NIH Public Access

Author Manuscript
Dev Psychol. Author manuscript; available in PMC 2011 December 5.
Published in final edited form as:
NIH-PA Author Manuscript

Dev Psychol. 2009 March ; 45(2): 431–446. doi:10.1037/a0014012.

The Cognitive Reserve Hypothesis: A Longitudinal Examination

of Age-Associated Declines in Reasoning and Processing Speed
Elliot M. Tucker-Drob1, Kathy E. Johnson2, and Richard N. Jones3
1 Department of Psychology, University of Virginia

2 Department of Psychology, Indiana University-Purdue University Indianapolis

3Institute for Aging Research, Hebrew SeniorLife, Harvard Medical School, Beth Israel
Deaconess Medical Center, Boston, Massachusetts

Abstract
The term cognitive reserve is frequently used to refer to the ubiquitous finding that, during later
life, those higher in experiential resources (e.g. education, knowledge) exhibit higher levels of
NIH-PA Author Manuscript

cognitive function. This observation may be the result of either (1) experiential resources playing
protective roles with respect to the cognitive declines associated with aging or (2) the persistence
of differences in functioning that have existed since earlier adulthood. These possibilities were
examined by applying accelerated longitudinal structural equation (growth curve) models to 5 year
reasoning and speed data from the no-contact control group (N=690, 65 to 89 years of age at
baseline) of the ACTIVE (Advanced Cognitive Training for Independent and Vital Elderly) study.
Vocabulary knowledge and years of education, as markers of cognitive reserve, were related to
levels of cognitive functioning, but unrelated to rates of cognitive change, both before and after
the (negative) relations between levels and rates were controlled for. These results suggest that
cognitive reserve reflects the persistence of earlier differences in cognitive functioning rather than
differential rates of age-associated cognitive declines.

Keywords
cognitive reserve; brain reserve; education; cognitive decline; aging

One particularly important and longstanding topic within the social, behavioral, and
NIH-PA Author Manuscript

cognitive sciences concerns the role that the environment plays with respect to our
interactions, behaviors, and cognitive functioning (Sternberg & Grigorenko, 1997). Levels
of educational attainment, and products of educational achievement such as knowledge and
literacy, are particularly meaningful indices of environmental quality that have well
established relations with cognitive performance throughout the lifespan (Salthouse, 1991).

Much research has focused on childhood development and on determination of the causal
direction of the education-cognition relation. There is a good deal of evidence to suggest that
this relation is reciprocal during childhood. (Ceci, 1996; Crano, Kenny, & Campbell, 1972;
Dickens & Flynn, 2001; c.f. Jensen, 1998). Researchers have also recently begun to
investigate the possibility that having had an enriched environment during earlier parts of
one’s life may play a protective role with respect to the cognitive deficits associated with
adult aging (Satz, 1993; Stern, 2002). Hypotheses addressing the education late-life

Address correspondences regarding this manuscript to: Elliot M. Tucker-Drob, Department of Psychology, University of Virginia, PO
Box 400400, Charlottesville, VA 22904-4400. Phone: 917-670-4641. [email protected].
 Tucker-Drob et al. Page 2

cognition relation have been generally referred to as cognitive reserve hypotheses1. The
major issue that has not yet been resolved with respect to this group of hypotheses, and
which is the focus of this paper, is whether the late-life relations between education and
NIH-PA Author Manuscript

cognitive performance reflect (1) a relation between the quality of earlier life environment
and rates of age-associated cognitive declines or (2) the persistence of education-cognition
relations that have existed since earlier adulthood.

Evidence from Prevalence and Incidence Studies of Dementia

Perhaps the most interesting findings regarding cognitive reserve hypotheses were by
Snowdon et al. (1996) who found that linguistic ability among nuns at a mean age of 22 was
predictive of their cognitive performance and the risk of Alzheimer’s disease approximately
58 years later. A similar study by Whalley, Starr, Athawes, Hunter, Pattie, & Deary (2000),
found that, compared to a matched control group of nondemented individuals, those
suffering from dementia at age 72 or older had scored significantly lower on mental ability
tests at 11 years of age. These findings have been bolstered by studies (e.g. The Canadian
Study of Health and Aging, 1994) reporting higher prevalence rates of dementia in lower
education groups. In fact, a meta-analysis of longitudinal studies, conducted by Valenzuela
and Sachdev (2006), found that reserve, as indexed by variables such as education,
occupation, premorbid IQ, and mental activities was associated with lower risks for incident
dementia.
NIH-PA Author Manuscript

Evidence from Longitudinal Studies of Cognitive Change

A number of longitudinal studies have examined the relations between hypothesized
protective factors and actual cognitive change in both normal and demented adults. Bosma,
van Boxtel, Ponds, Houx, and Jolles (2003), for example, found in a group of 708
individuals ages 50–80, 3 year declines in serial list recall, Stroop-color-word, and Mini
Mental State Examination scores were more shallow for those who reported higher levels of
educational attainment. Lyketsos, Chen, and Anthony (1999) reported similar findings with
respect to a longitudinal study spanning approximately 12 years. In this study, greater
declines (2.5 points compared to 1.4 points) in Mini-Mental State Examination scores were
associated with having 8 or fewer years of education. It is also notable that baseline scores
were systematically associated with magnitude of decline (a baseline score of 30 was
associated with a 1.23 point decrement, whereas a baseline score of 24 or less was
associated with a 2.3 point decrement). Finally, Manly, Touradji, Tang, & Stern (2003) used
literacy as an index of quality of education and demonstrated that the rate of decline in
world list recall performance over approximately 5 years was steeper for low literacy
individuals than for high literacy individuals. Literacy was also positively associated with
NIH-PA Author Manuscript

higher cognitive performance at all assessment occasions.

In seeming contradiction to these findings, Andel, Vigen, Mack, Clark, & Gatz (2006) found
that, among patients with a confirmed Alzheimer’s disease diagnosis, higher education was
related to faster declines in Mini Mental State Examination scores over an average of 2.5
years. Education was positively related to baseline scores. Interestingly, these results were
interpreted to mean that greater reserve is related to the postponement of dementia but also
to accelerated cognitive decline after the onset of dementia symptoms. Similar results have
been found by Stern, Albert, Tang, and Tsai (1999), Teri, McCurry, Edland, Kukull, and
Larso (1995), and Unverzagt, Hui, Farlow, Hall, and Hendrie (1998).

1The cognitive reserve hypothesis is sometimes distinguished from the similar brain reserve hypothesis. The brain reserve hypothesis
is discussed in detail by Christensen, Anstey, Leach, & Mackinnon (2008), who explain that “the hypothesis is that high premorbid
intelligence, education, an active, stimulating lifestyle, or a physically larger brain provide reserve capacity which protects the
individual from the negative effects of aging and disease on brain function.”

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

 Tucker-Drob et al. Page 3

Partially contradictory findings have also been reported by Christensen, Hofer, and their
colleagues (Christensen et al. 2001; Hofer et al., 2002; Mackinnon, Christensen, Hofer,
Korten, & Jorm, 2003), who found evidence for a relation between education and levels of
NIH-PA Author Manuscript

memory, speed, and verbal performance, but no evidence for a relation between education
and 7 year changes in performance. Moreover, in a series of dynamic longitudinal
investigations, Ghisletta and colleagues (Ghisletta & Lindenberger, 2003; Ghisletta &
Lindenberger, 2004; Ghisletta & Ribaupierre, 2005) have demonstrated that changes in
culture-based (crystallized, or pragmatic; e.g. vocabulary knowledge) abilities are predicted
by levels of process (fluid, or mechanic; e.g. processing speed) abilities better than changes
in process abilities are predicted by levels of culture-based abilities. For a comprehensive
review of relevant longitudinal studies published before 2000, see Anstey and Christensen
(2000).

Theories of Cognitive Reserve

A useful standard when comparing theoretical models is to focus on their conflicting
predictions. Among the most salient and testable predictions, in the case of cognitive reserve
models, are those concerning rates of decline in cognitive performance. Models predicting
moderation explain that high reserve people are better able to maintain their levels of
performance relative to low reserve people, who decline more rapidly, resulting in increased
performance differences over time. This prediction is similar to what has been termed
NIH-PA Author Manuscript

differential-preservation with respect to the mental exercise hypothesis (Salthouse, 2006;

Salthouse, Babcock, Skovronek, Mitchell, & Palmon, 1990). Alternatively, there are models
that predict stability of performance differences. These models explain that while high and
low reserve people differ in their levels of cognitive performance, their rates of decline in
performance are comparable. This prediction is similar to what has been termed preserved-
differentiation (Salthouse, 2006; Salthouse et al., 1990). Theories of active and passive
cognitive reserve processes are now introduced and evaluated with respect to their
implications for differential-preservations versus preserved-differentiation.

