Age, Cohort, and Period Effects on Metamemory Beliefs

Abstract

Questionnaires like the Metamemory in Adulthood instrument (MIA: Dixon, Hultsch, & Hertzog, 1988) have been used to examine longitudinal changes and cross-sectional age differences in multiple metamemory facets (e.g., memory self-efficacy). This study used three independent cross-sectional samples (N = 1555; ages 55–85) from the Victoria Longitudinal Study collected in 1986, 1992, and 2000 to evaluate period and cohort effects on eight MIA scales. Alternative general linear models analyzed age, cohort, and period effects, while subsequently assessing gender differences in metamemory beliefs. Period effects were detected on the MIA Internal Strategy and External Strategy scales; self-reported use of internal strategies decreased while use of external memory aids increased over the historical period. Reliable cohort (generational) differences were found for MIA Change, with the lowest levels of perceived change in individuals born between 1916 and 1925. MIA Task, measuring knowledge about memory, produced small age and cohort effects. Gender differences emerged in metamemory, especially for the Internal Strategy and External Strategy scales (women reporting higher strategy use). Gender differences were also seen for the Capacity, Locus, Anxiety, and Achievement scales, with women reporting higher perceived memory efficacy, control, memory anxiety, and greater motivation to have better memory, respectively. The historical trends in metamemory beliefs should be replicated with other measures and other populations; however, the results generally confirm conclusions from earlier cross-sectional studies regarding age sensitivity of metamemory beliefs from middle age to old age.

Introduction

Metamemory and Aging

Metamemory is a broad construct domain that encompasses several related but distinguishable facets, including beliefs and knowledge about the nature of memory, self-referent beliefs about memory ability and personal control, and use of memory strategies in everyday life (Dixon & Hultsch, 1983; Dunlosky & Metcalfe, 2009; Flavell, 1979; Herrmann, 1982; Hertzog & Hultsch, 2000).

Early work on adult development of metamemory featured the creation of questionnaires measuring self-reported memory ability, perceived change, and related constructs (see reviews by Dixon, 1989; Gilewski & Zelinski, 1986; Hertzog & Hultsch, 2000; Hertzog & Pearman, 2014). Older adults are typically found to report lower memory ability, lower personal control over memory, and greater change in their memory ability during adulthood, compared with young adults or middle-aged adults (e.g., Dixon & Hultsch, 1983; Gilewski, Zelinski, & Schaie, 1990; Hultsch, Hertzog, & Dixon, 1987; Jopp & Hertzog, 2007; Lachman, Bandura, Weaver, & Elliott, 1995; Lineweaver & Hertzog, 1998). Longitudinal studies indicate age-related reductions in perceived memory ability and control and increases in memory complaints during adulthood into old age (e.g., Hülür et al., 2014; Hülür, Hertzog, Pearman, & Gerstorf, 2015; Lane & Zelinski, 2003; Mascherek & Zimprich, 2011; Parisi et al., 2011, Rickenbach & Lachman, 2014; Valentijn et al., 2006).

The Metamemory in Adulthood Questionnaire (MIA), originally created by Dixon and Hultsch (1983) and empirically validated by item factor analysis is one of the most widely used metamemory questionnaires in gerontology. One of the MIA’s key features is that it includes separate scales for measuring multiple relevant metamemory facets, including subjective memory ability, perceived memory change, perceived control over memory, anxiety about memory, achievement motivation about memory, knowledge about memory and how it functions, and everyday memory strategies. Notably, MIA scales assessing ability, perceived change, personal control, and anxiety form a higher-order memory self-efficacy factor (Hertzog et al., 1987). They are the scales that typically show cross-sectional age differences in empirical studies (e.g., Hultsch et al., 1987). These scales have also been shown to manifest convergent validity with similar scales in the Memory Functioning Questionnaire (MFQ; Gilewski, Zelinski, & Schaie, 1990), another widely used metamemory questionnaire (Hertzog, Hultsch, & Dixon, 1989). The everyday memory strategies scales of the MIA also show convergence with a similar scale from the MFQ. The MIA scales measuring achievement motivation and knowledge are unique, in the sense that they are not found in other common metamemory questionnaires. The strategy, motivation, and knowledge scales correlate only weakly with those related to perceived memory ability, change, and control and, unlike the scales related to memory self-efficacy, typically do not yield adult age differences (Hultsch et al., 1987).

Life-span sciences emphasize the conceptual distinctions between outcomes that covary with how old a person is (age effects) and outcomes that covary with historical time, including generational (or cohort) effects and time period effects (see Alwin, 2009; Ram & Shores, 2018; Schaie, 1986; Shanahan, Elder, & Meich, 1997; Wohlwill, 1973). This study focuses on age, cohort, and period effects in metamemory.

Virtually nothing is known about whether there are historical shifts in metamemory beliefs in western cultures. The operating assumption in the literature appears to have been that cross-sectional age differences reflect phenomena tied to developmental influences rather than cohort differences. Life-span developmental research certainly questions this assumption. For one thing, it is well known that metamemory beliefs are tied to aspects of personality, including especially neuroticism and conscientiousness (e.g., Hülür et al., 2015; Pearman & Storandt, 2005) that may be prone to cohort effects (e.g., Milojev & Sibley, 2017; Twenge, 2000). Furthermore, metamemory beliefs probably arise in part because of monitoring memory and cognition in everyday life (Hertzog, Dixon, & Hultsch, 1990; Hertzog & Hultsch, 2000). Hence the fact that memory and other intellectual abilities manifest birth cohort differences (e.g., Hultsch, Hertzog, Dixon, & Small, 1998; Schaie, 2012; Zelinski & Kennison, 2009) could also lead to cohort differences in metamemory.

