Structure of the Aging Mind
In contrast to the popular notion that mental decline with age is inevitable, progressive, and general, research presents a more complex picture. Aging individuals vary greatly in the rates at which their cognitive functioning changes over the life span; the trajectory of cognitive aging is different for different cognitive functions (e.g., Baltes et al., 1999; Hultsch et al., 1998; Schaie, 1994, 1996; Willis, 1991); and as already noted, experience, including practice, physical exercise, and the status of sensory-motor systems and health, is associated with systematic differences in cognitive performance among older adults.
Much progress has been made over the past quarter-century in identifying which tasks prove particularly difficult for older adults and which do not, and in modeling age-related change in particular cognitive processes. Still lacking, however, are theories that can explain the overall pattern—why some functions are spared while others are impaired, and why patterns of age-related change differ across individuals. It is now possible to make significant progress in this direction. Improvements in measurement of cognitive and neural functioning, new methods of analyzing the data, and an expansion of longitudinal research now make possible substantial improvement in understanding the nature of cognitive aging, including the identification of mechanisms at the behavioral and neural levels that contribute to age-related changes and to differences between individuals and groups.
RECENT SCIENTIFIC ADVANCES
Variations in the Life Histories of Cognitive Functions
Accumulating data on cognitive functioning in later life are yielding a clearer understanding of the complex patterns of change. The Seattle Longitudinal Study (e.g., Schaie 1983, 1994, 1996) provides typical findings. Cross-sectional data from 5,000 adults from age 25 to 88 show consistent negative age effects on such latent abilities as inductive reasoning, spatial orientation, perceptual speed, and verbal memory (new learning). In contrast, numeric ability (simple arithmetic calculations) and verbal ability (synonyms and recognition tests of meaning) show improvement until midlife and then a plateau until the oldest tested age of 81. People at age 81 performed at a higher level on vocabulary tests than people at age 25. These data are for both speeded and unspeeded vocabulary tests combined. If only speeded vocabulary tests are considered, performance declines start in the 60s, although the rate of decline is still lower than for other cognitive functions. Longitudinal data in the Seattle study show an inverted U-shaped age function for most verbal and numerical abilities, with the highest scores achieved at ages from the 30s to the 60s, depending on the ability. World knowledge is usually found to be stable into old age, for example, as measured by the WAIS information subtest (Salthouse, 1982) or general information questions (Camp, 1989; Nyberg et al., 1996; Small et al., 1995; but see Hultsch et al., 1998).
Schaie's data are from psychometric tests that were not designed to identify mental processes and mechanisms underlying the cognitive functions being measured. Using laboratory techniques from cognitive psychology to investigate semantic memory processes, studies consistently report age invariance in semantic organization and processes, and in a variety of other language processes (see Kemper, 1992; Light, 1991; Burke, 1997). This extends even to discourse-level processes, because consistently higher ratings are given to older compared with younger adults' narratives (e.g., James et al., 1998; Kemper et al., 1990). Recent findings, however, demonstrate that some processes involved in language production decline in old age: older adults suffer more failures in retrieval of phonology (word finding failures) and orthography (spelling errors) than young adults, despite their superior vocabularies (Burke et al., 1991; MacKay and Abrams, 1998). This pattern of dissociation in age effects on language functions is at odds with descriptions of age invariance in crystallized or pragmatic functions. Together with dissociations in age effects in other cognitive domains (e.g., explicit versus implicit memory), these data pose a challenge to models of cognitive aging.
Cognitive functions are not only a matter of the speed and accuracy of information processing. The aging mind also involves cognitive contents, such as ideas of self and the meaning of life, and skills that go beyond speed
and accuracy of information processing, some of which are described by the concept of wisdom. Capability in such areas often increases through adulthood at least until the 70s, even when cognitive efficacy in the narrow sense is reduced (e.g., Baltes, 1997).
Recent research shows a less differentiated pattern of cognitive performance in advanced old age, that is, beyond age 85. In this age range, all mental abilities seem to decline for most people. Moreover, this research on the oldest old has also identified a pattern of age-related increases in the correlations among measures of cognitive functions, sensory-motor functions, and general health between ages 70 and 100. This phenomenon, often described as a dedifferentiation of cognitive functions (Baltes and Lindenberger, 1997), may be explainable in part in terms of change in sensory-motor and health status, as discussed in the next section.
Modulation of Cognitive Aging by Experience
The trajectories of cognitive aging are not the same for all individuals, even for highly specific cognitive functions. Life experiences matter.