Stern (2002) has proposed that cognitive reserve mechanisms are active processes through
which, in response to neurobiological degradation, the brain actively attempts to compensate
by using either brain networks or cognitive paradigms that are less susceptible to disruption.
Such active models would predict that, all else being equal, more highly educated people, or
individuals possessing more knowledge, would be able to postpone reaching clinical levels
of impaired cognitive functioning. Salthouse (2003) has proposed five possible mechanisms
by which this might occur: (1) knowledge can enhance memory in the form of richer and
more elaborate encoding and more effective retrieval cues facilitated by a superior
organizational structuring of information; (2) knowledge can result in easier access to
NIH-PA Author Manuscript

relevant information and better organized representations of the problem, resulting in

enhanced problem solving skills; (3) knowledge of past consequences of various alternatives
can provide an effortless means of making accurate predictions regarding future
consequences; (4) knowledge can enable reliance on previously compiled efficient
algorithms, rather than slow and controlled processes; and (5) knowledge of prior solutions
to familiar problems can reduce on-line processing requirements. All of these possible
mechanisms, however, should only result in differential rates of decline (moderation), if they
are increasingly relied upon, or if they are of increasing advantage, with increasing
neurobiological disintegrity. Rather, we argue, there is no strong reason to doubt that these
mechanisms are equally relied upon, or that the mechanisms are equally advantageous, at all
ages and levels of functioning. Therefore, active cognitive reserve models fit the stability, or
preserved differentiation class, and do not necessarily lead to the hypothesis of steeper
cognitive decline among those with less enriched environmental backgrounds.

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

 Tucker-Drob et al. Page 4

Passive models (Satz, 1993; Stern, 2002) maintain that in response to similar levels of
neurobiological degradation, high reserve individuals will experience less impairment than
low reserve individuals. Resilience of nervous system functioning is theoretically indexed
NIH-PA Author Manuscript

by brain size or synapse count, although Satz (1993) has proposed that education may be an
appropriate proxy. Such models often postulate the presence of a neuropathological
threshold that is higher for high reserve people, and only beyond which cognitive
impairment begins to take place. These models therefore predict that the rate of cognitive
decline will be slower, or at least delayed, for high reserve individuals who have not yet
reached their neuropathological threshold, even if their rate of neurobiological degradation
is comparable to that of low reserve individuals. However, this view is problematic for
examining normal age-associated declines because it is very clear that age-related effects on
cognitive performance begin in early adulthood and are continuous rather than abrupt
(Salthouse, 2004). A continuum-based passive model might instead predict that high reserve
individuals respond to neurobiological degradation to a lesser extent than do lower reserve
individuals (i.e. differential preservation), although some passive models (c.f. Figure 3 of
Stern, 2002), are difficult to distinguish from functional threshold models.

According to functional threshold models, high and low reserve individuals experience the
same rate of cognitive decline. However, low reserve individuals begin adulthood with
lower levels of cognitive performance, and therefore take less time to drop below a threshold
beyond which their level of functioning is considered clinically severe or pathological.
NIH-PA Author Manuscript

These models explain that findings linking reserve markers, such as education, with reduced
prevalence and incidence of dementia may be artifacts of arbitrary clinical cutoffs. While
such models have not been discussed in great detail in the cognitive reserve literature, they
are consistent with the diagnostic thresholds discussed by Stern (2002), and the diagnostic
criteria for various dementias described in the Diagnostic and Statistical Manual of Mental
Disorders (IV-TR; American Psychiatric Association, 2000).

Potential Artifacts in Change Research: Level-Slope Relations

Variables purported to index cognitive reserve tend to be correlated with levels of cognitive
functioning. This is, in fact, a necessary requirement of functional threshold models
described above. If, however, relations exist between levels of functioning and rates of age-
associated declines in functioning (i.e. level-slope relations), such phenomena are potentially
critical confounds with respect to moderation hypotheses.

There are both methodological and theoretical reasons to expect level-slope relations to
exist. In cases of poor instrument sensitivity, ceiling effects can prevent detection of changes
within the upper levels of functioning, and floor effects can prevent detection of changes
NIH-PA Author Manuscript

within the lower levels of functioning, thus resulting in spurious relations between initial
performance and change. This is of particular relevance to cognitive reserve research, where
many of the measures used are known to have low measurement ceilings (e.g. the Mini
Mental State Exam-see Anstey & Christensen, 2000; Verbal Learning tests- see Uttl, 2005).
Similar artifacts can occur with respect to the regression to the mean phenomenon
(Campbell & Kenny, 1999; Nesselroade, Stigler, & Baltes, 1980). In such cases, subjects
scoring at the extremes as the result of ordinary statistical noise, are likely to score closer to
their true, less extreme, scores when measured again. This results in spurious negative
correlations between initial level and change. From a theoretical standpoint, if there are
physical or biological limits to the upper and lower ends of functioning, then changes
towards these extremes will be smaller in magnitude than changes towards the mean,
regardless of instrument sensitivity (Ackerman, 2005; the “Law of Initial Values,” Wilder,
1967).

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

 Tucker-Drob et al. Page 5

Method
The current study seeks to test the specificity of the relation between popularly hypothesized
NIH-PA Author Manuscript

cognitive reserve variables (years of educational attainment and vocabulary knowledge) and
levels of reasoning and speed performance (“intercepts”) and rates of longitudinal changes
in reasoning and speed performance (“slopes”) in cognitively normal older adults. In
contrast to most examinations of cognitive reserve moderation hypotheses, the longitudinal
analyses presented in this study examine the degree to which cognitive reserve variables are
related to maturational changes both in conjunction with and independently of initial levels
of performance. Moreover, we examine maturational decline as a component separate from
that associated with the benefits of cumulative test experience (retest effects) associated with
repeated assessments.

Advanced Cognitive Training for Independent and Vital Eldery (ACTIVE)

The ACTIVE study is an ongoing randomized controlled trial conducted at six field sites
with the New England Research Institutes as the coordinating center. The primary goal of
the trial is to test the effects of three distinct cognitive training interventions on cognitive
function and cognitively demanding everyday functioning. This was achieved by random
assignment of individuals to one of three active treatment conditions, or a no-contact control
condition. In these analyses, we only use data from persons enrolled in the control condition.
Persons in the active treatment conditions were omitted. No-contact control data from
NIH-PA Author Manuscript

baseline, 12 week “post test”, first annual, second annual, third annual, and fifth annual
assessments were analyzed. For more detail on design see Jobe, et al. (2001), Ball et al.
(2002), and Willis et al. (2006).

Participants
ACTIVE participants were recruited from a number different settings, registries, and rosters
(e.g. state driver’s license and identification card registries, medical clinical rosters, senior
center and community organization rosters, congregate senior housing sites, local churches,
and rosters of assistance and service programs for the low-income elderly) with the goal of
enrolling a diverse sample of older adults who were living independently and in good
functional and cognitive status, but at risk for loss of functional independence. Recruitment
was focused on six metropolitan areas in the United States. Participants considered for the
current analyses were 704 persons aged 65 to 94 who were randomly assigned to a no
contact control group from a parent sample of 2,832 persons. Six participants were excluded
due to protocol violations.

Table 1 reports the number of participants in each 5 year age group present at each testing
NIH-PA Author Manuscript

occasion. Because only 8 participants were ages 90 and older, they were excluded from
analyses. This exclusion was necessary because of the multiple age group accelerated
longitudinal design implemented in the analysis. This resulted in a working sample of
N=690 (74% female) at baseline assessment with a five year, 6 occasion, retention rate of
54% (77% female). Reasons for dropout included death, withdrawal, site decision, and
unavailability. Scores on the Mini Mental State Exam (Folstein, Folstein, & McHugh, 1975)
ranged from 23 to 30 (the means for each age group ranged from 26.7 to 27.6, with standard
deviations ranging from 1.8 to 2.1; models excluding participants with Mini Mental State
Exam scores less than 25 did not change the patterns of results). Twenty-eight percent of
participants were nonwhite.

Education ranged from 6 to 20 years (mean = 13.4, SD = 2.7). Although negative in

magnitude, the absolute value of the age-education correlation was less than .1, suggesting a
nearly uniform level of selectivity across the age range. Corrections for cohort differences in

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

 Tucker-Drob et al. Page 6

education (c.f. Ronnlund & Nilsson, 2006) produced very similar patterns of results to those
reported in this paper.
NIH-PA Author Manuscript

Figure 1 depicts the distributions of years of education across individuals in each 5 year age
group. It can be seen that all age groups contained individuals ranging from less than a high
school education to greater than a college education, the largest proportion of individuals
having a complete high school or partial college education. While no groups contained large
proportions of participants with very low education (i.e. less than high school), this can be
viewed as a strength, because deprivation of basic schooling is likely to be accompanied by
other extreme environmental deficiencies that could confound results.

The consequences of this sample’s positive selection of well-educated participants were

considered in some detail. In particular it is possible that increments in educational
attainment have diminishing returns, such that the difference between having versus not
having a high school education has much greater cognitive consequences than the difference
between having versus not having a college education. There was some evidence for this, as
a significant quadratic trend (producing the above described pattern) was present in the
regression predicting vocabulary knowledge from years of education. However the size of
this effect was very small (the increment in R2 was .018). Moreover, when an Education2
term was included in the models, and when the models were fit to low and high education
groups separately (split at Education ≤ 12 and Education > 12), the results were very similar
NIH-PA Author Manuscript

to those reported here.

For additional information on sample characteristics and recruiting procedures see Ball et al.
(2002).

Measures
Measures were selected because they were deemed sensitive enough at the upper ranges of
performance to detect potential gains of training interventions, and at the lower ranges of
performance to detect declines associated with late-life aging (Ball et al., 2002).

Reasoning measures required participants to identify patterns in letter or word series

problems. The measures were: Word Series (Gonda & Schaie, 1985), Letter Series
(Thurstone & Thurstone, 1949), and Letter Sets (Ekstrom, French, Harman, & Derman,
1976).