Furthermore, stereotypes and beliefs about aging and cognition have been changing in recent historical time (e.g., Hummert, 2011). There have been societal changes in awareness about Alzheimer’s Disease, including its prevalence, multi-factorial mechanisms, global spread, societal costs, and its divergence from normal aging (Alzheimer’s Association, 2016; Prince et al., 2016). Historical changes in societal awareness, attitudes and beliefs about memory and aging could generate both cohort and period effects in metamemory beliefs and knowledge.

Lachman et al. (1995) showed that belief in the inevitability of cognitive decline was correlated with both subjective memory ability and concerns about Alzheimer’s Disease (see also Cutler & Hodgson, 1996; 2001). More recently, subjective memory decline, especially when coupled with worry or anxiety about personal decline, have been linked to actual memory change as well as to dementia risk (Jessen et al., 2014; Koppara et al., 2015; Snitz et al., 2014). Historical changes in societal attitudes and stereotypes about memory and aging could account for some of the variance in such concerns, and could in theory generate cohort and period effects in metamemory. Indeed, the available cross-sectional evidence regarding age differences in stereotypes about memory and aging (Hummert, 2011; Lineweaver, Berger, & Hertzog, 2009; Ryan & Kwong See, 1993) could reflect an as yet undetected mixture of age changes and cohort effects.

This study uses cross-sectional sequences (Baltes, 1968; Schaie, 1977) from the Victoria Longitudinal Study (VLS; Dixon & deFrias, 2004; Hultsch, Hertzog, Dixon, & Small, 1998) to explore whether there are cohort and period effects in metamemory knowledge and beliefs. There are to our knowledge no existing studies of whether subjective or metamemory beliefs, knowledge, and strategic behaviors have changed over historical time from the time period when the early metamemory questionnaires were developed in the 1980s. In order to address this lack of knowledge we analyzed MIA data from three independent cross-sectional samples of adults, ages 55 to 85, from the VLS collected in 1986, 1992, and 2000, modeling age, cohort, and period effects.

Methodological Approach

A well-known problem in cohort analysis is the linear dependency of time-based indices of these variables. A person’s current age, year of birth, and the current date are linearly dependent, creating issues for analyzing for age, birth cohort, and time period effects on dependent variables (e.g., Baltes & Nesselroade, 1970; Schaie, 1977). In order to break this linear dependency when analyzing developmental data, statistical models for age, period, and cohort effects must create sufficient constraints to uniquely identify the relevant parameters. Collecting longitudinal or cross-sectional sequences (Baltes, 1968) aids in the identification by sampling slices of age, cohort, or time in ways that begin to generate replicated evidence for each kind of outcome, although additional constraints must be imposed (Donaldson & Horn, 1992). Cross-sectional sequences, in which cross-sectional samples from a population are repeated over an extended time span, are ideal for investigating cohort and period effects. They avoid some of the inherent issues with longitudinal models (e.g., practice effects, experimental mortality) and are easily and appropriately analyzed with the general linear model. Longitudinal data are best-suited for scaling individual differences in within-person changes (Baltes & Nesselroade, 1979; McArdle & Nesselroade, 2014), but repeated measures on the same individuals are not needed for estimation of aggregated age, cohort, and period effects.

Briefly, our models address the linear dependency issue as follows. First, we model chronological age effects with age as a continuous variable, using linear and quadratic terms of a polynomial regression equation. This places overidentifying restrictions on age that, when combined with sequential sampling, helps to identify the cohort and period effects. Second, we create categorical birth cohort and time of measurement variables. In the cross-sectional sequences sampling design, period is perfectly confounded with date of sampling; hence we must assume equivalent and invariant sampling procedures for the three cross-sectional samples as well as stationary (unchanging) population processes (Baltes, Reese, & Nesselroade, 1988). In terms of the logic and language of quasi-experimental design, we assume no time × selection variation that would artifactually mimic period effects (Schaie, 1977; Shadish, Cook, & Campbell, 2002). Third, we assume that additive models suffice to capture the time-related phenomena (i.e., there are no interactions of age, cohort, and period). As described below, we explicitly checked this assumption for the age × cohort models. This approach allowed us to estimate age, cohort, and period effects simultaneously, with some additional attendant issues outlined in greater detail in the Methods section.

Gender Differences in Metamemory

An important but under-evaluated issue is whether men and women differ in their metamemory beliefs. Gender differences have not often been evaluated in studies of adult metamemory; even when studies contain gender-heterogeneous samples they may not analyze for gender differences. Hultsch et al. (1987), using both the MIA and MFQ, found that women rated themselves as more anxious about memory and reported more frequent use of everyday memory strategies. No gender differences were found for MIA Capacity and Change. Hertzog, Dixon, and Hultsch (1990) found smaller gender differences in MIA Capacity and in memory task performance predictions than observed in performance on word list recall and text recall (women outperformed men but their beliefs did not anticipate that difference). Stevens et al. (2001) found gender differences on MIA Capacity, and Zelinski and Gilewski (2004) reported small gender differences on the principal MFQ scale, Frequency of Forgetting (a measure of memory complaints closely related to MIA Capacity; Hertzog et al., 1989). Conversely, Hülür et al. (2015) found a nil relation of gender to memory complaints on a single Likert rating-scale item in a representative sample of over 15,000 adults in the Health and Retirement Survey. However, the importance of examining gender differences in related domains has been established (e.g., Tierney, Curtis, Chertkow, & Rylett, 2017; Weber et al., 2014) and observed in the related area of subjective cognitive decline (e.g., Peres et al., 2011). Given the large sample size available in this study, we opted to examine gender differences in the present eight facets of metamemory.

Primary and Secondary Research Aims

Our primary research aim was to explore the existence of period or cohort effects in metamemory during the historical time interval of 1986–2000. We had no specific a priori hypotheses, although we were particularly interested in the possibility that historical change would affect MIA scales related to memory self-efficacy already known to be sensitive to cross-sectional age differences (i.e., MIA Capacity, Change, Locus, and Anxiety).

