Journal of Gerontology: PSYCHOLOGICAL SCIENCES 2006, Vol. 61B, No. 6, P348–P354
Copyright 2006 by The Gerontological Society of America
Effects of Repeated Testing in a Longitudinal Age-Homogeneous Study of Cognitive Aging Valgeir Thorvaldsson,1 Scott M. Hofer,2 Stig Berg,3 and Boo Johansson1 2
1 Department of Psychology, Go¨teborg University, Sweden. Department of Human Development and Family Studies, Pennsylvania State University, University Park. 3 Institute of Gerontology, Jo¨nko¨ping University, Sweden.
Estimates of gains related to repeated test exposure (retest effects) and within-person cognitive changes are confounded in most longitudinal studies because of the nonindependent time structures underlying both processes. Recently developed statistical approaches rely on between-person age differences to estimate effects of repeated testing. This study, however, demonstrates how retest effects can be evaluated at the group level in an agehomogeneous population-based study by use of a sampling-based design approach in which level and change of cognitive performance of previous participants, measured at ages 70, 75, 79, 81, 85, 88, 90, 92, 95, 97, and 99 years, were compared with performances of survivors of a representative sample identified and drawn from the same original population cohort but invited for the first time at age 85 with subsequent measurements at ages 88, 90, 92, 95, 97, and 99. The comparisons revealed a trend toward retest effects on two out of five cognitive measurements. The study demonstrates how a design-based approach can provide valuable insights into continuous learning processes embedded in population average aging trajectories that are not confounded with cohort and mortalityrelated selective attrition.
E
STIMATES of cognitive change at the within-person level require repeated observations of cognitive function over time. Effects of repeated testing (i.e., retest, practice, warm-up effects) are often confounded with developmental or agingrelated change processes, because these distinct causal processes are usually indexed by identical samples of time (i.e., test intervals). Schaie (1965) and Baltes (1968) identified the essential confound between repeated testing and age within longitudinal designs having widely spaced assessments. Recently, however, several attempts have been made to statistically disentangle and estimate cognitive changes and retest effects simultaneously (e.g., Ferrer, Salthouse, McArdle, Stewart, & Schwartz, 2005; Ferrer, Salthouse, Stewart, & Schwartz, 2004; McArdle, Ferrer-Caja, Hamagami, & Woodcock, 2002; McArdle & Woodcock, 1997; Rabbitt, Diggle, Holland, & McInnes, 2004; Rabbitt, Diggle, Smith, Holland, & McInnes, 2001). These studies have defined retest effects as the number of occasions an individual has been exposed to a test at a particular time point, and they have relied on between-person age differences for identification of separate parameters for the estimation of both retest and ‘‘aging’’ effects. This can, however, be problematic, not only because retesteffect estimates are confounded with within-person aging changes, but also because age-heterogeneous differences do not converge to within-person changes because of other notable factors other than retest gains (see, e.g., MacDonald, Hultsch, Strauss, & Dixon, 2003; Sliwinski & Buschke, 1999). Other explanations for lack of convergence include cohort effects, selective attrition, population mortality selection, and the age range of the sample. Researchers have proposed alternative design-based approaches for the examination of retest effects. For example, repeated-measurement burst designs hold aging effects relatively constant within short-term intensive measurements and permit unbiased estimates of short-term retest gains (Hofer,
P348
Hoffman, Sliwinski, & Piccinin, 2005), and cross-sectional group comparisons across cognitive performances of previously assessed participants and a randomly assigned post-test control group allow these estimates at the group level (Baltes, 1968; Schaie, 1965, 1988). Our main objective in this study was to examine this latter design-based approach for the estimation of retest effects by taking advantage of the multiple sample, single-population-based design of the Gerontological and Geriatric Population Study (known as the H70 study).