Processing speed measures were timed, and required participants to identify and localize
information at 75% accuracy under varying levels of cognitive demand. They involved three
tasks from the useful Field of View measure (Owsley, Ball, McGwin, et al., 1998).
NIH-PA Author Manuscript

The vocabulary measure (Ekstrom et al., 1976) tested participants’ abilities to choose the
best synonym for a target word from a number of alternatives.

Equally weighted composite test scores representing reasoning and speed were created by
pooling all scores for all participants in the parent population and applying a Blom
transformation (Blom, 1958) to make them more normally distributed, as in Ball et al.
(2002). To facilitate interpretation, reasoning composite, speed composite, and vocabulary
scores were each then standardized with respect to the baseline scores for the youngest age
group using a T score metric (mean = 50, standard deviation = 10). All scores were scaled
such that higher values indicate superior performance (i.e., the Speed composite was
reversed).

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

 Tucker-Drob et al. Page 7

Analyses and Results

In the following section we present analyses and results directly relevant to cognitive reserve
NIH-PA Author Manuscript

hypotheses. To allow for lenient tests of cognitive reserve moderation (differential

preservation) hypotheses, alpha levels were set to .05.

Cross-sectional analyses
If it is the case that higher levels in variables hypothesized to index cognitive reserve, such
as education and knowledge, are related to shallower rates of cognitive decline, then one
should expect cross-sectional differences in cognitive performance between individuals
differing in their levels on those variables to be larger at older ages. This was tested by
examining age × education and age × baseline vocabulary knowledge interactions in
regressions predicting baseline speed and reasoning performance. In all such regressions,
age, education, and vocabulary knowledge were centered by subtracting their means.

Age × education interactions were examined using hierarchical regressions with age and
education entered in the first step, and product of age × education entered in the second step.
Consistent with the preserved differentiation hypothesis but inconsistent with the differential
preservation hypothesis, age and education significantly accounted for both reasoning and
speed performance, but the age × education interaction did not.
NIH-PA Author Manuscript

In examining age × vocabulary knowledge interactions, vocabulary and age were entered in
the first step, and the product of vocabulary × age was entered in the second step, of
hierarchical regressions predicting reasoning and speed. Both regressions resulted in
interaction terms that were statistically significant, but in the direction opposite to that
predicted by cognitive reserve moderation hypotheses. Figure 2 demonstrates these findings,
with higher vocabulary individuals exhibiting larger age differences than lower vocabulary
individuals.

Longitudinal Analyses
Hypotheses concerning age-related longitudinal changes in reasoning and speed abilities
were examined using latent growth curve modeling (LGM) techniques, which are employed
using the structural equation modeling framework. LGM techniques are similar to mixed
effects or hierarchical modeling techniques in that they allow for the estimation of fixed
effects, in the form of population level growth parameters, and random effects, in the form
of individual differences in growth parameters. Individual differences in variables
representing these parameters can then be further examined by regressing them onto
exogenous variables, such as education or vocabulary.
NIH-PA Author Manuscript

For the current analyses, we considered a number of alternative time-based, age-based, and
occasion-based growth curve models (see Appendix for a discussion of the strengths and
weaknesses of each of these models, and their results). Here we report a version of the
multiple group accelerated latent growth curve model (c.f. Chapter 6 of Duncan, Duncan,
Strycker, Li, & Alpert, 1999) allowing for retest effects. This model was chosen because (1)
retest effects can be maximally separated from maturational change by scaling retest effects
by occasions of test experience but scaling maturational change by approximate age at
testing, (2) whether the quality of change, and predictors of change, differ across age groups
can be formally tested, (3) the models can be fit with conventional structural equation
modeling methods, allowing for detailed fit indices, estimation of indirect effects with
standard errors, and, for ease of interpretation, standardized parameter estimates. We note
however, that the substantive conclusions drawn regarding the cognitive reserve hypothesis
were the same for all models considered.

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

 Tucker-Drob et al. Page 8

The multiple group accelerated latent growth curve model is schematically depicted in
Figure 3. In this diagram, squares represent observed variables, such as scores at each
measurement occasion, and circles represent latent, or unobserved variables, such as those
NIH-PA Author Manuscript

corresponding to maturational slope (s), intercept (i), and retest effects (r). Regression
coefficients are represented as one-headed arrows, and the variance terms of variables, or
their residuals, are represented as two-headed arrows attached to the specific variables. The
unit constant (allowing for the estimation of means) is represented as a triangle.

The model depicted in Figure 3 fits a similar 6 occasion model to each age group (5 year
cohort) individually, and across-group equality constraints are placed on the latent means,
latent variances, covariance and regression relations, and residual variances. Some of these
constraints can then be removed and the resulting chi-square and degrees of freedom change
documented to test for between-group differences. By parameterizing the basis coefficients
for the latent slope (s) to reflect cohort age (centered at 65), mean age related change in
cognitive functioning is assumed to be linear and reflected by cross-sectional age differences
(the convergence assumption; Bell, 1953), while still allowing for individual differences in
change to be determined by longitudinal information from repeated measurements. The
basis coefficients defining the shape of the retest curve (r) can then be freely estimated from
the data. The basis coefficients defining the growth curve intercept (i) are all fixed to 1,
allowing the intercept to be interpreted as performance at 65 years of age. Residual
variances are all constrained to be equal, reflecting the assumption of homoscedasticity over
NIH-PA Author Manuscript

time.

The growth curve portion of the model depicted in Figure 3 can be expressed as

(1)

or in expanded form as
NIH-PA Author Manuscript

where g indicates that a variable is group specific, and n indicates that a variable varies
across individuals. AgeBL,g represents the designated age (in years) of age group g at
baseline assessment (Age[o]g therefore corresponds to the designated age of the group at any
given occasion, centered at 65). Equation 1 explains that individual scores at each
measurement occasion Y[o]n can be fully accounted for by individual differences in growth
curve intercept (in), maturational slope (sn), retest effects (rn), and unexplained variance
(u[o]n).

Growth curve intercept, slope, and retest effects are regressed onto education and baseline
vocabulary knowledge (education and vocabulary were both centered at their means for each
age group, such that μE=0 and υV=0) according to the following equations:

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

 Tucker-Drob et al. Page 9

(2)
NIH-PA Author Manuscript

(3)

(4)

where υ’s represent regression intercepts, β’s represent regression coefficients, and e’s
represent residuals. These equations allow for identification of relations between individual
growth curve parameters and hypothesized cognitive reserve variables. Note that growth
curve slope and retest effects are regressed onto the growth curve intercept to determine the
degree to which relations are mediated through initial performance, and the degree to which
relations are independent of initial performance. Finally, vocabulary is regressed onto
education according to:

(5)
NIH-PA Author Manuscript

under the assumption that vocabulary knowledge is, in part, the product of education. To
account for the possibility that, as a result of social inequalities, years of education may have
different meaning for men and women, and whites and non-whites (Manly et al., 2003),
models were considered in which variables representing gender and race were included as
covariates in Equations 2 through 5. This addition did not change the overall pattern of
results, or the substantive interpretations drawn. We therefore report the simpler models.

All LGM analyses were conducted using the Mplus software package (Muthén & Muthén,
1998–2006) with Full Information Maximum Likelihood (FIML) estimation methods. FIML
estimates parameter values based on all available data, under the Missing at Random (MAR)
assumption that any systematic patterns of missingness are functions of the variables
included in the model, and that the patterns of and predictors of growth and change are the
same for complete and incomplete subgroups. Longitudinal growth curve methods
employing the MAR assumption to handle selective attrition, have been successfully applied
to simulated data (e.g. McArdle & Hamagami, 1992) and are regularly applied to real data
(e.g. Ferrer, Salthouse, McArdle, Stewart, & Schwartz, 2005; McArdle, Ferrer-Caja,
Hamagami, & Woodcock, 2002). In the current study, attrition analyses indicated that those
NIH-PA Author Manuscript

who were lower performing on cognitive measures at baseline were more likely to drop out
of the study, but that neither education nor vocabulary knowledge was related to study
dropout. Moreover, models fit to data from participants with complete data on all six
measurement occasions produced similar patterns of results to those reported here.
Therefore, while the MAR assumption remains untestable, these follow-up analyses suggest
that the substantive findings are robust.

Table 2 reports fit statistics for the unconditional reasoning and speed growth curve models
with the mean retest effect and retest curve shape constrained to be equal across age groups,
the mean retest effect allowed to differ across age groups, and the mean retest effect and
retest curve shape allowed to differ across age groups. Because these models were nested
within each other, chi square difference tests were used to select models. It can be seen that
the preferred reasoning model was one in which the mean and shape of the retest curve was
allowed to differ among the age groups, whereas the preferred speed model was one in

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

 Tucker-Drob et al. Page 10

which the shape of the retest curve was invariant across age groups, but the mean retest
effect was allowed to differ across age groups.
NIH-PA Author Manuscript

Figures 4 and 5 depict maturational components, retest components, and net trajectory of the
population average growth curves implied by the reasoning and speed models respectively.
It can be seen that, for both reasoning and speed, the largest retest benefit occurs during the
first retesting, and the younger groups tend to benefit more than the older groups. Moreover,
maturational age gradients for speed (−.70 T score units per year, p<.05) were considerably
steeper (pdifference<.05) than for reasoning (−.49 T score units per year, p<.05).