An important secondary aim was to examine whether cohort effects would be detected that would qualify or even wholly produce previously observed cross-sectional age differences in metamemory. The original impetus for cohort analysis in Schaie’s work was the conjecture that age differences in intelligence reflected cohort differences rather than true age-related changes, a source of considerable controversy (e.g., Baltes & Schaie, 1976; Horn & Donaldson, 1976). As in the literature on adult intellectual development (e.g., Schaie, 2012), we thought it more likely that cohort differences could lead to overestimation of age-related changes in metamemory. A major advantage of the use of cross-sectional sequences was also the ability to estimate any age-related effects in a relatively large sample.

Another secondary research aim involved assessing mean gender differences in metamemory. It seems safe to conclude that gender differences in metamemory beliefs are small in magnitude, based on the literature reviewed earlier. Hence evaluating them in a larger sample based on the VLS cross-sectional sequence should provide better estimates of gender differences and more power to detect gender differences if they exist.

Methods

Participants

Participants were community-dwelling adults (initially aged 53–91 years) drawn from the VLS (Dixon & de Frias, 2004). The VLS and all present data collection procedures were in certified compliance with prevailing human research ethics guidelines and boards. Informed written consent was provided by all participants. The study sample consisted of Wave 1 from three VLS samples. Sample 1 was collected in 1986–1988 (N = 484). Sample 2 was collected in 1992–1993 (N = 528), Sample 3 was collected in 2000–2002 (N = 570). To maintain comparability across samples, we excluded any adult with a reported age outside the interval of 55 to 85. The final study sample consisted of n = 1554 (M_Age = 68.4 (SD = 7.2); 65.0% female) older adults (see Table 1). Table 1 also reports descriptive statistics on years of education and two measures of self-reported health – absolute health and rated health relative to one’s same-age peers. We ran an age-cohort-period analysis of these three possible covariates. There was a reliable period effect in years of education, F (2, 1527) = 11.87, p < .001, but not in the two health-related variables (F < 1). Successive samples showed monotonic increases in years of education from 1986 to 2000.

Table 1.

Descriptive Statistics for Pooled Victoria Longitudinal Study (VLS) Sample and The Three Contributing VLS Samples.

	Study Sample	VLS Sample 1	VLS Sample 2	VLS Sample 3
N	1554	483	519	552
Age M (SD)	68.4 (7.2)	69.2 (5.8)	68.3 (7.3)	67.9 (8.1)
Range	55–85	55–85	55–85	55–85
Female Sex n (%)	1010 (65.0)	286 (59.2)	348 (67.1)	376 (68.1)
Education M (SD)	14.6 (3.1)	13.4 (3.1)	14.8 (3.1)	15.2 (2.9)
Range	3–26	6–23	6–26	3–24
Health Relative to Perfect M (SD)	0.80 (0.75)	0.82 (0.76)	0.76 (0.73)	0.82 (0.76)
Health Relative to Peers M (SD)	0.60 (0.71)	0.62 (0.70)	0.57 (0.70)	0.60 (0.72)

Design

The VLS has been described in detail elsewhere (Dixon & de Frias, 2004; Hultsch et al., 1998). We created four synthetic birth cohorts across the three samples as follows: Cohort 1: birth years 1898 thru 1915; Cohort 2: 1916 thru 1925; Cohort 3: 1926 thru 1935; Cohort 4: 1936 thru 1947. Table 2 shows the resulting cross-sectional sequence separated by birth cohort, reporting mean year of birth and age within the three times of measurement (samples).

Table 2.

Descriptive Statistics of Cross-sectional Sequence Separated by Birth Cohort.

	Cohort 1	Cohort 2	Cohort 3	Cohort 4
N (%)	263 (16.9)	591 (38.0)	421 (27.1)	279 (18.0)
Age M (SD)	76.4 (4.0)	70.6 (5.9)	66.1 (4.9)	59.8 (3.3)
Range	70–85	60–85	55–76	55–66
Female Sex n (%)	166 (63.1)	365 (61.8)	270 (64.1)	209 (74.9)
VLS Sample 1 (%)	192 (39.8)	267 (55.3)	24 (5.0)	-
VLS Sample 2 (%)	71 (13.7)	207 (39.9)	207 (39.9)	24 (6.6)
VLS Sample 3 (%)	-	117 (21.2)	190 (34.4)	245 (44.4)

Measures

The MIA questionnaire consists of everyday memory-related items that participants rate in terms of their beliefs, knowledge and use, with each item measured by a 5-point Likert scale (Dixon & Hultsch, 1983; Dixon, Hultsch, & Hertzog, 1988). The MIA consists of 108 items that are distributed into 7 scales: (a) Strategy, the use of memory strategies (18 items with high score = high use); (b) Capacity, ratings of one’s own memory performance and abilities (17 items with high score = high capacity); (c) Change, reflecting one’s perception of one’s own recent memory change (18 items with high score = stability); (d) Locus, perceived personal control over memory abilities (9 items with high score = internal locus); (e) Anxiety, knowledge of anxiety’s negative effects on memory and a personal tendency to become anxious in memory-demanding situations (14 items with high score = higher experienced anxiety); (f) Achievement, motivation to achieve in everyday memory performance tasks (16 items with high score = high achievement); and (g) Tasks, knowledge of how memory tasks affect memory performance (15 items with high score = high knowledge). We used the summative scales created by Dixon and Hultsch (1983) to facilitate comparisons to other studies using MIA. Reported internal validity measured by Cronbach’s α > 0.70 for all scales, ranged from .71 for Locus to 0.93 for Change (Dixon, Hultsch, & Hertzog, 1988). We divided the original MIA Strategy scale into MIA Internal Strategy (9 items) and MIA External Strategy (9 items). Estimated internal consistency for these two subscales was .76 and .75, respectively. A scale was treated as missing if it had any item with missing responses; missing values were recorded on fewer than 2% of the available cases.