METHODS
Participants and Sampling Design in the H70 Study The H70 study is an age-homogeneous population-based longitudinal aging study that started in 1971, when researchers identified all individuals living in the city of Go¨teborg, Sweden, in the birth cohort from July 1 1901 through June 30 1902 from the Swedish Revenue Office Register, at the approximate age of 70 years. Researchers then randomly selected approximately one third of this population, 1,148 individuals, for participation in the study (i.e., previous participants). From this group of participants, 460 individuals were randomly selected for a psychometric evaluation. The participation rate in this part of the study was 85%, leaving a sample of 392 individuals who participated at baseline, age 70, and subsequent follow-up at ages 75, 79, 81, 85, 88, 90, 92, 95, 97, and 99, conditional on mortality and occasional dropout. In 1986, researchers invited about two thirds of the original population who were not drawn for participation at baseline for participation and then randomly selected a subsample of 657 individuals for psychometric testing at ages 85, 88, 90, 95, and 99 (i.e., new participants). The participation rate was 60%, leaving a sample of 395 participants at baseline.
EFFECTS OF REPEATED TESTING
Table 1. Demographic and Health Variable Characteristics for Participants at Age 85 Variables Sex (% female)
Previous Participants New Participants (n ¼ 74) (n ¼ 268) 71.6
69.8
71.8 28.2
74.2 25.8
17.8 9.5
25.0 10.1
9.6 72.6 17.8 0.0
12.5 72.5 14.3 0.7
8.6 5.7 2.9
6.8 9.1 6.4
42.98 (17.05) 41.21 (15.95)
43.19 (17.26) 40.24 (17.62)
157.64 (20.91) 81.28 (12.08)
162.56 (24.96) 80.24 (12.29)
Education (%) 6 years .6 years Marital status (% married) Type of residence (% in institutions) Self-rated health (%) Very good Good Bad Very bad Diabetes (% diagnosis) Heart attack (% diagnosis) Stroke Vigorimeter grip strength [M (SD)] Right hand Left hand Mean blood pressure Systolic Diastolic Note: SD ¼ standard deviation.
This sampling design permits estimates of retest effects at the group level by comparing cognitive performances across groups of previous participants and new participants at age 85 and subsequent changes thereafter. Note that agerelated cohort effects and population mortality dynamics are held constant in this design because of the age homogeneity of the sample. In addition, selection biases are minimized in this design because of the population-based nature of the sample. We can then evaluate possible biasing effects caused by selective attrition and age of first invitation to the study by making comparisons of between-group homogeneity on various demographic and health measures. In this research, we used the Swedish Person Number System to randomly select the sample, to facilitate recruitment, and to retain contact with participants at subsequent measurements occasions. For the purpose of the present analysis, we excluded all individuals diagnosed with dementia according to the revised third edition of the Diagnostic and Statistical Manual of Mental Disorders (American Psychiatric Association, 1987) before measurement at age 85. In addition, we restricted all analyses to individuals that had at least one measurement occasion on the respective cognitive test at age 85. The number of participants and group differences on various demographic and healthrelated variables at age 85 are presented in Table 1; they indicate small but consistent differences in the direction of retention and selection effects on variables related to marital status, self-rated health, disease status, and physiological biomarkers such as grip strength. More detailed information about the procedure and recruitment in the H70 study can be found elsewhere (e.g., Rinder, Roupe, Steen, & Svanborg, 1975; Svanborg, 1977).
P349
Testing Procedure In the present study we used cognitive measurements from the psychometric battery of Dureman and Sa¨lde (1959), which was widely used in Sweden at the beginning of the H70 study. The Synonym test and both Digit Span tests were administered throughout all the measurement occasions, whereas the Block Design and Identical Forms tests were omitted when participants were at age 81. The same tests and a similar testing procedure were employed throughout subsequent measurement occasions. We excluded additional psychometric tests in the battery from the analyses because these tests were not administrated after participants reached the age of 85. Further information about these tests, their statistical properties, and usage in the H70 study can be found in Berg (1980). Synonyms test. —This test measures an individual’s ability to understand ideas expressed in words. Participants match a target word with one synonym among five choices. The test has a time limit of 7 minutes and a maximum score of 30 points. The words were presented in a magnified form if participants have problems reading the words as the result of a visual impairment. Block Design test. —This test is a measure of spatial ability. In this test the participants are presented with two color blocks. The task is to construct replicas of model designs that are presented to the participants. The test has a time limit of 20 minutes and a maximum score of 42 points. Figure Identification test. —This is a test of perceptual speed. In this test the participants match, as quickly as possible, a target figure with one identical figure that is placed in line among five others. The test has a time limit of 4 minutes and a maximum score of 60 points. Digit Span Forward and Backward tests. —These two tests measure short-term memory and working memory. Both tests consist of random number sequences that the examiner reads aloud at the rate of one per second. In the Digit Span Forward test, the participant’s task is to repeat the sequence exactly as the examiner reads it. When the sequence is repeated correctly, the examiner reads the next longer sequence of digits until the patient fails on the task or repeats a nine-digit sequence correctly. The maximum score in this test is 9 points. In the Digit Span Backward condition, the participant’s task is to repeat the digit sequence in reverse order. The test continues until the participant fails a sequence or correctly recalls eight digits. The maximum score on this test is 8 points.