Table 3 reports the fit statistics for the reasoning and speed growth curve models (the best
fitting reasoning and speed models selected from Table 2) with the relations (βEs, βEr, βvs,
βvr from Equations 2 through 5, and Figure 3) between hypothesized cognitive reserve
variables and latent change components held constant across age groups (fully constrained
models), and the change in fit after freeing these relations across groups (predictors freed
models). All other parameters in Equations 2 through 5 were constrained to be constant
across groups in both sets of models. The fit statistics indicate well fitting models, with all
RMSEAs less than .07, and CFIs all greater than .96. Because these fully constrained
models did not fit significantly worse (by nested chi square comparison) than the less
parsimonious predictors freed models, we accept them as the best representations of 5 year
changes in reasoning and speed performance, and their relations with education and
NIH-PA Author Manuscript

vocabulary. The results of these models are described below.

Tables 4 and 5 report the relations between hypothesized cognitive reserve variables and
growth curve variables for the reasoning and speed models respectively. Unstandardized
parameter estimates and, for ease of interpretation, standardized parameter estimates are
reported. Total effects (the sum of unique and mediated effects), direct (unique) effects, and
indirect (mediated) effects are reported. This allows for examinations of the relations
between cognitive reserve markers (education and vocabulary) and maturational changes
both with and without controlling for the intercepts, as well as the relation that is completely
attributable to the intercepts.

A number of observations are of note. First, the intercepts (i) were highly related to both
hypothesized cognitive reserve variables (education and vocabulary), but most of the
education effect was mediated through vocabulary. Second, maturational slopes (s) were not
significantly related to hypothesized cognitive reserve variables, whether or not the
significant influences of the intercepts were controlled for. Third, there were significant
negative relations between intercepts and maturational slopes, such that individuals at higher
levels of functioning declined more steeply. Finally, for speed, the retest component was
NIH-PA Author Manuscript

positively related to the intercept, but there were no significant direct or total relations
between the retest component and hypothesized cognitive reserve variables. For reasoning
there were no significant relations between the retest component and intercepts or the retest
component and hypothesized cognitive reserve variables.

Evidence for the Reliability of Change

Finally, we examined whether our failures to detect relations between hypothesized reserve
variables and longitudinal cognitive changes (maturational slopes) resulted from failures to
reliably assess these changes. The reliable assessment of change has been a longstanding
issue in aging research (for a discussion see Hertzog & Nesselroade, 2003). Hertzog,
Lindenberger, Ghisletta, and von Oertzen (2006) have recently called to attention that the
extent to which growth curve modeling has the capability to detect correlates of change is
still largely unknown, and arguing for the null hypothesis may therefore be problematic.
They explain that because growth curve slopes have the most potential to lack reliability, it

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

 Tucker-Drob et al. Page 11

is particularly difficult to identify correlates of slopes. Following this logic in reverse, we

reason that if we are able to detect moderate to large correlations between changes in
reasoning and speed, the latent slopes are likely to be sufficiently reliable (and variable) to
NIH-PA Author Manuscript

detect relations with hypothesized reserve variables.

To examine coupled change, the unconditional reasoning and speed models used in the
previous analyses were combined into a multiple-group bivariate latent growth curve model
(McArdle & Nesselroade, 2003). To maintain consistency with the models reported in tables
4 and 5, the basis coefficients defining the shape of the retest effects, the variances of the
intercepts, and the residual variances of the reasoning and speed scores for each occasion,
were fixed to the values that had been estimated with those models. To control for the
possibility that latent slopes may be related to each other by way of their relations to the
intercepts, latent slopes were regressed onto both latent intercepts, and their residuals
allowed to covary. Similarly, retest components were regressed onto latent intercepts, and
their residuals allowed to covary. The covariance between the latent intercepts was also
estimated. To maintain further consistency with the models previously reported, variance,
covariance, and regression parameters were constrained to be equal across age groups.

Parameters for the bivariate growth curve model are presented in Table 6. Consistent with
the positive manifold (Spearman, 1927) intercepts were strongly related to each other (r = .
626). Moreover, even after controlling for the influences of the latent intercepts, highly
NIH-PA Author Manuscript

significant relations were found between latent maturational slopes (r = .636), but not latent
retest components. We therefore conclude that the latent slopes were modeled with
sufficient reliability to have detected relations with hypothesized cognitive reserve variables,
had they been present. Our failures to detect predictors of retest components, however, may
be attributable to failures to reliably assess them (or lack of sufficient individual differences
in these components).

Discussion
Summary and Discussion of Findings
In this study, we examined the hypotheses that education and vocabulary knowledge, as
markers of cognitive reserve, are (1) associated with higher levels of functioning in old age,
and (2) associated with shallower rates of age-associated declines in functioning, for
variables indexing reasoning and speed. We found evidence for the former hypothesis, but
not for the latter.

Contrary to cognitive reserve theories predicting moderation or differential preservation,

both with and without accounting for the mediation of the intercepts, education and
NIH-PA Author Manuscript

vocabulary knowledge were unrelated to rates of maturational decline. These findings

suggest that the late life relations between education and vocabulary, as markers of cognitive
reserve, and cognitive functioning reflect the persistence of earlier-life differences in
cognitive functioning, and not differential rates of cognitive decline.

Processing speed, compared to reasoning, showed smaller level relations with education and
vocabulary, and steeper age gradients, suggesting that processing speed may be a cognitive
primitive that is less related to skill or acquired knowledge, and more reflective of age-
associated neurobiological degradation (c.f. Salthouse, 1996). If it is the case that reasoning
is more environmentally influenced than speed, and therefore potentially more amenable to
protective cognitive reserve processes, one might have expected rates of change in reasoning
to show greater positive relations to hypothesized cognitive reserve markers. We found no
evidence for this, as neither reasoning nor speed changes were significantly associated with
vocabulary or education in the predicted directions.

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

 Tucker-Drob et al. Page 12

The relations between education and levels of cognitive performance were substantially
mediated by vocabulary knowledge, suggesting that the benefits of education may be best
indexed by knowledge measures, for which confounds concerning quality (or “yearly
NIH-PA Author Manuscript

value”) of education are not of issue (c.f. Jones, 2003).

Based upon these findings we conclude that cognitive reserve models predicting stability,
rather than moderation, of performance differences are most plausible. In cases of active
cognitive reserve models, this would mean that, rather than increasing reliance on
knowledge-based paradigms and brain networks with age, individuals high in knowledge
and education rely upon these advantages consistently over their late adult lives. Contrary to
the predictions of passive cognitive reserve models, these findings suggest that individuals
high in knowledge and education exhibit similar degrees of age-associated cognitive
decrements as those less knowledgeable and educated.

Consistent with functional threshold predictions, cognitive reserve variables were found to
be related to levels of reasoning and speed performance, but unrelated to rates of decline in
reasoning and speed performance. Prevalence and incidence studies supportive of cognitive
reserve may therefore simply be artifacts of arbitrary clinical cutoffs for pathological
functioning and dementia. This is because, on average, more highly educated individuals
begin adulthood with higher levels of cognitive functioning. Therefore, they take longer to
reach clinically significant levels of functioning compared to lower educated individuals
NIH-PA Author Manuscript

whose abilities decline at similar average rates.

Previous studies supportive of cognitive reserve moderation hypotheses may be results of

poor psychometric properties of the cognitive measures used, or inappropriate use of the
instruments, such as those with pronounced ceiling effects. While the methods often used
assume interval measurement, this is often not the case, particularly at the extremes of the
scales (Embretson & Reise, 2000). In the current study, the instruments used were sensitive
to a wide range of abilities, and latent intercepts were considered as potential mediators of
latent slope relations. In fact, previous investigations (e.g. Christensen et al. 2001; Hofer et
al., 2002; Mackinnon et al., 2003) using sensitive measures and sophisticated modeling
techniques have also failed to find associates between cognitive reserve markers and rates of
change.

Limitations
It is important to acknowledge that this study included participants who were, on the whole,
at healthy levels of cognitive functioning. Our findings may therefore only be relevant to the
patterns of aging that precede senile dementia, but not the patterns ensuing from dementia.
Moreover, although there was little evidence that results qualitatively differed for low and
NIH-PA Author Manuscript

high education participants (see Participants section), very few participants with severely
impoverished educational backgrounds (e.g. less than 8 years of education) were included in
the sample, and we are therefore unable to make strong generalizations to such populations.
It is also possible that cognitive reserve mechanisms may act differently on different abilities
(e.g. memory), and we must therefore limit our conclusions to levels and rates of change in
reasoning and processing speed, the cognitive outcomes measured in this study. Finally, it is
important to keep in mind that, while popularly examined, educational attainment and
measures of acquired knowledge may only be surrogate markers for theoretical constructs
hypothesized to protect against cognitive decline. For instance it is possible that, rather than
being related to earlier-life environmental quality, those factors that truly serve to mitigate
late-life cognitive changes themselves emerge during late-life. It might therefore be
desirable for future examinations of potential moderators of cognitive aging to examine
hypothesized cognitive reserve markers that are measured at closer temporal proximities to
the senescent changes of interest.

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

 Tucker-Drob et al. Page 13

Conclusions
The current study yielded considerable evidence unsupportive of the notion that formal
education during earlier life is related to rates of decline in cognitive functioning during later
NIH-PA Author Manuscript

life. However, levels of cognitive performance were substantially related to levels of

educational attainment, as well as to vocabulary knowledge, a likely product of educational
achievement. Therefore, to the extent that education causally influences cognitive abilities
during childhood development (Ceci, 1996), these benefits seem to persist throughout the
lifespan and until late adulthood (Deary, Whalley, Lemmon, Crawford, & Starr, 2000). The
persistence of such benefits may also serve to protect against functional impairment (as
described by functional threshold models), and therefore have substantial implications for
everyday functioning in later life.