Procedure

Our statistical analysis approach for detecting age, period, and cohort effects was as follows. We treated age as continuous variable with linear and quadratic terms. Such an approach may miss some age-graded effects at specific points in the adult life span, such as changes in metamemory associated with age of retirement, but it is defensible if one assumes more continuous and subtle changes that derive from monitoring changes in everyday cognition over the adult life span. We also treated birth cohort and time period as categorical predictor variables (as in Table 2), accepting as error any within-sample variation in year of testing.

The cross-sectional sequence used in this study helps to remove the linear dependency among age, period, and cohort by creating an incomplete quasi-experimental factorial design (replicating the age range across birth cohorts) (Schaie, 1977). An underappreciated problem in cohort analysis is that even with a cross-sectional sequence sampling design that eliminates the complete linear dependency of age and cohort, the empirical correlation of these two variables will remain substantial. In the present sequential data set, the aggregate correlation of age and year of birth (pooling over time period) was still −.80. This level of near-collinearity creates well-known problems in estimating and testing regression coefficients. Statistically, we can algebraically identify the unknown age, cohort, and period parameters, but significance tests of the different sources may be excessively conservative due to mutual suppression (overlapping, nonunique partial covariances with the dependent variable). Furthermore, these characteristics make it highly likely that interaction effects of age and cohort cannot be detected, even if true interaction effects exist.

Our approach to this problem was to estimate three alternative general linear models: (a) an age-period (AP) model, omitting cohort parameters; (b) an age-cohort model (AC), omitting period parameters, and (c) an age-cohort-period (APC) model, including additive effects for all three sources. Given a primary research focus on age-related effects we made no effort to evaluate a period-cohort model, omitting age parameters. We fully expected that the APC model would result in conservative tests of the relevant effects, so we also evaluated whether age effects were seen in the AP and AC models, whether cohort effects were detected in the AC model, and whether period effects were observed in the AP model. We compared results across the three models to develop an interpretive filter for understanding the likely presence of each type of effect. All models were estimated in the general linear model (GLM) program of SPSS Version 24 (IBM Corp, 2016), using Type III sums of squares extraction to deal with effect tests in the nonorthogonal design.

In the AC model we first performed a preliminary evaluation of age × cohort interactions, and found none. As noted, this result is not definitive given empirical identification issues, but it can be taken as a justification to focus only on additive age, cohort, and period effects in our models. Our approach was similar to Schaie’s (1965) original cohort-sequential analysis approach (e.g., Schaie, 2012), but it differed in that (a) we do not assume a lack of period effects a priori, and (b) age is treated as a continuous variable.

In all three (AP, AC, APC) models, gender was also treated as a categorical predictor variable. We report inferential tests for Gender, Age, Cohort, and Period, as well as the partial η² effect-size statistic generated by the SPSS GLM program. As will be seen, the interrelations of age, period, and cohort generate variability in significance tests for the AP, AC, and APC models. In contrast, the effects of Gender were highly similar across the 3 models for a given MIA scale, given minimal correlations of gender with the other predictors (see Table 3). To facilitate evaluation of the gender effect sizes, we also report standardized mean differences using Cohen’s (1988) d statistic using fitted least-squares means for men and women estimated by GLM. We scaled d as (M_Women - M_Men ) / (MS_Error)^.5. Cohen (1988) suggested benchmarks for d of 0.2 for small effect size, 0.5 for a medium effect size, and 0.8 (and above) for a large effect size.

Table 3.

Correlations Among Metamemory in Adulthood Scales and Age, Year of Birth, Gender, and Education for Pooled Victoria Longitudinal Studies Samples (N = 1555), along with the aggregate M and SD for all variables.

Variable	1.	2	3	4	5	6	7	8
1. MIA Internal Strategy	1
2. MIA External Strategy	.48	1
3. MIA Capacity	−.01	−.18	1
4. MIA Change	−.08	−.17	.58	1
5. MIA Locus	.16	−.02	.35	.46	1
6. MIA Anxiety	.22	.21	−.47	−.49	−.20	1
7. MIA Achievement	.37	.16	.11	−.13	.24	.36	1
8. MIA Task	.27	.16	−.08	−.11	.08	.18	.26	1
9. Age	−.01	−.02	−.10	−.25	−.17	.06	.04	−.03
10. Year of Birth	−.06	.06	.10	.18	.10	−.06	−.04	−.02
11. Gender	−.14	−.23	−.10	−.01	−.07	−.13	−.09	.02
12. Education	−.02	−.02	.08	.10	.00	−.14	−.06	.03
13. Health	.03	.01	.14	.16	.17	−.17	−.01	.07
M	31.43	34.65	52.10	49.38	31.19	41.93	58.30	63.42
SD	4.88	5.28	9.21	11.01	5.01	8.79	7.19	6.20

Note: Pairwise N excluding missing data ranged from 1529 to 1546. Critical value of r to reject null hypothesis of 0 population correlation, two-tailed test with α = .01, N = 1529 is |r| ≥ .066.

Abbreviation: MIA – Metamemory in Adulthood Questionnaire.

Results

Correlations among MIA Scales

Table 3 reports Pearson correlations among the MIA scales and relevant predictors including, gender, chronological age, year of birth, education, and self-rated health. The subjective health variable was the sum of the two Likert-rating self-rated health variables, reversed so that high scores indicated excellent health.

As expected, the correlations among MIA scales were consistent with previous studies (e.g., Dixon & Hultsch, 1983; Hertzog et al., 1987; Hertzog et al., 1989) with the strongest correlations among the Capacity, Change, Locus, and Anxiety scales. The two strategy subscales were moderately correlated as well. Gender and age correlated with a number of MIA scales and represent a starting point for the GLMs using these predictors. Correlations with age were fully consistent with earlier studies using the MIA (Dixon, Hultsch, & Hertzog, 1988). There were small correlations of MIA scales with gender that will be analyzed in the context of the GLM analyses reported next. Correlations with education were very small in magnitude. For self-rated health, correlations were also small and only appreciably above zero for those MIA scales related to the higher-order memory-self efficacy construct (Hertzog et al., 1987): Capacity, Change, Locus, and Anxiety.