Analysis We fit piecewise (i.e., multiple slope) linear mixed models (e.g., Naumova, Must, & Laird, 2001) to the data with a dummy-coded time-invariant covariate indicating group differences (0 ¼ previous participants; 1 ¼ new participants) on intercept and linear change functions after participants reached the age of 85. We centered the intercept in all models
THORVALDSSON ET AL.
P350
Table 2. Summary of Piecewise Mixed Linear Models Identifying Pre- and Post-Age 85 Changes and Group–Retest Effects on Level of Performance at Age 85 and Post-Age 85 Changes Cognitive Test Synonym
Block Design
Figure Identification
Digit Span Forward
Digit Span Backward
Figure 1. Raw-score change plot trajectories for previous and new participants on the Synonyms test (verbal ability), from age 70 to 99. New participants were included into the study at approximately the age of 85.
at age 85. We determined the significance of fixed-effect parameters in nested models by use of a likelihood-ratio (LR) chi-square test based on full maximum likelihood estimation. The final estimated model and modeling procedure was as follows: yit ¼ b00 þ b10 ðpre-age 85it 85Þ þ b20 ðpost-age 85it 85Þ þ b01 ðgroupi Þ þ b11 ðgroupi 3 post-age 85it 85Þ þ ½c0i þ c1i ðpre-age 85it 85Þ þ c2i ðpost-age 85it 85Þ þ eit ; where yit is the estimated score on a certain test for individual i at time t,b00 is the average (fixed effects) intercept, b10 is the estimates average age-based linear change slope before age 85 and b20 is the average age-based linear change slope after age 85, b01 is the group covariate affecting the intercept, and b11 is the group covariate affecting the age-based linear change slope after age 85. Parameters that are bracketed in the third line of the equation indicate estimated random effects expressed as variation components of the intercept, and age-based linear change slopes before age 85 and age-based linear change slopes after age 85; eit refers to residual estimates. Our main emphases in the present analyses are on the group covariates estimates, because they allow direct identification of retest effects at the group level. Otherwise, we conducted the modeling procedure sequentially, beginning with (a) an unconditional means model; (b) inclusion of fixed and random linear slope functions after age 85; (c) inclusion of fixed and random linear slope functions before age 85; and (d) inclusion of dummy covariates for estimation of group differences on
Model Term Pre-age 85 Post-age 85 Group Group 3 Post-age Pre-age 85 Post-age 85 Group Group 3 Post-age Pre-age 85 Post-age 85 Group Group 3 Post-age Pre-age 85 Post-age 85 Group Group 3 Post-age Pre-age 85 Post-age 85 Group Group 3 Post-age
85
85
85
85
85
Estimatea
SE
— 0.44*** 2.47** — 0.38*** 0.51*** 1.68* — 0.51*** 0.51*** — — 0.03** 0.058*** — — 0.02* 0.057*** — —
— 0.076 0.959 — 0.057 0.069 0.803 — 0.063 0.069 — — 0.010 0.014 — — 0.009 0.013 — —
Notes: SE ¼ standard error. Pre-age 85 ¼ linear age-based change slope before age 85; Post-age 85 ¼ linear age-based change slope after age 85; Group ¼ dummy coded covariate at intercept centred at age 85 (0 ¼ previous participants; 1 ¼ new participants). The em dash refers to a nonsignificant model parameter that has been constrained to zero, determined by a full maximum likelihood function estimate. a Random effect estimates from these models can be found in the Appendix. *p , .05; **p , .01; ***p , .001.
intercept and linear slope function after age 85. We constrained parameters indicating nonsignificant improvement in fit to zero in subsequent models.
