Professional Documents
Culture Documents
20th century (e.g., Birren, 1964). Research on intelligence and intellectual development played a major
role in shaping the field of psychological gerontology (e.g., Botwinick, 1977). This chapter reviews what
is known and not yet known about adult intellectual development after decades of research on the
topic. Most of the information we have available concerns aspects of what Sternberg (1985) has defined
as academic intelligence (based on traditional psychometric tests of human abilities).
This chapter focuses on what is known about these types of human abilities and their correlates,
although I also briefly treat other aspects of intellect, such as practical intelligence and tacit knowledge.
Early studies of psychometric intelligence prior to 1940 determined that there were large differences in
performance on general tests of intellectual aptitude (see Salthouse, 1982 for an excellent summary and
review). Wechsler (1939) characterized the performance tests on the Wechsler Adult Intelligence Scale
(WAIS) as “don’t-hold” tests because of the lower performance on those subscales (e.g., WAIS Block
Design) by older adults in his cross-sectional norming studies of the test. Conversely, Wechsler found
that tests like WAIS vocabulary were typically shown to have much smaller age differences, causing
them to be characterized as “hold” tests. This basic idea, that one class of intellectual ability tests
manifests age decline whereas others do not, has been widely replicated and studied across a variety of
intelligence tests, and today represents a virtual “truism” about aging and intelligence. These findings
mirrored outcomes of studies using other tests to evaluate age differences in human abilities, studies
that spanned much of the 20th century (Salthouse, 1982). The concept of contrasting maintenance of
knowledge and verbal abilities, relative to other types of human abilities, has therefore figured
prominently in theoretical treatment of how aging affects intelligence. Cattell (1971) developed the
theory of fluid and crystallized intelligence, arguing that this basic pattern reflected two prototypic
classes of intellectual abilities. Fluid intelligence was seen as the fundamental ability to think, reason,
and process information, and prone to adult age decline as a function of biological aging processes (Horn
& Cattell, 1967; Horn & Hofer, 1992). Crystallized intelligence, on the other hand was seen as
determined by investment of fluid intelligence in knowledge acquisition, which was largely maintained
or even improved into old age (Horn & Cattell, 1967).
Baltes and his colleagues characterized the distinction as involving a decline in basic information-processing
mechanisms labeled the mechanics of cognition (e.g., Baltes, 1997). In contrast, experience with a culture leads to
acquisition of a broad class of declarative and procedural knowledge and skills about how to achieve goals in a
cultural context, labeled the pragmatics of intelligence. Although Baltes’ conceptualization emphasized mechanisms
that influence observed abilities, similar arguments were being made by Horn (e.g., Horn & Hofer, 1992) in
extended versions of fluidcrystallized theory. As a consequence, the differences between these theoretical
viewpoints are subtle at best. Can a two-curve model actually account for most of the age-related variance in adult
intellectual development? If so, it would be surprising, for several reasons. First, theories of psychometric abilities
generally acknowledge that a large number of intellectual abilities exist. Theoretical approaches based on the work
by Thurstone on primary mental abilities (e.g., Thurstone, 1938) typically argue for 30 or more primary abilities
(Carroll, 1993; Horn & Hofer, 1992). It would be surprising if all these abilities declined at the same rate in
adulthood. Second, contemporary hierarchical models of abilities typically acknowledge that fluid and crystallized
intelligence are distinct from other higher order ability factors. Horn (1985; Horn & Hofer, 1992) argued that, for
example, general visualization abilities, general auditory abilities, speediness, and secondary memory are all
empirically distinct from fluid intelligence. To the extent that these second-order factors are indeed differentiable
from fluid intelligence, one might expect their developmental curves in adulthood to also differ. Third, theories of
biological aging identify a large number of potential biological clocks, operating at different levels of basic
physiology, that appear to be associated with rates of biological aging. What do the empirical data tell us? The cross-
sectional age curves for episodic memory, spatial visualization, and measures of fluid intelligence and general
processing speed vary somewhat as a function of issues like how the tests are constructed and scaled, their
processing requirements, and the like. Yet there is surprising similarity in the curves across these different classes of
abilities. Certainly the ability that is typically found to have the largest crosssectional age differences is speed of