Behavioral and Cultural Factors
As noted in Chapter 3, there are systematic differences in cognitive status among aging members of social groups defined by such factors as socioeconomic status, education, occupation, culture, race, and ethnicity. A mechanism that may explain some of these differences is expertise, resulting from training and practice. Studies comparing experts in chess, aviation, typing, and piano playing report that cognitive functions based on expert knowledge are preserved in old age. Effects of expertise, however, are highly specific. Tasks that imitate areas of practice show preservation of function; tasks that measure general cognitive functioning do not. For example, Krampe and Ericsson (1996) reported standard age differences in general processing speed for expert and amateur pianists, but no age differences for experts in speed of finger movements while playing. Practice among experts was essential for the maintenance of skills. Thus, lifelong habits or practice may produce structural or process changes in the brain that are protective against decline of the practiced functions. Indeed, there is a growing body of evidence in neuroscience demonstrating that experience can produce changes in brain organization, for example, by expanding or reorganizing the functional region associated with a highly practiced behavior (e.g., Squire and Kandel, 1999; Gilbert and Wiesel, 1992; Buonomano and Merzenich, 1998; Greenough, 1976; Elbert et al., 1995; Pascual-Leone and Torres, 1993).
Cognitive performance among older adults is also associated with a variety of noncognitive experiential factors, including tobacco use, alcohol con-
sumption, diet, intake of antioxidants, and levels of physical activity (see Waldstein, Appendix E). Further investigation of the effect of the mechanisms that link experiential factors on cognitive aging promises to increase our understanding of the mechanisms involved in age-related cognitive changes.
Connections between Cognitive and Sensory-Motor Changes
Considerable evidence exists that peripheral sensory and motor systems decline with age. The sensory input to an older nervous system is attenuated, spatially blurred, and temporally smeared compared with that of a healthy young sensory system. The motor output of the aged individual is also often slower, less agile, and more variable than that of a younger person. Evidence is mounting that there is a correlation between sensory-motor decline and cognitive decline—that much of the interindividual variation in cognitive decline is associated with sensory-motor decline.
Much of the decline in sensory function with age is due to ''wear and tear" on these systems (see Kline and Scialfa, 1997; Schneider and Pichora-Fuller, in press, for reviews). For instance, the hair cells of the inner ear, which transduce sound vibrations into neural impulses, tend to wear out over repeated use, leading to loss of hearing sensitivity and poor ability to resolve frequency differences, which are the major hearing problems associated with aging (presbycusis). Once mammalian hair cells are damaged or lost, they cannot be replaced (see Tsue et al., 1994). Hair cells similar to those in the auditory system are sensory transducers for the vestibular system that controls balance and helps in locomotion. These also deteriorate over time. In vision, structural changes in the tissues of the cornea, iris, lens, and their fluids occur over time, one result being that less light reaches the retina for neural transduction. Similar forms of structural decline have been documented for taste, smell, the sense of balance, and proprioception. For these reasons, the information most older people receive through their senses is degraded relative to that received by younger people. Even when corrections are made (e.g., hearing aids, eyeglasses) for some forms of sensory loss, older people still tend to perform poorer on perceptual and attentional tasks than do younger people, especially in terms of temporal processing.
Strong correlations have been reported among declines in sensory, motor, and cognitive function. Mayer and Baltes (1996, see also Baltes, 1997; Baltes and Lindenberger, 1997) developed a measure of general intelligence as a part of the Berlin Aging Study. This measure of intelligence correlates strongly with age (from 70 to 105 years) with more than 32 percent of the variance accounted for. When simple measures of hearing, vision, and motor balance are treated as covariates, the correlation drops so that almost none
(less than 0.4 percent) of the variance in intelligence can be accounted for on the basis of age. Salthouse et al. (1998) also found strong correlations of cognitive and noncognitive measures with age, but also found some differences between the changes in cognitive and noncognitive measures.
Four explanations have been suggested for the correlated trajectories of cognitive and sensory-motor changes with age: (1) sensory-motor decline causes cognitive decline, perhaps by increasing central control of sensory-motor function; (2) cognitive decline adversely affects sensory performance; (3) secondary variables that covary with sensory-motor and cognitive decline are the basis for the correlation; and (4) sensory, motor, and cognitive systems are interrelated parts of a single system, so that as one part declines so must all other parts. It is important to discover which is correct. For instance, if sensory-motor decline causes cognitive decline, then aids to sensory-motor performance might protect against, or even reverse, cognitive decline.