RESULTS In order to demonstrate the age homogeneity in the analyzed sample, and the type of data the models were fitted to, we give an example of individual raw-score change trajectories in Figure 1 for one of the test (Synonyms test). Note that there is a limited value for an analysis of age-heterogeneous convergence modeling with these data, which is one of the reasons why the statistical retest-effect models are not applicable to the H70 study. In addition, these raw-score change trajectories indicate that we should expect relatively limited group differences on the Synonyms test in the intercept estimates, when participants were at age 85, and linear change functions after age 85. A summary of the fixed-effect estimates in the final models is presented in Table 2, and fixed effects are plotted in Figures 2– 6. Random-effect estimates in these models can be found in the Appendix. Retest effects identified by group differences on intercept and linear change slope function after age 85 were significant for Synonyms test and Block Design intercepts. New participants had, on average, a 2.47-point lower estimate at age 85 compared with previous participants on the Synonym test and a 1.68-point lower score on the Block Design test. Other retest estimates were nonsignificant, indicating average group similarity in level and change function. As we expected, postage 85 change estimates indicated steeper average decline
EFFECTS OF REPEATED TESTING
P351
Figure 2. Plotted fixed effect based on the final model for the Synonym test (verbal ability).
Figure 4. Plotted fixed effect based on the final model for the Figure Identification test (perceptual speed).
compared with pre-age 85 changes. The only exception to this pattern of findings was on the Figure Identification test, where the change estimates were similar. Interestingly, no average changes were identified before age 85 on the Synonym test measuring verbal ability. These changes were, however, relatively steep, indicating decline on the fluid ability measurements such as Block Design and Figure Identification tests measuring spatial ability and perceptual speed. These findings support previous studies that have suggested that crystallized abilities are relatively intact in a healthy aging population but that fluid abilities decline according to normal aging process. Change estimates on the Digit Span tests were relatively small, as we expected from their psychometric properties (small range and distribution) and the homogeneity of the sample in terms of health and other demographical variables.
DISCUSSION The present study illustrates how population-level retest effects can be evaluated relative to within-person cognitive
aging change estimates with a multiple-sample design-based approach (e.g., Baltes, 1968; Schaie, 1965). In the framework of many existing, ongoing, and future longitudinal cognitive aging studies, the processes of retest effects and within-person cognitive changes are unlikely to be separable at the withinperson level. This is because the estimation of cognitive change requires repeated observations, and the act of observing may alter the magnitude and nature of the process of interest. The results from the present study of group-based comparisons of previous and new participants in the H70 study indicated a relatively limited effect of retest on performance level of two of the five tests. A possible explanation for the observed test difference is that the test material that is presented to the participants in the Synonyms test and Block Design test are more likely to be remembered compared with the relatively simple procedures in the Figure Identification and both Digit Span tests. A possible explanation for the lack of retest effects in the present analysis refers to the advanced age interval of the estimated sample (i.e., Age3Retest interaction). Healthy older individuals
Figure 3. Plotted fixed effect based on the final model for the Block Design test (spatial ability).
Figure 5. Plotted fixed effect based on the final model for the Digit Span Forward test (short- term memory).