processing, such as identified by the Perceptual Speed factor (Carroll, 1993). Salthouse (1996) has evaluated
Perceptual Speed in a plethora of studies, typically finding the largest cross-sectional age differences for that factor
(see also Schaie, 1989). However, fluid intelligence shows considerable similarity in magnitude of estimated decline
to measures of episodic memory, working memory, and spatial visualization (e.g., Hertzog, 1989; Hultsch, Hertzog,
Dixon, & Small, 1998; Park et al., 1996; Salthouse, Pink, & Tucker-Drob, 2008). No one study has examined all the
relevant abilities in a truly representative sample of the adult population, and most observe at least some variation in
cross-sectional age slopes across abilities. Nevertheless, the available crosssectional evidence on the mechanics of
cognition is more or less consistent with the argument that abilities emphasizing cognitive mechanics decline in
adulthood. There are important exceptions – not all processing mechanisms decline, and not all aspects of
pragmatics are maintained (see Hertzog, 2008). Also, cross-sectional data disagree as to whether the cross-sectional
curves are linear or curvilinear – accelerating the magnitude of estimated decline in old age (e.g., compare Hultsch
et al., 1998, with Park et al., 1996, regarding episodic memory). Nevertheless, the negative correlation of age with
fluid intelligence, working memory, spatial visualization, and the like from early adulthood to old age is about –. 4.
There is evidence that the cross-sectional age curves for crystallized intelligence may differ as a function of the type
of knowledge being assessed. Work by Ackerman and colleagues has focused on tracking domain-specific
knowledge that may occur during and after the time that young adults begin to specialize vocational and personal
interests, crystallizing them into a pattern of preferences for information sought, acquired, digested, and assimilated
into existing knowledge structures (e.g., Ackerman, 2000; Beier & Ackerman, 2005). Ackerman’s argument is that
crystallized intelligence, as manifested in general cultural knowledge tests (like WAIS Information) or in
recognition vocabulary tests, underestimate acquisition of new knowledge during adulthood. Thus, although the
existing psychometric data suggesting longterm stability in verbal abilities and cultural knowledge diverges from the
pattern of negative age differences seen with fluid intelligence and other human abilities, it may not capture the
lifelong learning that occurs in the specific domains in which people invest time and effort to acquire knowledge.
Even within the domain of vocabulary, there may be activity-dependent differences in the types of word knowledge
that are acquired. Frequent crossword puzzle players show major cross-sectional age differences in esoteric
vocabulary terms that are correctly recognized, probably as a direct function of actual experience with encountering
these terms while solving puzzles (Hambrick, Meinz, & Salthouse, 1999). Be that as it may, there is little question
that abilities that reflect specific knowledge acquisition are maintained or improved, at least into the 60s. Beier and
Ackerman’s (2005) work on specificity of knowledge acquisition resonates with other evidence that people of
different ages also differ in historical life contexts that produce cohort differences in knowledge-based abilities.
Schaie (2005) has studied adult intellectual development for over 50 years, using hybrid cross-sectional and
longitudinal designs known as sequential strategies, enabling an evaluation of age changes across different birth
cohorts and epochs of historical time. One of Schaie’s findings is that there are large cohort differences in
vocabulary, which helps to explain why studies of age and cognition that use older vocabulary tests – particularly
with “advanced” and perhaps dated items – tend to find that older adults perform better than younger adults. Such
age differences probably reflect a combination of improvement with experience in the older adults, but lower
knowledge of esoteric word meanings in younger generations. By the same token, it is likely to be true that younger
adults have more word knowledge in domains they commonly employ, such as technical terms and jargon associated
with advanced technology (older adults are less likely to use new technology such as iPhones or iPods; Czaja et al.,
2006). Schaie (2005) has also shown that there are cohort differences favoring earlier born generations in simple
mental calculations such as two column addition. One could view this effect as being a societal consequence of the
use of computers and calculators, slowing the efficiency of mental arithmetic in more recent cohorts apt to rely on
technological support. In sum, the distinction in developmental functions between knowledge and experience-based
abilities, on the one hand, and fluid-like abilities, on the other hand, is consistent with a large body of crosssectional
evidence.