Hypotheses about common causes of sensory-motor and cognitive decline in old age call into question the common practice of treating sensory, motor, and cognitive systems as discrete systems underlying behavior (see Schneider and Pichora-Fuller, in press). They suggest that, contrary to the classic view in which sensory systems feed information to cognitive systems, which then dictate action to motor systems, the neurobiological substrate of behavior in old age is a single system with many interacting and highly overlapping subsystems. In this view, sensory transducers, hippocampal cells, and motor neurons are different points within a single system, and any change to one part of the system is highly likely to affect other parts.
In pursuing common cause hypotheses for correlations between sensory-motor and cognitive decline, genetics may provide some of the possible common causes. The studies of the role of APOE genotype in Alzheimer's disease show the potential of investigating such genetic variables. In addition, heritability studies of cognition in older adults offer a window into the phenomenon (e.g., McClearn et al., 1997; Johansson et al., 1999; Baltes et al., 1999). Other possible common causes may include patterns of life experience and somatic disease processes, as noted in the preceding and following sections.
Few studies have directly assessed the causal connections among sensory, perceptual, motor, and cognitive changes at any age. Moreover, existing knowledge is based mainly on laboratory tests of sensory, perceptual, and attentional abilities; much less is known about change in the activities of living as related sensory systems decline. There are important exceptions, however, such as the association found between useful field of view and car accidents in older adults (e.g., Owsley et al., 1998). A growing body of diverse literature models motor control as a function of environmental signals and internal models of task dynamics and temporal patterning (e.g., Ivry, 1996; Jordan, 1996; Levison, 1981). Also, recent research on the relations between motor and cognitive development in early life (e.g., Thelen and Smith, 1994;
L.B. Smith et al., 1999) suggests that motor control, because it involves problem solving, integration of multiple information sources, and the organization of dynamic internal representations, may be a productive model for understanding many aspects of cognitive development. Much is yet to be learned about how different neural systems (sensory, motor, cognitive) interact with one another and how those interactions change with age. This research may have fundamental implications for understanding the neurobiological basis of behavior in addition to improving understanding of cognitive aging.
The study of the relationship among cognitive, sensory, and motor decline requires a multidisciplinary approach. Investigators in the cognitive, sensory, and motor sciences have not typically collaborated, however. Such collaboration is needed to better understand the correlations among age-related changes in cognitive, sensory, and motor measures and their implications. A good example of the possibilities from research concerns age-related declines in hearing, which can have a significant impact on the lives of older people (Kline and Scialfa, 1997). Significant advances have been made recently in hearing-aid technology and in knowledge about attention, language comprehension, and cognitive processing, all of which seem to change in aging in correlation with auditory changes. If investigators studying audition work with those studying cognitive variables, such as language, the results might include better understanding of how these variables interact with age and innovations that might significantly improve older people's ability to communicate and function in their daily lives.
Cognitive Effects of Changes in Health Status
It has been proposed by some that age-related cognitive decline in many individuals is attributable to disease processes that affect the nervous system only indirectly, including cardiovascular diseases and diabetes, among others (Salthouse et al., 1990; Waldstein, Appendix E). Health problems negatively affect cognitive functioning and are more prevalent in the elderly (e.g., Perlmutter et al., 1988). In addition, the effects of neural changes on cognitive functioning may be moderated by health status or specific disease conditions.
High blood pressure is perhaps the most studied of the relevant health conditions. High blood pressure levels are adversely related to many neuropsychological measures of cognitive functioning (e.g., Elias and Robbins, 1991; Elias et al., 1990; Elias et al., 1993; for reviews see Waldstein, 1995; Appendix E). This relationship is quite apparent when normotensive subjects are compared with hypertensive ones, although there are moderating factors—for example, education has been reported to have a protective effect (Elias et al., 1987; see Waldstein, Appendix E). By some measures, blood pressure ac-
counts for half the variance in cognitive performance of older adults (Madden and Blumenthal, 1998; Elias et al., 1998). Cardiovascular disease, for which hypertension is an important risk factor, has also been shown to have a negative impact on cognitive functioning (e.g., Hertzog et al., 1978; Schaie, 1996). Various mechanisms have been proposed, including direct effects of elevated blood pressure, indirect effects through stress-induced cardiovascular and neuroendocrine responses, and third-variable explanations in which common genetic or environmental factors may predispose an individual both to hypertension and to cognitive decline (see Waldstein, Appendix E). Different explanations may apply for different subgroups of hypertensive individuals or at different points in the life span.