To check on VLS sample differences in MIA scale correlations, we analyzed the three samples in a multiple-group covariance structures model run in Mplus (Muthén & Muthén, 2012). Forcing equality of all population variances and covariances (Box’s M test) resulted in a minor loss of fit, χ² (72) = 118.61, p < .001, RMSEA = .032, 90% CI = [.024, .046]. A benchmark of .05 is often used as an indication of excellent model fit (Hu & Bentler, 1999). We then imposed a model that allowed MIA scale standard deviations to be freely estimated in each sample, but constrained the correlations to be equal across samples. This model fit well, such that the null hypothesis of equal correlations could not be rejected, χ² (56) = 54.50, p = .53. This result indicates that the sample differences in covariance matrices were isolated to the scale variances, and therefore suggests no period effects on MIA scale correlations. Indeed, the constrained-equal estimates of the correlations were virtually identical to those reported in Table 3.

Table 4 reports the estimated SDs for the three VLS samples from the Mplus model constraining the correlations to be equal across groups. In general, the sample differences in variability reflected higher variance in Sample 2 relative to Samples 1 and 3 on most MIA scales. The differences were most pronounced for the MIA scales Capacity, Change, and Locus, suggesting differences in scales related to memory self-efficacy, but there were no sample differences in variability on MIA Anxiety (also related to MSE; Hertzog et al., 1987). Differences in variability were smallest for the Internal Strategy scale.

Table 4.

Estimated variability in MIA Scales (fitted SDs with standard errors in parentheses) from a model forcing equal sample correlations) across the three VLS Samples.

MIA Scale		Sample 1 (1986)	Sample 2 (1994)	Sample 3 (2002)
Internal Strategy		4.82 (0.14)	4.95 (0.14)	4.77 (0.14)
External Strategy		5.57 (0.17)	5.54 (0.16)	4.67 (0.14)
Capacity		8.92 (0.25)	9.64 (0.28)	9.04 (0.25)
Change		10.74 (0.30)	11.84 (0.34)	10.47 (0.28)
Anxiety		8.89 (0.26)	8.78 (0.25)	8.69 (0.24)
Locus		5.16 (0.15)	5.54 (0.16)	4.54 (0.13)
Achievement		7.01 (0.21)	7.64 (0.22)	6.89 (0.19)
Task		6.14 (0.20)	6.39 (0.19)	6.02 (0.18)

Abbreviation: MIA – Metamemory in Adulthood Questionnaire.

Age, Cohort, Period, and Gender Effects in Metamemory

Given that we observed period effects on years of education, we ran preliminary APC models covarying on years of education. There were significant education effects on MIA Capacity, MIA Change, and MIA Anxiety in the expected direction (higher education was associated with greater perceived memory ability, lower levels of perceived memory change, and lower levels of anxiety about memory). However, consistent with the modest correlations of Education with MIA scales (see Table 3), covarying on Education did not affect any conclusions regarding age, cohort, and period effects on the MIA scales. Hence we report models that did not use Education as a covariate.

We also evaluated higher-order interactions. We tested for interactions of Age × Period in the AP models and for Age × Cohort interactions in the AC cohort-sequential models and found none. We also evaluated models with gender-related interactions (e.g., gender × period), generally without detecting them. As is common in moderated regression analysis (Cohen, Cohen, West, & Aiken, 2003), including these interactions typically eliminated the effects of gender. Hence, we report results for models with additive and not moderated regression effects.

Table 5 provides significance tests and effect sizes for the alternative GLMs for each MIA scale. Given correlations among the age, period, and cohort variables, which would work against detecting unique effects, the Type I error rate was set at 5%. Figure 1 overlays fitted polynomials for age for the key AP and AC models, contrasting how fitting cohort effects, in particular, influenced estimated age effects for the eight MIA scales. We describe and interpret the alternative models for each MIA scale in turn.

Table 5.

Statistical Test and Effect Size Statistic (η_p²) for Gender, Age(Linear), Age(Quadratic), Period, and Cohort sources of variance in Alternative Age, Cohort, and Period Models for the Metamemory in Adulthood Questionnaire Scales.

MIA Scale		Model	df (Error)	Gender	Age(Linear)	Age(Quadratic)	Period	Cohort
				F / ηp2	F / ηp2	F / ηp2	F / ηp2	F / ηp2
Internal Strategy
		AP	1536	33.79*** / .022	< 1 / --	< 1 / --	8.76** / .011	-- / --
		AC	1535	33.44*** / .021	4.67* / .009	< 1 / --	-- / --	4.74* / .009
		ACP	1532	32.76*** / .021	< 1 / --	< 1 / --	3.03* / .004	< 1 / --
External Strategy
		AP	1542	81.75*** / .050	< 1 / --	< 1 / --	3.26* / .004	-- / --
		AC	1541	82.31*** / .051	1.17 / .001	< 1 / --	-- / --	1.14 / .002
		ACP	1539	81.26*** / .050	< 1 / --	< 1 / --	1.85 / .002	< 1 /--
Capacity		AP	1542	13 79*** / .009	12.45*** / .008	2.15 / .001	1.09 / .001	-- / --
		AC	1541	14.30*** / .009	1.92 / .001	< 1 / --	-- / --	1.49 / .003
		ACP	1539	14.03*** / .009	< 1 / --	< 1 / --	< 1 / --	< 1 / --
Change		AP	1537	< 1 / --	84.22*** / .052	9.97** / .006	< 1 / --	-- / --
		AC	1536	< 1 / --	32.90*** / .021	3.97* / .003	-- / --	3.14* / .006
		ACP	15.34	< 1 / --	13.14*** / .008	4.55* / .003	< 1 / --	2.97* / .006
Locus		AP	1542	8.37** / .005	41.86*** / .026	7.58** / .005	1.46 / .002	-- / --
		AC	1541	8.98** / .006	29 34*** / .019	2.71 / .002	-- / --	1.93 / .004
		ACP	1539	8.99** / .006	7 47** / .005	2.50 / .002	< 1 / --	< 1 /--
Anxiety		AP	1540	26.41*** / .017	4.00 / .003	2.99 / .002	2.17 / .003	-- / --
		AC	1539	25.81*** / .016	< 1 / --	1.79 / .001	-- / --	2.36 / .005
		ACP	1537	26.55*** / .017	1.03 / .001	2.51 / .002	1.64 / .002	2.01 / .004
Achievement		AP	1535	11 84*** / .008	1.95 / .001	3.74 / .002	2.11 / .003	-- / --
		AC	1534	12 43*** / .008	< 1 / --	4.04* / .003	-- / --	2.51 / .005
		ACP	1532	12.04*** / .008	< 1 / --	3.35 / .002	< 1 / --	1.56 / .003
Task		AP	1536	< 1 / --	1.61 / .001	< 1 / --	3.35* / .004	-- / --
		AC	1535	< 1 / --	5.77* / .004	< 1 / --	-- / --	3.15* / .006
		ACP	1433	< 1 / --	< 1 / --	< 1 / --	< 1 / --	1.50 / .003