Longitudinal Evidence Regarding Levels of Adult Intellectual Development
As noted earlier, Schaie and colleagues (e.g., Schaie, 2005) have assembled the largest extant database with
combined longitudinal and cross-sectional intelligence test data. A reasonable question to ask, then, is whether these
data produce radically different conclusions regarding age changes in adult intellectual development, relative to the
crosssectional data. On the one hand, Schaie’s (2005) data clearly indicate that cohort differences are not confined to
aspects of knowledge and crystallized intelligence. He also observes substantial generational differences on a tests
of fluid reasoning and spatial relations. Others have noted the changes during the 20th century in performance on
tests of reasoning and fluid intelligence, as manifested in the so-called Flynn effect (Flynn, 2007; Raven, 2000). The
impact of these cohort effects is primarily in attenuating the estimated changes in intelligence from ages 20 to 50,
but they also reduce the magnitude of estimated age change in late life as well (Zelinski, Kennison, Watts, & Lewis,
2009). Certainly the STAMAT Verbal Meaning test shows a prolonged period of maintenance, relative to the other
abilities, but it too manifests evidence of longitudinal decline in old age. Separate evidence, however, suggests that
this pattern of apparent decline is an artifact of the speeded properties of the STAMAT Verbal Meaning test (e.g.,
Hertzog, 1989). In fact, all of the STAMAT tests are substantially influenced by speed of processing, in part because
of limited item difficulty, even for the Letter Series and Space tests. The pattern of mean ability changes based on
sequential data can be separated into three parts. The first is the similarity of age changes across different aspects of
cognitive mechanics. The second is the conclusion that meaningful age-related changes in cognitive mechanics
occur after mid-life and accelerate in magnitude in late life. The third is the presence of substantial cohort effects on
variables measuring different aspects of cognitive mechanics that inflate estimates of age changes made from cross-
sectional data. Regarding cohort effects, there is broad agreement across studies that there are few cohort effects in
general informationprocessing speed, including the Perceptual Speed factor identified by psychometric tests (e.g.,
Hultsch et al., 1998; Schaie, 1990). However, the limited available data from studies other than Schaie’s Seattle
Longitudinal Study confirm substantial cohort effects on tests of reasoning (Raven, 2000; Zelinski & Kennison,
2007; R¨onnlund & Nilsson, 2008) and visuospatial ability (R¨onnlund & Nilsson, 2008; Zelinski & Kennison,
2007). These effects attenuate estimated age changes in cognition. For example, Zelinski and Kennison (2007)
found that six-year effect sizes in reasoning, spatial ability, and episodic memory were reduced in old age by
between 0.2 and 0.3 standard deviations (SD) by controlling on cohort differences. Interestingly, some studies report
few cohort effects on crystallized intelligence while finding larger effects on abilities more related to cognitive
mechanics (see Zelinski et al., 2009; cf. Alwin, 2009). The conclusion that declines in cognitive mechanics are
subtle before age 50 and accelerating thereafter is broadly consistent with reported results from a number of other
longitudinal studies of cognition and intellectual abilities in adulthood, including the Long Beach Longitudinal
Study (Zelinski & Kennison, 2007), the Victoria Longitudinal Study (Hultsch et al., 1998), and the Betula
Longitudinal Study (R¨ onnlund, Nyberg, B¨ackman, & Nilsson, 2005). These studies all suggest curvilinear
patterns of average age changes from the period of midlife through old age, with an acceleration in the rate of aging
effects on fluid intelligence, episodic memory, and spatial visualization and other fluid-like abilities after age 65.
Salthouse (2009) has argued that the type of longitudinal gradients produced by Schaie (2005) are contaminated by
practice effects on the tests, an internal validity threat (Shadish, Cook, & Campell, 2002) that is problematic for
longitudinal designs (Schaie, 1977). Because individuals are repeatedly given the same tests, they may show some
savings in generating problem answers. If it were the case that younger adults manifest larger practice effects (an
age X practice interaction), perhaps due to retention of prior test answers, then the contamination by practice would
produce shallower age slopes. One way to address the problem of practice effects has been to incorporate effects of
number of occasions of measurement as a proxy for exposure that would benefit from practice. Models that use this
approach also tend to increase the magnitude of age-related decline and estimate an earlier onset of reliable age-
related decline (e.g., Ferrer, Salthouse, Stewart, & Schwartz, 2004; Rabbitt, Diggle, Holland, & McInnes, 2004).