Diabetes is among the other disease conditions that appear to be related to cognitive functioning. Most research suggests that diabetes is related to such measures of cognitive functioning as verbal memory, sensory-motor speed, cognitive flexibility, and concept formation (Dey et al., 1997; Elias et al. 1997; Mochizuki et al., 1998; Naor et al., 1997; van Boxtel et al., 1998), although there is also evidence to the contrary (Muqit and Ferdous, 1998; Worrall et al., 1996).
Numerous possible causal mechanisms may explain correlations between disease conditions and particular types of cognitive functioning, and research is just beginning in this area. In addition to having some potential to illuminate some of the causes of cognitive decline, this line of research brings an added benefit: it can strengthen research on other aspects of cognitive aging by promoting controls for health conditions and interpretations that take such conditions into account.
In addition, there is a considerable body of research on the relationship between self-rated health and cognition (e.g., Field et al., 1988; Hultsch et al., 1993; Perlmutter et al., 1988; Perlmutter and Nyquist, 1990; Salthouse et al., 1990). For example, Perlmutter and Nyquist (1990) found that self-reported health accounted for a significant proportion of the variance in cognitive performance (e.g., digit span and fluid intelligence), even after age-related differences in health were statistically controlled. These associations may reflect cognitive effects of subclinical health conditions, although other explanations are also possible. Some researchers, however, have failed to find a link between self-assessed health and cognition (e.g., Salthouse et al., 1990).
Changes in mental health status, such as depression, may also influence cognitive functioning in older people. In addition to the association of depression with poor physical health (Wells et al., 1989) and elevated rates of mortality (Murphy et al., 1987), depressed individuals are also characterized by impaired cognitive functioning (Wright and Salmon, 1990). The assumption made by most researchers in this area is that the poor cognitive performance is related to other symptoms of depression, rather than representing a direct influence of depression on brain functioning. For example, the symp-
toms of fatigue, diminished energy and motivation, reduced cognitive effort, and increased rumination and self-focus all combine to impair learning and memory performance in depressed individuals (Gotlib and Hammen, 1992). Although there is some recent evidence indicating that depressed individuals actually demonstrate better memory for negative than for positive information (Gotlib et al., in press), it is clear nevertheless that depression interferes globally with cognitive functioning.
Improved explanations of the relationships of somatic disease to cognitive decline may reveal opportunities to use health interventions to improve cognitive functioning. Although only limited evidence exists of cognitive improvement resulting from specific health interventions (see Waldstein, Appendix E), the research is still in its infancy. Opportunities for intervention may be of particular importance for low-income and minority groups that have below-average use of health care services and high rates of chronic illnesses (black Americans, for example, have higher rates of hypertension, diabetes mellitus, and coronary heart disease) (Ferraro and Farmer, 1996; Harper and Alexander, 1990; Marquis and Long, 1996; Miles and Bernard, 1992).
As the above discussion indicates, various noncognitive indicators, including blood pressure, sensory-motor performance, and peak expiratory flow (Albert et al., 1995), account statistically for large proportions of the variance in rates of cognitive decline in normal aging. These proportions are so large as to indicate overdetermination, a condition in which several different factors appear to explain the same variations in cognitive performance. The strongly correlated trajectories of so many variables constitute an important puzzle for research on the structure of the aging mind: Which variables are causally prior to which? Which correlations reflect the operation of underlying common causes? Which correlations present opportunities for noncognitive interventions that can help preserve cognitive function?
Developments in Measurement and Theory
Growing sophistication in theoretical understanding of cognitive functions, advances in measurement of neural phenomena and cognitive functions, and the availability of analytic techniques from related fields are making possible new advances in explaining the patterns of cognitive aging and linking changes in cognitive function to changes in the brain.
IMPROVED MEASUREMENT OF NEURAL AND COGNITIVE PHENOMENA
Advanced neuroimaging and electrophysiological techniques for measuring on-line brain function, such as functional magnetic resonance imagery (fMRI), position emission tomography (PET), magnetoencephalography
(MEG), transcranial magnetic stimulation (TMS), electroencephalography (EEG), event-related brain potentials (ERP), and event-related optical signal (EROS), several of which so far are underutilized for the study of age-related cognitive changes, make breakthroughs possible in understanding brain-behavior links. Techniques such as fMRI and single-unit recording of neural activity are providing unprecedented levels of spatial and temporal resolution in observations of the brain, and other new and emerging techniques may hasten progress. Functional MRI, for example, can provide measurements in the brain with a time resolution of less than 1 second and spatial resolution of about 2 ram; technological advances promise further improvements in resolution (Le Bihan and Karni, 1995; Albright, Appendix G). Such techniques allow for much closer observation of neural phenomena than ever before, making possible much closer analysis of the relationships between neural and cognitive processes (e.g., Gabrieli, 1998).