These findings argue for an integrated approach to developmental psychology that continues
to emphasize both early life and late life phenomena (Baltes, 1987; Craik & Bialystok,
2006). Developing a fuller understanding of the education-cognition relation in late life
certainly requires further untangling of the causal components of the education-cognition
relation during childhood, and determination of the effects of continuing education and
cognitive training (e.g. Willis et al., 2006) during adulthood. Moreover, research focusing on
lifespan longitudinal data (e.g. Deary et al., 2000; McArdle et al., 2002) will continue to
prove useful in better understanding and characterizing the processes of change that occur
within individuals as they grow and mature.
NIH-PA Author Manuscript

Acknowledgments
ACTIVE is supported by grants (U01AG14260; U01AG14263; U01AG14276; U01AG14282; U01AG14289;
U01NR04507; U01NR04508) from the National Institute on Aging and the National Institute of Nursing Research.
Elliot M. Tucker-Drob was supported as a trainee by a grant (T32AG020500) from the National Institute on Aging,
and Richard N. Jones was supported by the Harvard Older Americans Independence Center grant (P60AG008812)
from the National Institute on Aging. We would like to acknowledge the valuable suggestions by Timothy A.
Salthouse, John R. Nesselroade, and those at the Center for Developmental and Health Research Methodology at
the University of Virginia.

References
Ackerman, PL. Ability determinants of individual differences in skilled performance. In: Sternberg,
RJ.; Pretz, J., editors. Cognition and Intelligence: Identifying the Mechanisms of the Mind. NY:
Cambridge University Press; 2005. p. 142-159.To appear
American Psychiatric Association. Diagnostic and statistical manual of mental disorders. 4.
Washington, D.C: Author; 2000. text revision
Andel R, Vigen C, Mack WJ, Clark L, Gatz M. The effect of education and occupational complexity
NIH-PA Author Manuscript

on rate of cognitive decline in Alzheimer’s patients. Journal of the International Neuropsychological

Society. 2006; 12:147–152. [PubMed: 16433954]
Anstey K, Christensen H. Education, activity, health, blood pressure and apolipoprotein E as predictors
of cognitive change in older age: A review. Gerontology. 2000; 46:163–177. [PubMed: 10754375]
Ball K, Berch DB, Helmers KF, Jobe JB, Leveck MD, Marsiske M, Morris JN, Rebok GW, Smith
DM, Tennstedt SL, Unverzagt FW, Willis SL. Advanced Cognitive Training for Independent and
Vital Elderly Study Group (ACTIVE): Effects of cognitive training interventions with older adults:
A randomized controlled trial. Journal of the American Medical Association. 2002; 288:2271–2281.
[PubMed: 12425704]
Baltes PB. Theoretical propositions of life-span developmental psychology: On the dynamics between
growth and decline. Developmental Psychology. 1987; 23:611–626.
Bell RQ. Convergence: An accelerated longitudinal approach. Child Development. 1953; 24:145–152.
[PubMed: 13141335]
Blom, G. Statistical Estimates and Transformed Beta Variables. New York: Wiley; 1958.

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

 Tucker-Drob et al. Page 14

Bosma H, Boxtel MPJ, Ponds RWHM, Houx PJH, Jolles J. Education and age-related cognitive
decline: The contribution of mental workload. Educational Gerontology. 2003; 29:165–173.
Campbell, DT.; Kenny, DA. A Primer on Regression Artifacts. New York: Guilford; 1999.
NIH-PA Author Manuscript

Ceci, SJ. On Intelligence: A bio-ecological treatise on intellectual development. 2. Cambridge, MA:

Harvard University Press; 1996.
Christensen, H.; Anstey, KJ.; Leach, LS.; Mackinnon, AJ. Intelligence, education, and the brain
reserve hypothesis. In: Craik, FIM.; Salthouse, TA., editors. The handbook of aging and cognition.
3. Hillsdale, N.J: Lawrence Erlbaum Associates; 2008.
Christensen H, Hofer SM, Mackinnon AJ, Korten AE, Jorm AF, Henderson AS. Age is no kinder to
the better educated: Absence of an association established using latent growth techniques in a
community sample. Psychological Medicine. 2001; 31:15–27. [PubMed: 11200953]
Craik, FIM.; Bialystok, E. On structure and process in lifespan cognitive development. In: Bialystock,
E.; Craik, FIM., editors. Lifespan Cognition: Mechanisms of Change. Oxford University Press;
New York: 2006.
Crano WD, Kenny DA, Campbell DT. Does intelligence cause achievement? A cross-lagged panel
analysis. Journal of Educational Psychology. 1972; 63:258–275.
Deary IA, Whalley LJ, Lemmon H, Crawford JR, Starr JM. The stability of individual differences in
mental ability from childhood to old age: Follow-up of the 1932 Scottish Mental Survey.
Intelligence. 2000; 28:49–55.
Dickens WT, Flynn JR. Heritability estimates versus large environmental effects: The IQ paradox
resolved. Psychological Review. 2001; 108:356–369.
NIH-PA Author Manuscript

Duncan, TE.; Duncan, SC.; Strycker, LA.; Li, F.; Alpert, A. An introduction to latent variable growth
curve modeling. Mahwah, NJ: Erlbaum; 1999.
Ekstrom, RB.; French, JW.; Harman, H.; Derman, D. Kit of Factor-Referenced Cognitive Tests rev.
Princeton, NJ: Educational Testing Service; 1976.
Embretson, SE.; Reise, S. Item response theory for psychologists. Mahwah, NJ: Lawrence Erlbaum
Associates, Inc; 2000.
Ferrer E, Salthouse TA, McArdle JJ, Stewart WF, Schwartz BS. Multivariate modeling of age and
retest in longitudinal studies of cognitive abilities. Psychology and Aging. 2005; 20:412–422.
[PubMed: 16248701]
Folstein MF, Folstein SE, McHugh PR. ‘Mini-mental state.’ A practical method for grading the
cognitive state of patients for the clinician. Journal of Psychiatric Research. 1975; 12(3):189–198.
[PubMed: 1202204]
Ghisletta P, Lindenberger U. Age-based structural dynamics between perceptual speed and knowledge
in the Berlin Aging Study: Direct evidence for ability dedifferentiation in old age. Psychology and
Aging. 2003; 18:696–713. [PubMed: 14692858]
Ghisletta P, Lindenberger U. Static and dynamic longitudinal structural analyses of cognitive changes
in old age. Gerontology. 2004; 50:12–16. [PubMed: 14654721]
Ghisletta P, de Ribaupierre A. A dynamic investigation of cognitive dedifferentiation with control for
NIH-PA Author Manuscript

retest: Evidence from the Swiss Interdisciplinary Longitudinal Study on the Oldest Old.
Psychology and Aging. 2005; 20:671–682. [PubMed: 16420141]
Gonda, J.; Schaie, KW. Schaie-Thurstone Mental Abilities Test: Word Series Test. Palo Alto, CA:
Consulting Psychologists Press; 1985.
Hofer SM, Christensen H, MacKinnon AJ, Korten AE, Jorm AF, Henderson AS, Easteal S. Change in
cognitive functioning associated with apoE genotype in a community sample of older adults.
Psychology and Aging. 2002; 17:194–208. [PubMed: 12061406]
Hertzog C, Nesselroade JR. Assessing psychological change in adulthood: An overview of
methodological issues. Psychology and Aging. 2003
Hertzog C, Lindenberger U, Ghisletta P, Oertzen Tv. On the power of multivariate latent growth curve
models to detect correlated change. Psychological Methods. 2006; 11(3):244–252. [PubMed:
16953703]
Jensen, AR. The g factor: The science of mental ability. Westport, CT: Praeger/Greenwood; 1998.

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

 Tucker-Drob et al. Page 15

Jobe JB, Smith DM, Ball K, Tennstedt SL, Marsiske M, Willis SL, Rebok GW, Morris JN, Helmers
KF, Leveck MD, Kleinman K. ACTIVE: A cognitive intervention trial to promote independence
in older adults. Controlled clinical trials. 2001; 22:453–479. [PubMed: 11514044]
NIH-PA Author Manuscript

Jones RN. Racial bias in the assessment of cognitive functioning of older adults. Aging and Mental
Health. 2003; 7:83–102. [PubMed: 12745387]
Lyketsos CG, Chen LS, Anthony JC. Cognitive Decline in Adulthood: An 11.5-year follow up of the
Baltimore epidemiologic catchment area study. American Journal of Psychiatry. 1999; 156:58–65.
[PubMed: 9892298]
Mackinnon A, Christensen H, Hofer SM, Korten A, Jorm AF. Use it and still lose it? The association
between activity and cognitive performance established using latent growth techniques in a
community sample. Aging, Neuropsychology, and Cognition. 2003; 10:215–229.
Manly JJ, Touradji P, Tang MX, Stern Y. Literacy and memory decline among ethnically diverse
elders. Journal of Clinical and Experimental Neuropsychology. 2003; 25:680–690. [PubMed:
12815505]
McArdle JJ, Ferrer-Caja E, Hamagami F, Woodcock RW. Comparative longitudinal structural
analyses of growth and decline of multiple intellectual abilities over the lifespan. Developmental
Psychology. 2002; 38:115–142. [PubMed: 11806695]
McArdle JJ, Hamagami E. Modeling incomplete longitudinal and cross-sectional data using latent
growth structural models. Experimental Aging Research. 1992; 18:145–166. [PubMed: 1459161]
McArdle, JJ.; Nesselroade, JR. Growth curve analyses in contemporary psychological research. In:
Schinka, J.; Velicer, W., editors. Comprehensive Handbook of Psychology, Volume Two:
Research Methods in Psychology. New York: Pergamon Press; 2003. p. 447-480.
NIH-PA Author Manuscript