However, this modeling approach is controversial (see the exchange between Salthouse, 2009, Schaie, 2009, and
Nilsson, Stern¨ang, R¨ onnlund, & Nyberg, 2009). A model that uses all available data in a standard longitudinal
panel and then jointly estimates age changes and practice effects (under the convergence assumption – see McArdle
& Bell, 2001) confounds the estimates of practice effects with other influences that are not modeled, including
historical period (time), experimental mortality (attrition), and selection X period interactions. Sliwinski, Hoffman,
and Hofer (2010) argue that such models inevitably assign within-person changes that deviate from cross-sectional
trends to estimates of practice, morphing the estimated age effects away from within person change toward between-
person differences. As pointed out by Nilsson et al. (2009), studies that use an independent samples comparison
group to estimate practice effects report far less impressive practice adjustments than studies like Ferrer et al.
(2004).
Age Changes in the Factor Structure of Intelligence Tests
Another important question about aging is whether it influences the underlying factor structure of human abilities. A
leading developmental hypothesis has been the dedifferentiation hypothesis (deFrias, L¨ovd´en, Lindenberger, &
Nilsson, 2007). It states that shared causes of age effects across different kinds of human abilities will produce
increased correlations among ability factors. In the extreme, such changes could lead to a reduced number of distinct
human abilities. Factor analytic questions of this type cannot be separated from issues of how broadly or narrowly
tests are selected. A unifying perspective on this issue derives from hierarchical models of abilities, such as in
Carroll (1993). This view suggests that one can evaluate factor structure at a relatively narrow level (how different
tests define primary abilities, such as inductive reasoning or working memory), at a second-order level (how
different primary abilities define higher order factors like fluid intelligence, general speed of processing, or spatial
visualization), or at the highest levels (how second-order factors define a highest order general intelligence factor).
At the primary ability or second-order level, one can also evaluate the correlations among ability factors, treating
these correlations as an index of differentiation. In addressing these questions one can run into difficulty separating
measurement invariance and suboptimal measurement properties of tests from changes in relationships among
constructs. For example, use of speeded tests of intelligence may produce a substantial degree of dedifferentiation
that is attributable to the global effects of speed of processing on test performance, rather than because the
underlying ability constructs are becoming more correlated (Hertzog & Bleckley, 2001). The best available evidence
suggests that the factor structure of intelligence is not materially affected by aging. A large number of confirmatory
factor analytic studies, using both cross-sectional and longitudinal data, indicate that the same human abilities can be
identified in young adulthood, middle age, and old age (e.g., Anstey, Hofer, & Luszcz, 2003; Hertzog & Schaie,
1986; Hertzog, Dixon, Hultsch, & Mac-Donald, 2003; Hultsch et al., 1998; Brickley, Keith, & Wolfe, 1995; Lane &
Zelinski, 2003; Schaie et al., 1998). In all cases, the hypothesis of configural invariance (i.e., that the same variables
load on the same factors at all ages; Meredith & Horn, 2001) has been supported. In most cases, the evidence
supports the stronger hypothesis of metric invariance, that the unstandardized factor pattern weights, or factor
loadings (Meredith & Horn, 2001), are equivalent across time in longitudinal studies or are equivalent across age
groups. This is a broad generalization, and there are some interesting exceptions. Nevertheless, the developmental
changes that occur in adulthood do not appear to radically alter the underlying nature of human abilities. On the
other hand, the evidence regarding whether adult development results in increasing correlations among human
ability factors is mixed. Some studies have not found such effects (e.g., Zelinski et al, 2009; Bickley et al., 1995),
whereas other studies have (deFrias et al., 2007; Hertzog & Bleckley, 2001; Hertzog et al., 2003; Hultsch et al.,
1998; Schaie et al., 1998; Verhaeghen & Salthouse, 1997). However, major increases in factor correlations may be
restricted to old age (deFrias et al., 2007; Schaie et al., 1998). One methodological concern with agecomparative
factor analysis is that aggregation over long epochs of age is often needed to generate sufficient sample sizes for
factor analysis of cross-sectional data. For example, one might pool data from people within the ages of 20 and 39,
40 and 59, 60 and 79 to create “young,” “middle aged,” and “old” age groups. Aggregation over wide age spans
(such as 20 years) can create spurious increases in factor correlations because of the inflating influence of age-
heterogeneity on variable correlations (Hofer, Flaherty, & Hoffman, 2006). Given greater average age change after
age 60 that is similar across variables, factor correlations in the oldest group would be inflated. Forming narrower
age spans, if possible given the sample size, avoids this effect. In sum, factor analytic evidence indicates subtle
changes, if any, in the factor structure of human abilities. Thus, quantitative comparisons of ability test scores may
not be compromised by age-related changes in the measurement properties of the tests (Baltes & Nesselroade,
1970).