Note: F-tests below 1.0 are deemed too small to evaluate and are shown simply with < 1. For AP Model, no Cohort effect was estimated; For AC Model, no Period effect was estimated. η_p² indicates partial η² effect size statistic available in SPSS GLM.

Abbreviations: MIA – Metamemory in Adulthood Questionnaire; AP – Age-Period Model; AC – Age-Cohort Model

^*p < .05

^**p < .01

^***p < .001

Figure 1.

Simple chronological age curves estimated for the Age-Period and Age-Cohort models for each of the eight MIA scales. Panel a: Internal Strategy; Panel b: External Strategy; Panel c: Capacity; Panel d: Change; Panel e: Anxiety; Panel f: Achievement; Panel g: Locus; Panel h: Task.

Note: Period signifies the fitted age curve from the Age-Period Model; Cohort signifies the fitted age curve for the Age-Cohort Model.

MIA Internal Strategy

The pattern of results from the three models suggested the existence of period effects in reported use of internal strategies. As documented in Table 5, there were robust effects of Period in both the AP and APC models. The AC model revealed a small cohort effect that was not significant in APC model. Fitted LS marginal means for the three time periods revealed decreasing reliance on internal strategies from 1986 to 2000 (see Figure 2a). The linear trend was significant, est. (ψ) = −0.78, SE = 0.22, p < .001, Bonferroni-adjusted critical p = .025. The quadratic trend was not reliable, t < 1.

Figure 2.

Fitted Least Squares Marginal Means for Time Period for the MIA Internal Strategy and MIA External Strategy Scales from the Age-Period Model.

Note: Error bars denote 1 SE of the fitted mean.

There were robust gender differences in all three models with women reporting more internal memory strategies. Fitted least-squares means from the APC model were estimated at M_Women = 32.0, SE = 0.17; M_Men = 30.5, SE = 0.22, Cohen’s (1988) d = 0.31.

MIA External Strategy

There was a robust effect of Period in the AP model with no hint of a Cohort effect in the AC model. However, unlike the Internal Strategy scale, Period effects were not significant in the APC model. We regarded this pattern as evidence for a period effect of weaker magnitude, consistent with the effect sizes reported in Table 5. Fitted LS marginal means for Period in the AP model revealed increasing reliance on external memory aids from 1986 to 2000 (see Figure 2b), with a significant linear trend, est. (ψ) = −0.69, SE = 0.24, p = .004, Bonferroni-adjusted critical p = .025. The quadratic trend was not reliable, t < 1.

Gender differences in external strategy use were reliable, with women reporting more frequent use of external aids, M_Women = 35.5, SE = 0.18; M_Men = 33.0, SE = 0.24, d = 0.49.

MIA Capacity

Robust linear age effects were manifested in AP model but eliminated in the AC model. Figure 1c shows that the linear age effect was attenuated in magnitude but similar in shape in the APC model relative to the AP model. However, Cohort was not significant in either the AC and APC models. No period effects were observed in either the AP or APC models.

Women reported significantly higher memory ability, M_Women = 52.6, SE = 0.32; M_Men = 50.8, SE = 0.42, d = 0.20.

MIA Change

As noted, the Capacity and Change scales were substantially correlated. Yet the pattern of age, period, and cohort effects were quite dissimilar. MIA Change yielded highly reliable linear and quadratic age effects in all three models, with largest effect sizes in the AP model, a smaller effect in the AC model, and the smallest effect size in the APC model. As can be seen in Figure 1d, the age curve for Change reflected accelerating perceptions of change (more decline) with advancing age. Furthermore, this scale produced reliable cohort differences in both the AC and APC models, which correspondingly reduced the age effect size. However, attenuation in the fitted age curve by adding cohort effects was minimal (see Figure 1d), suggesting the reduced effect size was due to overlapping (shared) variance by age and cohort. The pattern of cohort effects was for less perceived change in Cohort 2 (birth years 1916 thru 1925) relative to Cohort 1 (birth years 1898 thru 1915) (see Figure 3), with significant Tukey-adjusted pairwise comparisons. None of the other pairwise comparisons were reliable. The same qualitative pattern of fitted means for cohorts was observed in the APC model, but none of the adjusted comparisons were statistically reliable.

Figure 3.

Fitted Least Squares Marginal Means for Cohort Groups for the MIA Change scale, estimated from the Age-Cohort Model.