McArdle JJ, Woodcock JR. Expanding test-rest designs to include developmental time-lag
components. Psychological Methods. 1997; 2:403–435.
Muthén, LK.; Muthén, BO. Mplus user’s guide. 4. Los Angeles,CA: Muthén & Muthén; 1998–2006.
Nesselroade JR, Stigler SM, Baltes PB. Regression toward the mean and the study of change.
Psychological Bulletin. 1980; 88:622–637.
Owsley C, Ball K, McGwin G, et al. Vision impairment, eye disease, and injurious motor vehicle
crashes in the elderly. Opthalmic Epidemiology. 1998; 5:101–113.
Ronnlund M, Nilsson LG. Adult life-span patterns in WAIS-R Block Design performance: Cross-
sectional versus longitudinal age gradients and relations to demographic factors. Intelligence.
2006; 34:63–78.
Salthouse, TA. Theoretical Perspectives on Cognitive Aging. Hillsdale, NJ: Lawrence Erlbaum
Associates; 1991.
Salthouse TA. The processing-speed theory of adult age differences in cognition. Psychological
Review. 1996; 103:403–428. [PubMed: 8759042]
Salthouse, TA. Interrelations of aging, knowledge, and cognitive performance. In: Staudinger, U.;
Lindenberger, U., editors. Understanding Human Development: Dialogues with Lifespan
Psychology. Berlin: Kluwer Academic Publishers; 2003.
NIH-PA Author Manuscript

Salthouse TA. What and when of cognitive aging. Current Directions in Psychological Science. 2004;
13:140–144.
Salthouse TA. Mental Exercise and mental aging: Evaluating the validity of the use it or lose it
hypothesis. Perspectives on Psychological Science. 2006; 1:68–87.
Salthouse TA, Babcock RL, Skovronek E, Mitchell DRD, Palmon R. Age and experience effects in
spatial visualization. Developmental Psychology. 1990; 26:128–136.
Satz P. Brain reserve capacity on symptom onset after brain injury: A formulation and review of
evidence for threshold theory. Neuropsychology. 1993; 7:273–295.
Snowdon DA, Kemper SJ, Mortimer JA, Greiner LH, Wekstein DR, Markesbery WR. Linguistic
ability in early life and cognitive function and Alzheimer’s disease in late life. Findings from the
Nun Study. Journal of the American Medical Association. 1996; 21:528–532. [PubMed: 8606473]
Spearman, C. The abilities of man: Their nature and measurement. New York: Macmillan; 1927.
Stern Y. What is cognitive reserve? Theory and research application of the reserve concept. Journal of
the International Neuropsychological Society. 2002; 8:448–460. [PubMed: 11939702]

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

 Tucker-Drob et al. Page 16

Stern Y, Albert S, Tang MX, Tsai WY. Rate of memory decline in AD is related to education and
occupation: Cognitive reserve? Neurology. 1999; 53:1942–1947. [PubMed: 10599762]
Sternberg, RJ.; Gringorenko, E. Intelligence, Heredity, and Environment. New York: Cambridge
NIH-PA Author Manuscript

University Press; 1997.

Teri L, McCurry SM, Edland SD, Kukull WA, Larson EB. Cognitive decline in Alzheimer’s disease:
A longitudinal investigation of risk factors for accelerated decline. Journal of Gerontology:
Medical Sciences. 1995; 50A:M49–M55.
The Canadian Study of Health and Aging. Risk factors for Alzheimer’s disease in Canada. Neurology.
1994; 44:2073–2080. [PubMed: 7969962]
Thurstone, LL.; Thurstone, TG. Examiner Manual for the SRA Primary Mental Abilities Test (Form
10–14). Chicago: Science Research Associates; 1949.
Unverzagt FW, hui SL, Farlow MR, Hall KS, Hendrie HC. Cognitive decline and education in mild
dementia. Neurology. 1998; 50:181–185. [PubMed: 9443477]
Uttl B. Measurement of individual differences: Lessons from memory assessment in research and
clinical practice. Psychological Science. 2005; 16:460–467. [PubMed: 15943672]
Valenzuela MJ, Sachdev P. Brain reserve and dementia: a systematic review. Psychological Medicine.
2006; 36:441–454. [PubMed: 16207391]
Whalley LJ, Starr JM, Athawes R, Hunter D, Pattie A, Deary IJ. Childhood mental ability and
dementia. Neurology. 2000; 55:1455–1459. [PubMed: 11094097]
Wilder, J. The law of initial values. Bristol, England: John Wright & Sons; 1967. Stimulus and
response.
NIH-PA Author Manuscript

Willis SL, Tennstedt SL, Marskiske M, Ball K, Elias J, Koepke KM, Morris JN, Rebok GW,
Unverzagt FW, Stoddard AM, Wright E. Long-term effects of Cognitive Training on Everyday
Functional Outcomes in Older Adults. Journal of the American Medical Association. 2006;
296:2806–2814.

Appendix: Alternative Growth Curve Models

A number of growth curve models were considered as alternatives to those described by
Equation 1. Here we present these models and discuss their respective strengths and
weaknesses.

One set of models considered assumes that rates of change are person-specific functions of
the time since baseline evaluation, and/or the number of occasions of measurement. The
simplest of these models can be written as

(A1)
NIH-PA Author Manuscript

where Y[o] n is the observed score of person n at occasion o, in is the growth curve intercept
of person n, Time[o] is the time (in years) that has elapsed since the baseline assessment (i.e.
0, .23, 1.23, 2.23, 3.23, 5.23), sn is the growth curve slope of person n, and u[o]n is the
unexplained component for person n at occasion o. This model assumes only one source of
variance in change, and therefore confounds change that may be due to age-related
maturation, and change that may be due to the experience of repeated testing. Change is
assumed to occur linearly, and therefore any potential nonlinear change in cognitive
performance over time is not captured. Moreover, because performance is a function of time
since baseline evaluation, rather than age at testing, the model fails to capitalize on age at
testing as a source of information (for a discussion see McArdle et al., 2002).

Assumptions about linear change in the above model can be relaxed by allowing the shape
of the growth curve to be produced by the data. Such a growth curve can be written as

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

 Tucker-Drob et al. Page 17

(A2)
NIH-PA Author Manuscript

where rn represents nonlinear change, scaled by occasion (i.e. B[o] = 0, 1, B[1], B[2], B[3],
B[5]). Again, only one source of variation in change is assumed and age-related information
is neglected.

An elaboration of the above two models is one that includes allowances for both linear and
nonlinear sources of change. This model can be written as

(A3)

We again label the person-specific linear component, scaled by time since baseline
evaluation, as sn, and the person-specific nonlinear component, scaled by occasion, as rn.
However, because age-related information is still neglected, and sn and rn are therefore on
similar scales, it may not be justified to conceptualize these sn and rn components as
individually representative of maturation and retest effects respectively.

A second set of models considered assumes that rates of change are person-specific
NIH-PA Author Manuscript

functions of age at testing, and/or the number of occasions of measurement. A simple form
of such a model can be written as

(A4)

where Age[o]n corresponds to the age of the individual at any given occasion, centered at
65. As with the time-based linear model, this model does not account for nonlinear change.
However, it is able to incorporate the age at testing as an additional source of information,
under the assumption that age-related differences are informative about age-related changes
(Bell, 1953).

Maturational components and retest components can be considerably separated by

combining the models described by Equations A2 and A4 to one in which maturation is
scaled based on the age of the individual, but retest effects are scaled according to occasion
of measurement (e.g. Ferrer et al., 2005; McArdle et al., 2002). This model can be written as
NIH-PA Author Manuscript

(A5)

This model is very similar to that described by Equation 1, with the difference being that the
age basis is specific to the individual rather than the cohort (i.e. subscript g is removed from
Equation 1), and no cohort differences in the shape or magnitude of the retest effects are
permitted.