Disease and brain pathology. The findings of Raz et al. (2008) do not necessarily imply that neurobiological aging
in the brain drives cognitive changes. The morphological changes in the brain can also be caused by disease, such as
cardiovascular disease and dementia. Sliwinski et al. (2003) conducted a fascinating study in this regard, using data
from the Bronx Longitudinal Study (Sliwinski & Buschke, 2004). The study involved a prospective design of the
incidence of dementing illnesses in a nondemented control group collected as part of a larger study of Alzheimer’s
disease and related disorders. Individuals in this group were measured cognitively at regular intervals, but they were
also assessed for dementia. Over time some of the participants were clinically diagnosed as having dementia, and
this allowed Sliwinski and colleagues to compare cognitive change in the preclinical phase with change in those
individuals who did not convert to dementia. As might be expected, individuals who had not yet been diagnosed
with dementia (but undoubtedly had contracted the disease) showed greater change in episodic memory during their
preclinical phase, compared to individuals who did not later receive a dementia diagnosis. Even more interesting,
however, was the fact that the aggregate control sample manifested individual differences in cognitive change, as
well as correlations of changes across cognitive variables. However, the magnitude of individual differences was
reduced by controlling on later dementia diagnosis, as were the correlations of change among different variables.
Furthermore, within the dementia group, organizing the time scale by point of diagnosis rather than chronological
age eliminated the individual differences in rates of cognitive change. What does this pattern imply? It would appear
that in this sample, the presence of preclinical dementia was a major source of individual differences in change.
Because people vary in the age at which the disease is contracted and later diagnosed, organizing the data by age
(without knowledge of the disease and its progression) produces larger individual differences in rates of change.
Given that other prospective studies of Alzheimer’s disease, vascular dementia, and other dementing illnesses
indicate a fairly long preclinical period in which cognition may be affected (e.g., B¨ackman & Small, 2007), it
would appear that a major influence on individual differences in cognitive change in old age is the presence or
absence of dementia. Furthermore, a number of studies have directly linked magnitudes of longitudinal changes in
cognitive abilities to different kinds of disease, including cardiovascular and cerebrovascular disease, late-onset
diabetes, and their precursors, or risk factors, such as obesity, hypertension, poor cholesterol profiles, and the like
(Spiro & Brady, 2008).
Disease and terminal decline. A focus on disease effects on cognition raises an additional set of important questions
about aging and intellectual development. To what extent are the average curves for cognitive abilities and age
misleading, in the sense that they are not representative of the actual developmental trajectories of individuals?
