as in recognition, older adults also made more incorrect R responses to unstudied items (.26 vs. .09: t(75) = 3.77, p < .01).

A final set of analyses compared performance in the recognition and cued-recall tests. Compared to recognition, cued recall showed a sharp increase in the proportion of false alarms (R+K to unstudied targets) for both the young (.55 vs. .14; t(75) = 8.85, p < .001) and old (.63 vs. .22; t(75) = 7.98, p < .001). This elevation suggests that, relative to recognition, self-generation of items in cued recall increases their acceptance as old (Toth, 2005). However, this false-memory via self-generation effect was notably different as a function of age. For the young, the increase in false alarms from recognition to recall was mainly in their F responses (.12 to .46) with R showing a much smaller increase (.02 to .09). This makes sense given that the items triggering these responses were unstudied and thus could not give rise to true recollective detail. Older adults, by contrast, showed a much larger increase in false R responses (.08 to .26) along with an increase in false F responses (.15 to .37). A 2×2 ANOVA on false R responses revealed main effects of Age (F(1,150) = 24.88, MSe = .021, p < .001) and Test (F(1,150) = 31.15, MSe = .021, p < .001), as well as the Age-by-Test interaction (F(1,150) = 5.14, MSe = .021, p < .05).

In sum, older adults showed lower accuracy in both recognition and cued recall, as well as higher levels of false recollection. Both age groups showed higher false memory in cued recall, relative to recognition, suggesting that generating potential target items in cued recall increases feelings of oldness for those items. Older adults were particularly vulnerable to this effect, showing very high levels of false recollection in cued recall. This pattern suggests that, not only do older adults have impaired recollection, they also have difficulty discriminating between recollection and familiarity.

JOLs and JOL Accuracy

Older adults’ mean JOL ratings were numerically higher than those by the younger adults in both recognition (51.97 vs. 45.31) and cued-recall (53.51 vs. 43.15) but ANOVA revealed no significant differences.

We analyzed JOL accuracy in three ways. The first used a back-sorting procedure similar to those used in neuroimaging studies (e.g., Wagner et al., 1998). Figures 1 (recognition) and 2 (cued recall) show mean JOL ratings for items subsequently rated R, F, or N². As can be seen, JOLs at encoding were higher for items that would later receive a “Recollect” judgment at test, a pattern that held for both age groups and both tests; however, JOL differences between R items and the other two categories were generally greater for younger, compared to the older, adults. These impressions were confirmed in two ANOVAs with Age and Test Judgment (R/F/N) as factors. The ANOVA for recognition revealed effects of Age (F(1,69) = 4.57, MSe = 620.00, p < .05) and Test Judgment (F(2,138) = 59.66, MSe = 3979.68, p < .001) both qualified by a significant interaction (F(2,138) = 5.75, MSe = 383.22, p < .005). Planned comparisons revealed that R items garnered reliably higher JOLs than F or N items in both the young (p’s < .001) and old (p’s < .001). By contrast, JOLs did not differ for F and N items in either age group (t’s < 1). The Age-by-Judgment interaction occurred because, while R items showed no age difference (t < 1), older adults gave higher JOL ratings than the young adults for F and N items (t(73) = 3.03, p < .005; t(71) = 2.05, p < .05). Stated differently, older adults’ JOLs showed less discrimination between items that were later recollected and those that were not.

The ANOVA for cued recall also revealed main effects of Age (F(1,49) = 7.18, MSe = 4554.86, p < .05) and Test Judgment (F(2,98) = 24.86, MSe = 4151.37, p < .001) although here the interaction was not reliable (F(2,98) = 2.12, MSe = 353.24, p = .13). Planned comparisons showed that, as in recognition, test items given R judgments received higher JOLs at study than those judged F or N, for both the young (p’s < .001) and old (p’s < .005). Although JOLs were numerically higher for F items compared to N items, neither difference was reliable (p = .08 for both age groups). Age comparisons showed the same pattern found in recognition: No difference in JOLs for items judged R (t(75) = 1.29; p = .20) but higher JOL ratings by older adults for items later judged F (t(71) = 2.93; p < .01) and N (t(51)=2.56; p < .05).

In sum, JOLs were reliably higher for recollected items compared to those that were only familiar or not remembered. This pattern obtained for both tests and both age groups, although older adults showed less JOL discrimination between recollected and non-recollected items.

A second set of JOL analyses used Goodman-Kruskal gamma correlations which provide an index of how well participants’ JOLs predicted their memory performance on an item-by-item basis. Gammas were computed in two ways. The first was designed to mimic standard old/new recognition by coding both R and F (“old”) responses as 2 and N responses (“new”) as 1. For the second “Recollect Only” method, only R responses were coded as 2 with both K and N responses coded as 1. This coding was designed to assess whether JOL accuracy would be enhanced when memory performance was restricted to instances of recollection. A 2 (Age) × 2 (Test) ANOVA on the Standard-Recognition gammas revealed only a main effect of Test [recognition = .24, recall = .09; F(1,144) = 13.70, MSe = .060, p < .001]. Notably, the effect of Age was not significant (p = .31). By contrast, the corresponding Recollect-Only analysis showed an effect of Age [young = .32, old = .22; F(1,150) = 7.35, MSe = .048, p < .01] along with an effect of Test [recognition = .34, recall = .20: F(1,150) = 14.27, MSe = .048, p < .001]. Thus, whereas age differences were not observed with standard recognition gammas, such differences did emerge when memory performance was exclusively defined in terms of recollection.