Results
Tables A1 and A2 report fit statistics and parameter estimates for the above described
models fit to the reasoning and speed data, along with the conditional relations described by
Equations 2 through 5. These output are also provided for the model described by Equation
1, with the shape and mean of the retest component, mean maturational component, mean

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

 Tucker-Drob et al. Page 18

intercept, conditional relations, and all residual variance terms constrained to be equal across
groups.
NIH-PA Author Manuscript

A number of observations are of note. First because the time and occasion basis models are
nested within one another, as are the age and occasion basis models, we can directly
compare their fits. In both cases, the models allowing for both linear and nonlinear sources
of change fit the data best. Second, the time + occasion basis models produce curves (i.e.
basis coefficients) that are difficult to interpret as distinctly representative of retest and
maturation components, as the nonlinear components peak at the second assessment but
subsequently decline (which is uncharacteristic of retest components, which are likely to
accumulate with experience and decay more slowly with time), and the linear components
both have non-significant positive means (inconsistent with maturational decline).
Alternatively, the age + occasion basis models are more readily interpretable as retest and
maturation components, as the nonlinear components have positive means and
approximately asymptotic shapes, and the linear components have negative means. Finally,
it is of particular note that for all models, the relations between hypothesized cognitive
reserve variables and intercept and change components are very similar, overwhelmingly
inconsistent with moderation hypotheses but consistent with stability hypotheses.
NIH-PA Author Manuscript
NIH-PA Author Manuscript

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

 Tucker-Drob et al. Page 19
NIH-PA Author Manuscript
NIH-PA Author Manuscript

Figure 1.
Distribution of education by age group.
NIH-PA Author Manuscript

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

 Tucker-Drob et al. Page 20
NIH-PA Author Manuscript

Figure 2.
Left: Plot of Age × Vocabulary interaction predicting cross-sectional reasoning
performance. Right: Plot of Age × Vocabulary interaction predicting cross-sectional speed
performance. In both plots high and low vocabulary are at one standard deviation above the
mean, and one standard deviation below the mean, respectively.
NIH-PA Author Manuscript
NIH-PA Author Manuscript

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

 Tucker-Drob et al. Page 21
NIH-PA Author Manuscript
NIH-PA Author Manuscript

Figure 3.
Path diagram depicting the multiple group accelerated latent growth curve model employed.
BL, PT, Y1, Y2, Y3, and Y5 represent vocabulary or speed performance at baseline, 12
week post test, first annual, second annual, third annual, and fifth annual assessments
respectively. Beneath these labels are the designated ages of each age group at each
occasion. See text for further description.
NIH-PA Author Manuscript

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

 Tucker-Drob et al. Page 22
NIH-PA Author Manuscript
NIH-PA Author Manuscript

Figure 4.
Model implied population components of 5 year longitudinal change in reasoning
performance.
NIH-PA Author Manuscript

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

 Tucker-Drob et al. Page 23
NIH-PA Author Manuscript
NIH-PA Author Manuscript

Figure 5.
Model implied components of 5 year longitudinal change in speed performance.
NIH-PA Author Manuscript

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

 NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript

Table 1
Sample size by age group and assessment.

Age Group Baseline Post Test Annual 1 Annual 2 Annual 3 Annual 5

65–69 182 163 138 130 126 108

70–74 223 202 172 169 155 134
Tucker-Drob et al.

75–79 156 149 124 110 100 86

80–84 92 82 63 54 39 30
85–89 37 36 30 27 23 15
Total 690 632 527 490 443 373

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

Page 24
 NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript

Table 2
Model Fit Comparisons – Unconditional Models. The best-fitting models are in bold.

Goodness-of-fit indices
Model X2 df CFI TLI AIC BIC RMSEA Δ X2 Δ df p of Δ

Reasoning
Tucker-Drob et al.

Fully Constrained 163.019 121 .990 .994 18186.253 18249.766 .053 baseline
Free Retest Mean 155.239 117 .991 .994 18186.472 18268.132 .049 7.78 4 >.05
Free Retest Mean and Shape 126.101 101 .994 .996 18189.335 18343.582 .042 29.138 16 <.05
Speed
Fully Constrained 189.688 121 .969 .981 20329.851 20393.303 .064 baseline
Free Retest Mean 167.514 117 .977 .985 20315.677 20397.259 .056 22.174 4 <.001
Free Retest Mean and Shape 148.728 101 .978 .984 20328.890 20482.990 .059 18.786 16 >.05

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

Page 25
 NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript

Table 3
Model Fit Comparisons – Conditional Models. The best-fitting models are in bold.

Goodness-of-fit indices
Model X2 df CFI TLI AIC BIC RMSEA Δ X2 Δ df p of Δ

Reasoning - Fully Constrained 205.092 160 .991 .992 26113.245 26340.080 .045 baseline
Tucker-Drob et al.

Reasoning - Predictors freed 187.435 144 .991 .991 26127.588 26427.010 .047 17.657 16 >.05
Speed - Fully Constrained 259.264 176 .966 .973 28497.663 28651.911 .059 baseline
Speed - Predictors freed 248.302 160 .964 .968 28518.701 28745.536 .063 10.962 16 >.05

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

Page 26
 NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript

Table 4
Key Parameter Estimates for Reasoning Model (baseline, fully constrained, model from Table 3).

Total Effects Direct Effects Mediated by i and/or vocab

Estimate/ Standardized Estimate/ Standardized Estimate/ Standardized
Parameter Standard Parameter Parameter Standard Parameter Parameter Standard Parameter
Parameter Estimate Error Estimate Estimate Error Estimate Estimate Error Estimate
Tucker-Drob et al.

i on education (βei) 1.464* 10.154* .455* .663* 4.419* .206* .801* 8.778* .249*
i on vocab (βvi) .473* 11.321* .518* .473* 11.321* .518*
s on i (βis) −.030* −4.683* −.668* −.030* −4.683* −.668*
s on education (βes) −.019 −1.481 −.133 .013 .995 .090 −.032* −3.838* −.224*
s on vocab (βvs) −.007 −1.753 −.178 .007 1.539 .167 −.014* −4.335* −.346*
r on education (βer) −.055 −1.152 −.214 −.101 −1.848 −.396 .046 1.613 .182

r on vocab (βvr) .014 .947 .191 −.003 −.145 −.035 .016 1.591 .226

r on i (βir) .035 1.609 .437 .035 1.609 .437

vocab on education (βev) 1.695* 13.884* .481* 1.695* 13.884* .481*

*
Note: p<.05. i, s, and r are latent growth curve intercept, maturational slope, and retest component respectively.

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

Page 27
 NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript

Table 5
Key Parameter Estimates for Speed Model (baseline, fully constrained, model from Table 3).

Total Effects Direct Effects Mediated by i and/or vocab

Estimate/ Standardized Estimate/ Standardized Estimate/ Standardized
Parameter Standard Parameter Parameter Standard Parameter Parameter Standard Parameter
Parameter Estimate Error Estimate Estimate Error Estimate Estimate Error Estimate
Tucker-Drob et al.

i on education (βei) .712* 4.230* .240* .261 1.390 .088 .451* 4.771* .152*
i on vocab (βvi) .266* 5.079* .316* .266* 5.079* .316*
s on i (βis) −.020* −2.515* −.819* −.020* −2.515* −.819*
s on education (βes) −.009 −.619 −.127 .011 .755 .153 −.021* −2.471* −.280*
s on vocab (βvs) −.009 −1.829 −.431 −.004 −.766 −.172 −.005* −2.289* −.259*
r on education (βer) .029 .415 .053 −.054 −.670 −.098 .084* 2.012* .150*
r on vocab (βvr) .037 1.643 .233 .015 .635 .095 .022* 2.192* .137*
r on i (βir) .082* 2.437* .434* .082* 2.437* .434*
vocab on education (βev) 1.698* 13.900* .482* 1.698* 13.900* .482*

*
Note: p<.05. i, s, and r are latent growth curve intercept, maturational slope, and retest component respectively.

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

Page 28
 NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript

Table 6
Fit Indices and Key Parameter Estimates for Bivariate Latent Growth Model.

Parameter Parameter Estimate Estimate/Standard Error Standardized Parameter Estimate

Model Fit Indices

X2 601.767
Tucker-Drob et al.

df 423
CFI .974
TLI .980
RMSEA .055
Parameter Estimates
Reasoning slope on Reasoning intercept (sR on iR) −.027* −4.148* −.519*
Reasoning slope on Speed intercept (sR on iS) −.003 −.298 −.045

Speed slope on Speed intercept (sS on iS) −.014 −1.606 −.343

Speed slope on Reasoning Intercept (sS on iR) −.018* −2.261* −.467*

Reasoning retest on Reasoning intercept (rR on iR) −.016 .764 .322

Reasoning retest on Speed intercept (rR on iS)

Speed retest on Speed intercept (rS on iS) .086* 2.331* .568*

Speed retest on Reasoning intercept (rS on iR) −.008 −.257 −.059

Reasoning intercept with Speed intercept (iR with iS) 45.734* 15.178* .626*
Reasoning slope with Speed slope (sR with sS) .102* 4.033* .636*
Reasoning retest with Speed retest (rR with rS) .172 .397 .320

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

*
Note: p<.05. i, s, and r are latent growth curve intercept, maturational slope, and retest component respectively. The subscripts R and S refer to reasoning and speed respectively.
Page 29
 NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript

Table A1
Fit Indices and Parameter Estimates for Alternative Growth Curve Models of Reasoning

Age and Occasion

Basis Structural
Time and Occasion Basis (Structural Equation Growth Equation Growth
Model) Age and Occasion Basis (Random Effects Growth Model) Model
Tucker-Drob et al.