Means, even if generated from longitudinal data, are simply best guesses as to the level of function, on average, at a
particular age. We link the means of different ages with a line (or fit a curve to the data), but this does not imply that
the developmental pathways of individuals have the shape implied by the shape of the aggregate mean curve. The
population of adults might be quite heterogeneous in nature, with the major changes in psychological functioning,
including cognition, occurring during a period of decline preceding death (e.g., Berg, 1996; Bosworth, Schaie, &
Willis, 1999). Indeed, time to death may be a more important way of indexing cognitive loss in old age than
chronological age (Singer et al. 2003). Some new and impressive data on this score come from models of
longitudinal data that jointly use time to death and age to organize the data (Ram et al., 2010). The modeling
approach is fairly complex, requiring estimation of a change point (Hall, Sliwinski, Stewart, & Lipton, 2000), at
which the slope of decline prior to a change point is lower than the slope immediately prior to death. Thorvaldsson
et al. (2008) used this method to demonstrate accelerated cognitive decline occurring about seven years before death
in the Swedish Goteborg Longitudinal Study data. Wilson, Beck, Bienias, and Bennett (2007) found evidence for a
shorter period of terminal decline of about four years. Terminal decline was associated with the apolipoprotein E ε4
allele, a genetic polymorphism thought to be associated with risk for Alzheimer’s disease (AD). Laukka,
MacDonald, and Backman (2008) also concluded that a substantial proportion of the variance in terminal cognitive
decline might be due to emergence of dementia, but there was evidence of decline in individuals who did not
develop AD. Undoubtedly future research will clarify the extent to which other disease factors play a role in
terminal cognitive decline, including vascular disease and organ failure (e.g., renal dysfunction, see Buchman et al.,
2009). In light of the evidence for terminal decline effects, the possibility exists that the curvilinear age trends for
cognitive function in late life are actually an artifact of aggregation over individuals with different functions. This
idea was nicely illustrated by Baltes and Labouvie (1973), who showed that a combination of (1) a change point
function of stable level of cognition, followed by terminal decline, and (2) a variable onset of the terminal decline
that was correlated with advancing age could produce aggregate curvilinear functions that did not capture the
functional form of individual change (see Figure 9.2). The aggregate function could be influenced by the increasing
risk of terminal decline, with its curvature reflecting an averaging of persons in terminal decline with persons who
are still stable.
Exercise and an engaged life style. A critical question regarding adult intellectual development is whether health-
promoting behaviors such as exercise, nutrition, and an active lifestyle promote better developmental outcomes
(Hertzog, Kramer, Wilson, & Lindenberger, 2009). Over the last decade, compelling evidence has emerged that
aerobic exercise in middle age and old age promotes enhanced cognitive function in older adults. Colcombe and
Kramer’s (2003) metaanalysis evaluated aerobic exercise intervention studies in older adults and compared exercise
groups’ cognitive performance to performance in a groups doing toning and stretching only. Short-term aerobic
exercise resulted in substantial improvements in tasks assessing executive functioning and controlled attention
(domains highly correlated with fluid intelligence; Salthouse et al., 2008). The data are broadly consistent with
cross-sectional studies suggesting an association of self-reported exercise with human abilities (e.g., Eggermont et
al., 2009), but the intervention effects help to argue for a causal influence of exercise on cognition. Unfortunately,
there are at present no longitudinal studies that contrast longer term adherence with exercise regimens and degree of
cognitive change in adulthood. Does engaging in intellectually stimulating activities also promote better cognitive
outcomes? Salthouse (2006) expressed skepticism on this score, given that his cross-sectional data on self-reported
activities have failed to observe age X activity interactions (see Hertzog et al., 2009, for a critique of this argument).
Certainly, simple cross-sectional correlations of activities and intelligence are insufficient grounds for arguing that
activities help preserve cognitive functioning, because individuals with high intelligence tend to manifest higher
levels of intellectual engagement in early adulthood (Ackerman & Heggestad, 1997). However, longitudinal
evidence is needed, given the potential lack of sensitivity of cross-sectional data to change alluded to earlier.
Longitudinal studies have often found relationships of self-reported intellectual engagement with cognition (e.g.,
Schooler, Mulatu, & Oates, 1999; Wilson et al., 2003; see Hertzog et al., 2009 for a review). However, as noted by
Hultsch, Hertzog, Small, and Dixon (1999), longitudinal correlations of activities with cognitive change could still
be due to latelife cognitive changes leading to curtailed activity (MacKinnon et al., 2003). There are fewer
intervention studies with activities, but there is at least some indication that encouraging older adults to engage in
stimulating activities may have cognitive benefits (Carlson et al., 2009; Stine-Morrow et al., 2007; Tranter &
Koutstaal, 2008). In one recent study, participation in a complex videogame environment led to short-term
improvements in attentional control and executive function (Basak, Boot, Voss, &Kramer, 2008). This outcome is
consistent with intervention studies that target executive control (Hertzog et al., 2009), producing more transfer of
training than is typically observed when training focuses on teaching specific processing strategies (e.g., Ball et al.,
2002). The evidence favors an impact of activities on cognitive function, but there is still some disagreement and
controversy on this point.