Behavioral research has developed a rich array of laboratory techniques that isolate and measure specific mental operations that are fundamental to cognition. These techniques offer greater sensitivity and analytic power than traditional neuropsychological tests, which are designed to detect impairments but not to identify underlying processes and mechanisms. For example, experimental techniques have been used to demonstrate the distinct neural bases of implicit versus explicit memory in cognitive neuroscience research with patients (e.g., Gabrieli et al., 1995; Shimamura and Squire, 1984) and using imaging techniques (e.g., Uecker et al., 1997). The research demonstrates that implicit and explicit memory are differently affected by aging (Fleischman and Gabrieli, 1998; LaVoie and Light, 1994). Similarly, techniques for on-line evaluation of language comprehension processes have isolated semantic and syntactic processes required for comprehension (e.g., Marslen-Wilson and Tyler, 1980). These techniques have been central to evaluating the neural basis of comprehension in research with patients (e.g., Kempler et al., 1998) and using imaging techniques (Caplan et al., 1998, 1999). On-line measures of comprehension have provided evidence consistent with the maintenance of semantic comprehension processes in old age (e.g., Madden, 1988; Stine and Wingfield, 1994; see Burke, 1997; Light, 1991), although there is less agreement about syntactic processes (Caplan and Waters, 1999; Kemper and Kemptes, 1999; Stine-Morrow et al., 1996).
In animal research, tasks have been developed that are selectively sensitive to the effects of damage to the hippocampus, the amygdala, the caudate nucleus, the cerebellum, and the frontal cortex in rats. These tasks have since been adapted for the mouse and their research applications are being disseminated to researchers. For example, the Cold Spring Harbor Laboratories established an annual course in mouse behavior in 1998.
Because of the variety of behavioral tasks that are available, it is increasingly possible to use behavioral observations to identify specific brain regions
in which age-related changes that affect cognition are occurring. This information will be important to future efforts to intervene at the molecular and cellular level, because it will tell where in the brain to look for the substrate of observed behavioral effects.
As neuroscience research attains higher levels of resolution, it will become possible to identify particular neural circuits believed to be associated with particular cognitive functions. To achieve understanding of brain-behavior links at this level, it is important to build theory and to identify or construct behavioral measures that fit the structure of cognition and can be localized with comparable resolution to neural observations. This implies a search for fine-grained measures of specific cognitive processes and of the operation of particular neural systems. It also implies a continuing co-evolution of behavioral measures and brain measures and a continuing effort to refine both kinds of measures in order to further clarify brain-behavior links.
An important point to recognize with regard to the above issues of measurement is the central role of experimental animals in research on aging. Neural observations are more feasible in animal models, and many findings are likely to generalize across species. Biological research during the past decade has shown the extraordinary extent to which cellular and molecular mechanisms are conserved through evolution. Indeed, even at the level of brain systems and brain-behavior relationships, one finds considerable parallel across species. Recent advances in molecular biology have caused the mouse to become important for behavioral studies. At the present time, one can expect useful work on brain and behavior, in the context of aging research, to be carried out in the mouse, rat, and monkey, and perhaps in other animal species as well.
Advances in Theory Development
Cognitive behavioral science is making progress on the theoretical side. An example is Baddeley's influential theory of working memory, which postulates interrelated components for speech and visual information with separate storage and rehearsal mechanisms. This theory has motivated investigations of working memory in patients (Vallar and Baddeley, 1984) and using imaging techniques (e.g., Smith and Jonides, 1997) that have supported distinctions in the theory and have identified the neural basis for hypothesized working memory mechanisms. A number of studies following the theory have investigated age differences in components of working memory (e.g., Wingfield et al., 1995).
Research based on theoretically justified measures of specific cognitive functions will make it possible to examine more closely the links between functioning of particular neural circuits and performance on the behavioral indicators, as well as between the behavioral indicators and performance of
life tasks. This line of research can go far to clarifying the mechanisms linking age-related changes in neural circuitry to change in cognitive functioning and performance. It is likely also to identify opportunities to intervene either at the neural or behavioral levels so as to maintain performance of life activities in the face of neural decline.
Underutilized Analytical Techniques
A number of mathematical techniques that can be used to characterize the structure and evolution of behavior over time are maturing to a point at which they may be of great benefit in the study of cognitive aging. Four of these are dynamical systems theory, hidden Markov models, connectionist models, and dynamic factor analysis.