Note: Error bars denote 1 SE of the fitted mean. Minimum-Maximum Birth Years -- Cohort 1: 1898 thru 1915; Cohort 2: 1916 thru 1925; Cohort 3: 1926 thru 1935; Cohort 4: 1936 thru 1947

Unlike Capacity, there was no indication of gender differences in perceived memory change, M_Women = 49.2, SE = 0.37; M_Men = 49.0, SE = 0.50, d = 0.03.

MIA Locus

Significant linear age effects were found in all three models, with perceived control differences from middle age through old age. Similar to the Change scale, age differences were largest in the AP model and attenuated in the AC and APC models. However, as seen in Figure 1e the shape of the curves was not reduced by including cohort effects in the AC model, relative to the AP model. Neither Period nor Cohort tests were statistically reliable.

There were significant gender differences in Locus, with women reporting more control over memory than men, M_Women = 31.4, SE = 0.17; M_Men = 30.6, SE = 0.23, d = 0.16.

MIA Anxiety

A modest age effect, reflecting increasing anxiety with increasing age was detected in the AP model but it was eliminated when cohort effects were included in the model. Cohort and period effects were not statistically significant in any model.

The Anxiety scale produced robust gender differences, with women reporting more anxiety about memory, M_Women = 42.9, SE = 0.30; M_Men = 40.5, SE = 0.40, d = 0.28.

MIA Achievement

A small quadratic age effect was detected in the AC model, with achievement motivation increasing more in later life. However, this effect was not reliable in the AP and APC models. There were no reliable Period and Cohort effects.

Women were reliably higher than men in achievement motivations regarding memory, M_Women = 58.8, SE = 0.25; M_Men = 57.5, SE = 0.33, d = 0.19.

MIA Task

The models for this scale were somewhat divergent. The AP analysis indicated a small period effect, but there was little hint of that in the APC model. The AC model generated small Age and Cohort effects. Taken at face value, the AC model suggested that a nil cross-sectional age correlation (see Table 3) was decomposed by the GLM into cohort effects (with higher knowledge for Cohort 2 (birth years 1916 thru 1925) relative to later born cohorts; see Figure 4) and a small negative age effect (a linear decrease in knowledge with age).

Figure 4.

Fitted Least Squares Marginal Means for Cohort Groups for the MIA Task scale, estimated from the Age-Cohort Model.

Note: Error bars denote 1 SE of the fitted mean. Minimum-Maximum Birth Years -- Cohort 1: 1898 thru 1915; Cohort 2: 1916 thru 1925; Cohort 3: 1926 thru 1935; Cohort 4: 1936 thru 1947

There were no gender differences in knowledge about memory, M_Women = 63.2, SE = 0.22; M_Men = 63.5, SE = 0.29, d = −0.04.

Discussion

This analysis of metamemory beliefs revealed a set of interesting findings regarding age, period, and cohort effects, as well as evidence of reliable gender effects in these beliefs.

Regarding period effects, we obtained clear evidence for historical period differences in responses in the strategy-related MIA scales. Reports of use of mnemonics and other memory strategies, as captured by Internal Strategy, decreased from 1986 to 2000. In contrast, reported External Strategy use manifested the opposite pattern, with increasing use of external aids over the same time period. Although internal strategies can produce powerful improvements in domain-specific memory performance, external aids are arguably of more practical and generalizable use for everyday memory problems encountered by older adults (Dixon & de Frias, 2007). Note that the separation of internal and external memory strategies was critical to detect these opposing influences that might have been overlooked had the original MIA Strategy scale that combines these two subscales (Dixon et al., 1988) been used instead.

These period effects were relatively small in magnitude, but they may signify the start of historical changes in the nature of external aid usage that has accelerated in the early 21^st century given recent explosions in smartphone and tablet technology. Technological advances may lead to increasing reliance on outsourcing everyday prospective memory to digital devices. Lee et al. (2018) recently reported historical changes from 1994 to 2013 in attitudes towards technology and computers that maintained but reduced the well-known “digital divide” – older adults being less likely to adopt computer technology (Czaja et al., 2006), including time-management applications that would be considered external aids. Lee et al. (2018) also found cohort effects in attitudes and reported use in computers. It is possible that continuing historical change in technology availability, adoption, and use will lead to further changes in self-reported use of external memory aids and also to cohort differences in reported use of external memory aids that parallel Lee et al.’s (2018) findings about computer use, at least until today’s younger technology-savvy cohorts become middle-aged and older adults.

The MIA Change scale generated the most clear-cut evidence of cohort effects in metamemory beliefs. This effect co-occurred with substantial age differences in perceived memory change; indeed, the cohort difference did not greatly attenuate regression parameters for age. Part of the reason for this was probably the nonmonotonic pattern of cohort differences, with the 1921 birth cohort showing the least perceived change (see Figure 3). One is tempted to discount the effect due to its non-monotonicity, but a variety of phenomena could generate this pattern. For instance, individuals born after World War I probably witnessed more graceful cognitive aging in family members given advances in medical practices, but later cohorts may have become more aware of risks for Alzheimer’s disease. As is typically the case with cohort analyses, there is, at present, no way to know the mechanisms underlying these descriptive results.

Another cohort effect detected in this study involved MIA Task, but only in the AC model. This effect co-occurred with estimated age-related effects. Taken at face value, the cohort effects, reflecting higher knowledge in earlier born cohorts, could be concealing (through classical statistical suppression) age-related changes in self-rated knowledge about memory.

The general impression from these models is that there are only small influences of cohort and period on metamemory beliefs. One implication of this findings is that the age effects on metamemory beliefs observed here, when jointly analyzing for age, period, and cohort effects, are consistent with the cross-sectional age differences already reported in the literature. Thus, these age differences in metamemory knowledge and beliefs are not attributable to confounded cohort differences. To the contrary, the hypothesis that cohort effects are the true source of cross-sectional age differences on the MIA scales previously reported to manifest substantial age differences (Capacity, Change, Locus) appears to have been falsified. The age differences in metamemory previously identified in the literature are truly correlated with chronological age more than year of birth, and perhaps (but not necessarily), with adult developmental changes (see Wohlwill, 1973).