Linear Basis by Latent Basis by Linear + Latent Latent Basis by Multiple Age Group
Time (Equation Occasion Basis b (Equation Linear Basis by Age Occasion Linear + Latent Basis Linear + Latent Basis
A1) (Equation A2) A3) (Equation A4) (Equation A2) b (Equation A5) (Equation 1)

Model Fit Indices

X2 313.407 71.282 13.262 - 71.282 - 243.804
Degrees of Freedom 29 25 20 - 25 - 180
−2×Log Likelihood 26464.93 26222.806 26164.786 23129.632 22893.840 22720.648 26051.956
Free Parameters 13 17 22 13 17 22 30
CFI .943 .991 1 - .991 - .987
TLI .945 .990 1 - .990 - .990
RMSEA .119 .052 0 - .052 - .051
AIC 26490.93 26256.806 26208.785 23155.631 22927.840 22764.648 26111.957
BIC 26549.907 26333.929 26308.592 23214.608 23004.964 22864.455 26248.057
Regression Parameter Estimates
i on education (βei) .719* .722* .748* .714* .722* .566* .639*
i on vocab (βvi) .411* .404* .405* .483* .404* .525* .481*
s on i (βis) .010 - .011 −.015* - −.026* −.025*
s on education (βes) −.036* - −.034 .004 - .017 .015

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

s on vocab (βvs) −.006 - −.006 .001 - .003 .004

r on education (βer) - −.126 −.088 - −.126 −.096 −.111

r on vocab (βvr) - −.027 −.010 - −.027 .017 .013

r on i (βir) - .067* .028 - .067* .008 .016

vocab on education (βev) 1.670* 1.671* 1.671* 1.671* 1.671* 1.673* 1.694*
Basis Coefficients
BL on s 0a 0a 0a ageBL,n-65a 0a ageBL,n -65a ageBL,g-65a

PT on s .23a 0a 0.2a ageBL,n -64.77a 0a ageBL,n -64.77a ageBL,g -64.77a

Page 30
 NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript

Age and Occasion

Basis Structural
Time and Occasion Basis (Structural Equation Growth Equation Growth
Model) Age and Occasion Basis (Random Effects Growth Model) Model

Linear Basis by Latent Basis by Linear + Latent Latent Basis by Multiple Age Group
Time (Equation Occasion Basis b (Equation Linear Basis by Age Occasion Linear + Latent Basis Linear + Latent Basis
A1) (Equation A2) A3) (Equation A4) (Equation A2) b (Equation A5) (Equation 1)
Tucker-Drob et al.

YR1 on s 1.23a 0a 1.23a ageBL,n -63.77a 0a ageBL,n -63.77a ageBL,g -63.77a

YR2 on s 2.23a 0a 2.23a ageBL,n -62.77a 0a ageBL,n -62.77 a ageBL,g -62.77a

YR3 on s 3.23a 0a 3.23a ageBL,n -61.77a 0a ageBL,n -61.77a ageBL,g -61.77a

YR5 on s 5.23a 0a 5.23a ageBL,n -59.77a 0a ageBL,n -59.77 a ageBL,g -59.77a

BL on r 0a 0a 0a 0a 0a 0a 0a
PT on r 0a 1a 1a 0a 1a 1a 1a
YR1 on r 0a .690* .654* 0a .690* 0.966* 0.952*
YR2 on r 0a .767* .677* 0a .767* 1.264* 1.236*
YR3 on r 0a .868* .733* 0a .867* 1.625* 1.584*
YR5 on r 0a .410* .102 0a .410* 1.596* 1.529*
Residual Variance Estimates
reason res. variance (σy2) 9.089* 9.352* 8.124* 9.495* 9.353* 8.066* 8.070*
i residual variance (σi2) 46.397* 45.497* 45.069* 44.701* 45.494* 48.061* 46.566*
s residual variance (σs 2)
.215* - 0.281* .100* - .077* .090*
r residual variance (σr2) - −3.456* −1.403 - −3.458* .986* .912*
vocab residual variance (σv2) 73.302* 73.311* 73.300* 73.289* 73.312* 73.294* 73.075*

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

Latent Variable Mean Estimates
mean i (υi) 48.798* 47.459* 45.069* 50.587* 47.459* 52.555* 51.287*
mean s (υs + βis × υi) .073 - .125 −.140* - −.561* −.527*
mean r (υr + βir × υi) - 2.534* 2.468* - 2.534* 2.518* 2.515*

*
Note: p<.05.
a
indicates that a parameter was fixed.
Page 31
 NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript
b
denotes the best fitting model by nested model comparisons relative to the two previous models. ageBL,n is age at first testing. ageBL,g is approximate group age at first testing. All parameters are
unstandardized. Because age at testing varied considerably across individuals, all Age and Occasion Basis Random Effects Growth Models were fit with the RANDOM option of Mplus, which has a
provision for individually varying times of observation (TSCORES).
Tucker-Drob et al.

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

Page 32
 NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript

Table A2
Fit Indices and Parameter Estimates for Alternative Growth Curve Models of Speed

Age and Occasion

Basis Structural
Time and Occasion Basis (Structural Equation Growth Equation Growth
Model) Age and Occasion Basis (Random Effects Growth Model) Model
Tucker-Drob et al.

Linear Basis by Latent Basis by Latent Basis by Multiple Age Group

Time (Equation Occasion Linear + Latent Basisb Linear Basis by Age Occasion Linear + Latent Basisb Linear + Latent Basis
A1) (Equation A2) (Equation A3) (Equation A4) (Equation A2) (Equation A5) (Equation 1)

Model Fit Indices

X2 243.121 40.999* 18.44 - 40.999* - 278.669*
Degrees of Freedom 29 25 20 - 25 - 180
−2×Log Likelihood 28884.044 28681.922 28659.362 25510.938 25352.956 25124.986 28449.068
Free Parameters 13 17 22 13 17 22 30
CFI .926 .994 1 - .994 - .96
TLI .929 .994 1 - .994 - .969
RMSEA .103 .030 0 - .030 - .063
AIC 28910 28715.9 28703.4 25536.9 25387 25168.99 28509.068
BIC 28969 28793 28803.2 25595.9 25464.1 25268.79 28645.168
Regression Parameter Estimates
i on education (βei) .494* .528* .509* .276 .528* .214 .257

i on vocab (βvi) .194* .166* .169* .267* .166* .309* .269*

s on i (βis) .013 - .014 −.013 - −.021* −.019*
s on education (βes) .010 - .016 .020 - .009 .01

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

s on vocab (βvs) −.005 - −.004 −.002 - −.005 −.004

r on education (βer) - −.142 −.129 - −.142 .002 −.043

r on vocab (βvr) - −.009 .002 - −.009 .012 .015

r on i (βir) - .228* .170* - .228* .062 .070*

vocab on education (βev) 1.675* 1.675* 1.675* 1.675* 1.675* 1.676* 1.698*
Basis Coefficients
BL on s 0a 0a 0a ageBL,n -65a 0a ageBL,n -65a ageBL,g -65a

PT on s .23a 0a 0.2a ageBL,n -64.77a 0a ageBL,n -64.77a ageBL,g -64.77a

Page 33
 NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript

Age and Occasion

Basis Structural
Time and Occasion Basis (Structural Equation Growth Equation Growth
Model) Age and Occasion Basis (Random Effects Growth Model) Model
Linear Basis by Latent Basis by Latent Basis by Multiple Age Group
Time (Equation Occasion Linear + Latent Basisb Linear Basis by Age Occasion Linear + Latent Basisb Linear + Latent Basis
A1) (Equation A2) (Equation A3) (Equation A4) (Equation A2) (Equation A5) (Equation 1)

YR1 on s
Tucker-Drob et al.

1.23a 0a 1.23a ageBL,n -63.77 a 0a ageBL,n -63.77a ageBL,g -63.77a

YR2 on s 2.23a 0a 2.23a ageBL,n -62.77 a 0a ageBL,n -62.77a ageBL,g -62.77a

YR3 on s 3.23a 0a 3.23a ageBL,n -61.77 a 0a ageBL,n -61.77a ageBL,g -61.77a

YR5 on s 5.23a 0a 5.23a ageBL,n -59.77a 0a ageBL,n -59.77a ageBL,g -59.77a

BL on r 0a 0a 0a 0a 0a 0a 0a
PT on r 0a 1a 1a 0a 1a 1a 1a
YR1 on r 0a .944* .915* 0a .944* 1.218* 1.216*
YR2 on r 0a .959* .897* 0a .959* 1.487* 1.482*
YR3 on r 0a .940* .853* 0a .940* 1.776* 1.763*
YR5 on r 0a .561* .358 0a .561* 1.800* 1.786*
Residual Variance Estimates
speed res. variance (σy2) 24.795* 25.047* 23.228* 26.215* 25.051* 22.890* 22.959*
i residual variance (σi2) 68.978* 56.735 57.609* 62.403 56.729 61.798* 59.613*
s residual variance (σs2) .333* - .421* .093 - .017 −.004

r residual variance (σr2) - −8.147* −5.013 - −8.160* 2.261* 2.360*

vocab residual variance (σv2) 73.285* 73.288* 73.286* - 73.288* 73.280* 73.065*

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

Latent Variable Mean Estimates
mean i (υi) 47.739* 45.810* 45.825* 51.703* 45.810* 53.006* 51.291*
mean s (υs + βis × υi) .140* - .077 −.346* - −.802* −.776*
mean r (υr + βir × υi) - 3.214* 3.196* - 3.214* 3.216* 3.185*

*
Note: p<.05.
a
indicates that a parameter was fixed.
Page 34
 NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript
b
denotes the best fitting model by nested model comparisons relative to the two previous models. ageBL,n is age at first testing. ageBL,g is approximate group age at first testing. All parameters are
unstandardized. Because age at testing varied considerably across individuals, all Age and Occasion Basis Random Effects Growth Models were fit with the RANDOM option of Mplus, which has a
provision for individually varying times of observation (TSCORES).
Tucker-Drob et al.

Dev Psychol. Author manuscript; available in PMC 2011 December 5.

Page 35