Dynamical systems theory characterizes the properties of different patterns of stable behavior, as well as transitions among such patterns. In this approach, behavior is typically represented as a continuous trajectory in a state space, i.e., a space whose dimensions are the important variables needed to describe and predict behavior. Stable behavior is not modeled as simply a constant, but is viewed as resulting from the interaction of various abstract forces and perturbations to a behavioral system. These forces may push a system toward a point in the state space, toward a particular oscillation or limit cycle, or toward some more elaborate pattern of stable behavior. Some perturbations can be compensated for by the system; others result in a loss of stability, and perhaps achievement of a new stable state. This abstract, holistic style of description has been applied to various types of behavior ranging from physiological subsystems (e.g., heartbeat; breathing; see Glass and Mackey, 1988) to coordinated limb movements (e.g., Kelso, 1995). In the developmental domain, this approach has been used to describe changes in perceptual-motor performance in children. It has shown how particular changes in underlying behavioral dynamics combine with each other and with characteristics of the environment to produce stable behavioral patterns (Thelen and Smith, 1994).
A number of intriguing possibilities exist for applying this mathematical approach to older adults. For example, it might be applied to developmental declines in the ways it has been applied to developmental advances. It could clarify the implications for overall performance of declines in particular behavioral dynamics and identify specific types of remediation or environmental modification that could maintain performance despite such declines. Also, because dynamical systems theory characterizes the stability of behavior in terms of a process rather than simply approximating it as a constant, it allows the examination of varieties of stability that differ both qualitatively and quantitatively. Some patterns may be too stable (i.e., rigid) and inhibit adaptive changes in behavior; other patterns may be insufficiently stable and result in
loss of control. In other words, there may exist optimum levels and types of stability for specific functions (e.g., Beek, 1989; Thelen, 1999). Also, because dynamical systems theory is applicable across many realms of measurement, it may facilitate comparisons of adaptivity at the neural and behavioral levels.
Hidden Markov models typically represent behavioral patterns as a network of discrete states with various probabilities of transition to subsequent states. Such models may be particularly helpful in designing technological interventions for processes that exhibit discrete transitions. For example, a person's interactions with a computer-controlled device might be characterized by the values of a limited set of discrete state variables corresponding to the physical state of the device and the cognitive state of the person (see Fisher, Appendix D). Interactions that can be characterized by finite sets of possible external inputs and available actions can be thoroughly analyzed in terms of transitions from the current state into some new state. The pattern of state transitions will differ across individuals and may provide useful information in diagnosing problems that a particular individual is having in controlling a device (e.g., Miller, 1985; Fisher, Appendix D). The Markov network may indicate what subgoals a person is trying to achieve and what strategies are being used. This information may provide the basis for designing appropriate changes in the device structure and/or computer prompts to adaptively improve the action patterns of older adults. Similar techniques may also be applicable to more continuous control tasks, such as driving a motor vehicle, to describe the transitions between different discrete subgoals in continuous movement patterns and/or co-occurring discrete subtasks that accompany continuous control of vehicular movement (e.g., Baron and Corker, 1989; Levison, 1993). Other modeling techniques involving Bayesian inference and/or neural networks (e.g., Jacobs and Jordan, 1993; see below) may also be useful for modeling how complex tasks are partitioned into subtasks and/or the effects of multiple competing goals on action. All of these techniques address the temporal microstructure of action.
Connectionist models have shown a dramatic resurgence of interest during the 1980s and 1990s. These models include various approaches known as neural networks, neural models, parallel distributed processing systems, localist models, and spreading activation models, and have in common a network of nodes connected by weighted pathways. Connectionist models are sensitive to biological constraints, as nodes can be loosely associated with neurons and their connections with synapses. Connectionist models have been used to simulate brain processes, for example, topographic map formation in the brain (Reggia et al., 1992), the development of receptive fields (Linsker, 1986), and the physiological basis of EEG (Lagerlund and Sharbrough, 1988). At the psychological level, connectionist models have been used to simulate cognitive operations, such as associative memory (Anderson, 1983) and sentence production (Dell, 1986). These models, developed to
account for normal brain and cognitive functions, have been widely used as the basis for modeling brain disorders (for example, focal cortical lesions and disconnection syndromes), as well as cognitive disorders (for example, dyslexia, amnesia, and aphasia) (Reggia et al., 1994). In models of cognitive disorders, patterns of symptoms or deficits emerge from damage to the normal system, and these symptoms are compared with the performance of patients on cognitive tasks.