The MIA Capacity, Change and Locus scales all appear to be connected to a higher-order memory self-efficacy construct (e.g., Hertzog et al., 1987). Hence a reasonable conclusion is that the cohort effect seen only for MIA Change appears to be specific to that scale, not to overall memory self-efficacy. It may well be the case that age stereotypes about aging and memory have persisted despite secular changes in what is known about aging, and that these stereotypes have been internalized in similar ways by people of different birth cohorts (Barber, 2017; Hummert, 2011).

In terms of age differences in metamemory, one should not necessarily conclude that older adults accurately monitor their own cognitive status or actual change in cognition over time. Perceptions of memory change in adulthood are also influenced by implicit theories about aging and memory (e.g., Lineweaver & Hertzog, 1998; Ryan & Kwong See, 1990) that may become activated as one grows older (e.g., McDonald-Miszczak, Hertzog, & Hultsch, 1995; McFarland, Ross, & Giltrow, 1992). Hertzog, Hülür, Gerstorf, and Pearman et al. (2018) recently reported longitudinal model results challenging the view that individual differences in actual memory change drive individual differences in perceived memory change. Instead, current perceptions that one’s memory has changed were most closely aligned with current memory self-efficacy.

We also detected period effects in years of education, with increasing levels of educational attainment across the three VLS samples from 1986 to 2000. However, these differences could not account for the period effects in metamemory beliefs, in part because education had relatively low correlations with metamemory beliefs, particularly the Internal Strategy and External Strategy scales that produced period effects. Moreover, those MIA scales that had small but reliable correlations with education (MIA Change, Locus, Anxiety, and Capacity) did not manifest period effects. Societal changes in levels of education do not necessarily generate societal differences in awareness of age-related memory and how it changes during adulthood.

Obviously the present findings cannot be generalized to other measures of memory beliefs or to different time periods and birth cohorts. In particular, we were only able to evaluate period effects over a relatively narrow time period (1986 to 2000) given available VLS data. It could well be the case that period effects would be more prevalent and higher in magnitude if it were possible to trace them back or forward over several additional decades. Nevertheless, the present cohort analyses featured a broad range of birth years (1898–1947) spanning from pre-World War I to post-World War II, a half century of dramatic historical change. Given the benchmark results of the present study, an interesting comparison for future research would be to test metamemory beliefs in samples drawn from more recently born cohorts, including the baby boom generation (1950s and 1960s). It would also be valuable and important to see replication or extension of the effects we report here using other sequential samples, even if with different metamemory questionnaires. Hopefully this paper will motivate other researchers to address these questions in those longitudinal studies that have the relevant data.

Our analyses detected gender differences in several aspects of metamemory beliefs measured by the MIA, largely consistent with the pattern observed in Hultsch et al. (1987). As is generically typical for sex and gender comparisons in psychological constructs, the effects were generally small. Nevertheless, they could reflect gender-related social or cultural differences (Weber et al., 2014). Notably, the effects for the strategy-related scales were more substantial, with women reporting more internal and external strategy use than men. With small mean differences and overlapping distributions, many potentially interesting gender effects may have been missed in studies using smaller samples. For instance, a sample of 100 persons with 60% females has only a .16 probability (power) to reject the null hypothesis for a small population effect size of d = 0.2 in an independent samples t-test. An increase in sample size to 300 increases power to .39. With a sample size of 1500, close to the N of this study, power to detect a 0.2 SD effect size jumps to .97.

Higher ratings of memory ability and control over memory for women may not be surprising, given that women are consistently found to have higher performance on laboratory tests of episodic memory (Herlitz & Rehnman, 2008). It is interesting then that women were also higher in self-rated anxiety about memory, which would tend to be associated with lower, not higher, levels of memory self-efficacy and control (Bandura, 1997). This outcome may reflect the strong relationship of neuroticism and negative affect to self-rated memory seen in the literature (see Hertzog & Pearman, 2014, for a review), which may act to override the gender differences in memory self-efficacy. Women typically manifest higher levels of neuroticism, depression, and negative affect than men, on average, even in adult samples (Fiske, Wetherell, & Gatz, 2009).

In contrast to the mean effects detected by our GLMs, MIA scale correlations were invariant over time. It appears that the period effects observed for the MIA Internal Strategy and External Strategy scales influence only the level of those variables and not the correlations they manifest with other MIA scales. The pattern of MIA scale correlations was consistent with previous reports (Hertzog et al., 1989; Dixon et al., 1988) and also with the Hertzog et al. (1987) factor model creating higher order MSE out of Capacity, Change, Locus, Anxiety. It is useful to note, then, that the differential patterns of mean effects in this study for these scales argue for the differentiability of these MIA scales; they are not simply alternate forms of some higher-order self-efficacy construct. Although they share variance consistent with being indicators of memory self-efficacy, they also show unique patterns of relations to other variables that indicate differentiation between these aspects of metamemory as well.

In sum, we assembled from the VLS archives a relatively large data set representing three separate samples of older adults, each ranging at intake from 55 to 85 years old. The baselines for the samples occurred in three successive decades (1980s, 1990s, and 2000s). The present analyses detected interesting patterns of age, period, and cohort effects in metamemory beliefs—a domain not previously examined in this manner. To our surprise, the results indicated that the MIA scales most closely related to a memory self-efficacy construct – MIA Capacity, Locus, Change, and Anxiety, were not the main source of period or cohort effects. Instead, the strongest period effects appeared to be associated with the strategy-related MIA scales. Given the selectivity and novelty of our results, we recommend that other studies with similar data conduct similar pooled or replication analyses to evaluate cohort and period effects in metamemory beliefs. In addition, although our sample represented a nearly 50-year band of birth years (1898–1947), it would be interesting to extend this approach to more recent birth cohorts such as the late baby boom generation.