A final set of analyses examined Pearson correlations between list-wide JOLs (the mean JOL for each individual) and memory accuracy (hits–false alarms). Note that these correlations were computed across subjects and thus provide an index of how well a person’s aggregate JOLs predict their performance relative to others. If JOLs are driven by details that later contribute to recollection, then list-wide JOLs should show the strongest relations with items given R responses at test. Consistent with this, young adults’ list-wide JOLs were related to the accuracy of their R judgments [recognition: r(36) = .41, p < .05; cued-recall: r(36) = .46, p < .01]. Interestingly, significant negative relations were found between young adults list-wide JOLs and their F judgments [recognition: r(36) = −.36, p < .05); cued-recall: r(36) = −.49, p < .01]. Based on the assumption that recollection and familiarity have independent influences on memory, we computed a more process-pure measure of familiarity [F/(1−R); see Jacoby, Yonelinas, & Jennings, 1997]. Correlations between JOLs and this purified measure of familiarity were much smaller and nonsignificant (r’s = .13 & .20). Note that this same general pattern obtained for the older adults although none of the individual correlations achieved significance.

Aging and JOL Accuracy

Prior research has sometimes failed to observe age-related deficits in JOL accuracy (e.g., Connor et al., 1997; Hertzog et al., 2002). In contrast, age deficits in JOL accuracy were clearly observed in the present study in both the back-sorting data and gamma correlations. Part of the discrepancy may reflect the degree to which recollection drives overall memory performance. In the present study, participants were explicitly instructed to generate images or personally meaningful associations at study; and the requirement to make Recollect/Familiar judgments at test likely increased the use of these details. Given age-related deficits in recollection and evidence that memory monitoring is tied to memory for detail (Kelley & Sahakyan, 2003), age deficits in JOL accuracy may only be observed in tasks that emphasize recollection. Conversely, age deficits are unlikely to be observed when either JOLs at study, or memory responses at test, are driven by more automatic factors. Evidence supporting this account can be found in the present study, in that age differences in the gamma correlations only emerged in the “Recollect Only” analysis; that is, when only R responses were coded as hits. When F responses were also coded as hits (“Standard Recognition”), no significant age differences were observed.

There are at least two ways in which age-related deficits in recollection could undermine memory monitoring. One is that, relative to the young, older adults may fail to recollect details that influenced their JOL ratings at study. If an item evokes sufficient detail to garner a high JOL but that detail is not recollected at test, it would likely get classified as F or N. This would explain why older adults’ JOLs for F and N judgments were higher than those for the young (see Figures 1 & 2). A second possibility concerns false memory. That is, items that do not garner high JOLs at study may nevertheless be falsely “recollected” at test, and thus get classified as R. Support for this explanation is provided by our finding that, relative to the young, older adults made significantly more false R responses to unstudied items in both recognition and cued recall. Given such high rates of false recollection, it is likely that some of the older adults’ “correct” R responses to old items were not based on details that were part of the initial encoding.

Recollection, JOL Accuracy, and Theories of Monitoring

Although we agree with others that metacognitive judgments are driven by many factors, including heuristic factors such as fluency, the present study encourages more research on the use of contextual details in mediating the accuracy of memory judgments. In the case of JOLs, people are likely capable of generating a variety of details that may influence their judgments including intra- or extra-list associations, images, idiosyncratic elaborations, or personal experiences. An interesting question is whether different detail types have systematic influences on JOLs. Personal experiences, for example, might naturally engender high JOLs. Alternatively, the type of detail may be less important than its distinctiveness. Regardless, the use of contextual details in making JOLs, and the recollection of those details at test, can be viewed as a form of “privileged access” that drives metacognitive accuracy (Lovelace, 1984).

Of course, recollection is not infallible as shown by numerous demonstrations of false recollection in both young and older adults (e.g., Jacoby, Bishara, Hessels, & Toth, 2005; Roediger & McDermott, 1995). Nevertheless, a strong case can be made that, compared to more automatic forms of memory, recollection is likely to be more reliable across situations and may be especially important for memory monitoring (Gallo et al., 2007; Kelley & Sahakyan, 2003). Indeed, although JOLs are influenced by a number of factors, whether these factors will affect JOL accuracy depends on the degree to which people notice and correctly interpret their influence, and whether they are operating at both study and test. Contextual, idiosyncratic detail is likely to fair well on both accounts. That is, the more distinctive a cue, the more likely it will be used in forming a JOL and the higher the probability that it will be recollected at test.

An early contrast in the memory monitoring literature was between direct-access (trace-strength) and inferential (cue-utilization) views. Along with others (e.g., Begg et al., 1989), we find the inferential view most compelling. However, it is interesting to note how recollection, perhaps more than any other factor, shows some compatibility with the direct-access view. For example, in Jacoby’s (1998) “direct-retrieval” model, recollection is thought to provide relatively error-free (albeit limited) access to the past. Similarly, episodic memory has been characterized as a form of mental time travel whereby a person “consciously recollects some prior episode or state as it was previously experienced” (Wheeler et al., 2001, p. 333). These notions try to capture the remarkable ability of humans to relive or recover highly specific details associated with prior events. We believe such details may play a critical role in the monitoring of memory.