The connectionist approach has been extremely useful in cognitive research, producing better theoretical understanding of brain function and cognitive processes and generating hypotheses about mechanisms underlying important cognitive behaviors. It has led to the identification of fundamental principles of cognitive processing, such as the necessity of parallel processing to achieve the computational power required for cognition, and it has provided a means for both generating and evaluating hypotheses about the functional deficits that underlie cognitive disorders. However, the connectionist approach has had little impact so far on cognitive aging research. Behavioral research on cognitive aging has produced an accumulation of rich datasets, but theory is insufficient to organize and explain them. Connectionist models have been developed for a wide range of cognitive behaviors and impairments' and offer a promising approach for identifying mechanisms that can explain existing data on the pattern of cognitive functioning in old age and generate hypotheses for future research.
Dynamic factor analysis provides an expansion of cross-sectional factor analysis, which is normally used to identify a small set of distinct behavioral variables that clarifies the pattern underlying a larger number of measures. With dynamic factor analysis, these variables can be analyzed over time in ways that explore temporally lagged relationships among them. Because of the demands of repeated behavioral measurement, dynamic factor analysis is typically used to analyze changes occurring on time scales of days to months. Dynamic factor analysis has been productive for studying temporal variability within individuals in such personality characteristics as perceived control of one's environment (e.g., Eizenman et al., 1997) and emotional positivity and negativity (e.g., Shifrin et al., 1997). The finding that intraindividual variability in perceived control correlates with mortality in older adults over a five-year period (Eizenman et al., 1997) suggests that the method is useful for studying relationships involving life experience variables. Dynamic factor analysis may be useful for analyzing cognitive and behavioral adaptations in aging individuals that occur over appropriate time scales, such as in response to dementia, stroke, or stressful life events (e.g., loss of a spouse, confinement in a hospital or nursing home).
Modeling techniques such as these are important not only for their potential ability to represent particular cognitive phenomena. They may also offer first steps toward the development of process models that would im-
prove understanding of the chains of causation—probably quite complex—that link life experiences, physical and mental health, social and technological context, sensory-motor phenomena, and neural changes to the life course of cognitive capability and performance. It is also important to emphasize that progress in the modeling of cognitive processes depends on continued improvement in understanding of those processes that comes from improved psychometric measures and from experimental research on perception and cognition.
RESEARCH INITIATIVE ON THE STRUCTURE OF THE AGING MIND
The NIA should undertake a major research initiative to improve understanding of the structure of the aging mind, including the identification of mechanisms at the behavioral and neural levels that contribute to age-related change in cognitive functioning.
Research has established that the effect of aging on cognitive functioning and performance of associated life tasks varies for different cognitive operations and for different individuals. It has also revealed that in advanced old age, the correlations increase among measures of different cognitive functions. Furthermore, research has linked a variety of experiences to cognitive performance in older adults, including physical exercise, diet, cognitive training, expertise, and the provision of environmental support; it has also documented relationships of cognitive functions to body systems, including sensory-motor functioning and chronic diseases. The mechanisms underlying these relationships are not yet known; hypotheses are still at an early stage of development.
The recommended research initiative would aim to specify the patterns of variation in cognitive functioning during the aging process and to identify the mechanisms, at levels of analysis from the molecular to the cultural, that contribute to age-linked stability and change. By identifying these mechanisms, the initiative would contribute to the search for effective interventions to assist older adults in maintaining cognitive functioning and performance. For instance, it might document benefits from diet, exercise, cognitive activity, and other interventions and determine whether certain of these have more general or more lasting effects than others.
The research initiative should include studies of the full variety of phenomena identified in this chapter as well as related ones. It should build on a base of psychometric and experimental research on cognitive processes, which should continue to receive NIA support. We believe, however, that the most rapid advances are likely to result from encouraging researchers to expand their portfolio of research approaches by applying and integrating promising methodologies that are either new or underutilized in cognitive aging re-
search, along with continuing research with well-established methods. We therefore recommend that the initiative emphasize three method-based research strategies and their integration.
1. Relating high-resolution measures of neural functioning to measures of cognitive functioning in the aging mind.
The research initiative should support investigator-initiated research that will measure and analyze the reciprocal relations between brain and cognition. This research should utilize the new high-resolution techniques for measuring neural functioning and should link these observations to measures of cognitive functioning that are capable of isolating specific mental operations. The research will investigate age differences in the cortical components and behavioral indicators of cognitive processes as well as the effects of interventions and experience on cortical organization and behavior.
For this research approach to achieve its potential, the research initiative should support studies that address key methodological issues associated with brain imaging techniques. One such issue that requires immediate research is the decoding of the vascular signal of fMRI in ways that allow meaningful comparisons across the life span. As is well known, fMRI signals measure neural function indirectly by measuring blood flow; because blood flows differentially to active regions of the brain, the fMRI is presumed also to reflect neural activity (Le Bihan and Karni, 1995). However, the precise correlation between neural activity and the vascular signal is not known, and it may vary with age because of atrophy and other neuronal changes related to aging, as well as changes in vascularization (D'Esposito et al., 1999; Taoka et al., 1998). Eventually, longitudinal studies with humans may aid interpretation. For the next decade or more, however, the best source of insight may be information from animals, particularly the awake, behaving monkey, in which the relationship between single unit activity and the fMRI signal can be directly explored in the same tasks (Albright, Appendix G). Recently, the feasibility of this approach was demonstrated in the monkey (Stefanacci et al., 1998), and an improved technique for doing such work was demonstrated in monkeys using a custom-designed, vertical bore magnet that allows the monkey to sit in a conventional primate chair within the magnet (Logothetis et al., 1999).
Another methodological problem deserving immediate attention concerns ways to improve the correspondence between behavioral measures and neural observations, such as those that new measurement techniques can provide. Recent advances in the behavioral measurement of specific cognitive functions have already been noted. The research initiative should support studies to identify or develop focused behavioral indicators that connect closely to high-resolution neural observations. It is to be expected that simple behaviors will yield more easily to this approach than complex ones, which are more likely to involve distributed neural activity.
This line of research requires two kinds of knowledge that rarely exist in the same investigator: knowledge of cognitive techniques for isolating mental operations and of neuroimaging or electrophysiological techniques for measuring on-line brain function. We therefore recommend that the NIA support workshops to provide basic training to investigators. The goal of the workshops would be to provide training on techniques for measuring and decomposing phenomena at both the cognitive-behavioral and the neurological levels in order to encourage investigators to develop relevant research applications. We also recommend that the NIA support research to develop or modify behavioral and neural measures so as to improve the correspondence between these types of measurement.
2. Elaborating theory-based and mathematical models of the aging mind.
The research initiative should support theory-driven research that develops models of cognitive aging effects that will increase our understanding of patterns of stability and change among mental processes in the aging mind, as well as complex interactions among cognitive and other systems. Existing findings from cognitive aging research provide a wealth of data that demonstrate both impaired and preserved cognitive functions in old age. Theories to explain this pattern, however, have been slow to develop.
This research strategy should take advantage of theoretical and mathematical tools that have proved useful in related fields. To illustrate, connectionist models of language and knowledge representation and production system models of executive functions and memory have been very influential in improving understanding of basic cognitive processes; they can be productively extended to account for age-related changes in these processes. Also, research should be encouraged that applies various types of statistical and mathematical models (e.g., structural equation models, dynamic factor analysis, Markov models, dynamical systems models, adaptive control theory) to understanding short-term variability, long-term stability and change, and multisystem causal linkages involving or affecting cognitive functioning in older adults (e.g., sensory-motor functioning and cognition).
3. Conducting and analyzing large-scale, multivariate studies of the aging mind.
To achieve the objectives of this research initiative, it will be necessary to expand the use of large-scale, multivariate, longitudinal studies. It is necessary to expand and improve on previous longitudinal research by including variables reflecting high-resolution cognitive and neural measures; indicators of health status and sensory-motor functioning; and measures of relevant life experience. Analysis of multiple measures can help explain patterns of correlations, such as that in which several different physiological variables appear
to explain the same variations in cognition. It is also important to examine a broad representative sample of the population, sometimes oversampling in subgroups whose health status or responses to life experiences are expected to illuminate important theoretical questions, and to encompass a wide age range. Moreover, by following individuals into very old age, promising new findings suggesting the existence of unexpected linkages between cognitive functioning and survival could be investigated. The requirements for long-term longitudinal studies are discussed in more detail in Chapter 5.
The conjunction of improved measurement, advances in modeling, and the comprehensive collection of longitudinal data on cognitive functioning and associated factors can have a synergistic effect in advancing knowledge. Longitudinal studies can make a quantum improvement by employing new fine-grained neural and behavioral measures; renewed attention to modeling can make better sense of the patterns that underlie associations between these neural and behavioral measures and that can be drawn from the longitudinal data.