P352
THORVALDSSON ET AL.
can be expected to acquire new information, although to a lesser degree than younger individuals. Inferences and generalizability of these finding might therefore be limited to the oldest age segment. Furthermore, comparisons between the two groups in demographic and health-related variables indicated that the original H70 sample, selected at age 70, had retained its representativeness of the population up to 85 years of age with minor differences in the expected direction of retention and selection characteristics. This does, at least to some extent, justify inferences of retest effects in the demonstrated group comparisons but also emphasizes that a careful interpretation of these effects is required for inferences at the within-person level. The strength of this comparative analysis of retest effects lies in the fact that, in the beginning of the study, in 1971, the entire population of individuals who were 70 years of age living in the city of Go¨teborg was identified. Then, one third of the population was randomly selected for participation with a new sample drawn from the same population cohort at age 85, permitting a comparison of survivors from the original population that differ primarily on repeated test exposures. This is an important design component, because it minimizes confounding effects that are due to time of sampling, cohort, age, and selective attrition differences. The identification of retest effects as the number of occasions each individual has been exposed to a test at a particular time point is also intrinsically confounded with selective population mortality. Previous studies have customarily tackled this problem by using a statistical framework to separately identify the effects of retest from other processes of change. However, in the present study, we controlled these separate effects by using a sampling-based design approach with mortality selection expected to be similar across the two selected samples from the same population. This is an important design component, because it allows us to evaluate retest effects when controlling for the effects of mortality-related selective attrition. We should note that selective attrition might be confounded in estimates from age-homogeneous samples if attrition results from negative experiences associated with previous testing occasions. Such potentially biasing factors are equally confounded in age-heterogeneous longitudinal studies estimating retest effects by number of individual test exposures, but such confounds usually cannot be assessed. Retention and selection effects caused by age of sampling are other factors that have to be considered with an agehomogeneous design with population resampling. For example, it is possible that frail individuals that are asked to participate in a study at age 85 are more likely to participate if they have the experience of participating at earlier occasions compared with those without previous experiences from participation. There was, however, little evidence for selective attrition of this type in the H70 study, as the individuals retained over the longitudinal period compared similarly with the new sample of 85-year-olds from the same population. Therefore, in the H70 study this is an unlikely explanation because comparisons between previous participants and new participants on various health-related variables did not indicate significant differences between the groups. The values observed on variables such as self-rated health and diabetes indicating poorer health of the previous participant could be expected to work in the opposite direction to retest gains, because retest gains would be expected
Figure 6. Plotted fixed effect based on the final model for the Digit Span Backward test (working memory).
to be smaller for less healthy individuals (retained in the study). Another issue seldom addressed in the discussion of retest effects in longitudinal studies is the potential for medical interventions related to study assessments and necessitated for ethical reasons, especially in studies that emphasize clinical examinations, such as the H70 study. The effects of intervention improve the overall health status of the participants, and this will lead to improvements in health-related cognitive abilities and possibly mask the identification of retest effects. Retest effects have been the focus of recent innovations in the statistical analysis of age-heterogeneous longitudinal studies. However, these methods necessitate the use of between-person differences in test performance for the estimation of agingrelated changes that are sufficiently independent from effects of test exposure (number of occasions tested). Results from these models often provide estimates of ‘‘age changes’’ that are similar to what would have been obtained from an analysis of between-person age differences available at the first occasion of testing. There are major confounding factors associated with the uncritical use of between-person age differences, and we are in agreement with Baltes (1968) that ‘‘it must be recognized that the problem of testing effects is not dealt with in a repeated measurement design over the factor age’’(p. 162). Although retest effects are not formally confounded with between-person age differences at the first test occasion, there are many reasons that between-person age differences are a poor proxy for withinperson age changes and that they should not be utilized in this manner. There are many critical assumptions (e.g., no cohort effects, no selective attrition, no mortality selection for agebased between-person sampling) that underlie statistical models for estimating retest effects that rely on the noncomparability of between-person age differences and within-person age changes (Hofer & Sliwinski, 2006). From a conceptual perspective, gains resulting from repeated test exposure cannot be separated from individual-level change that is due to other processes unless the time scales for such changes are separable. Within a given study, this encourages the analysis of individual differences in gains that feature relative performance changes to the common testing cohort.
EFFECTS OF REPEATED TESTING
The major limitation of most models for estimating retest effects, including the approach presented here, is that no information is provided at the individual level but only at the group level. However, retest effects are by definition a withinperson phenomenon and between-person differences should be expected, because individuals learn at different rates as a function of age, ability, and other characteristics. Researchers cannot overcome this obstacle by statistically estimating retest effects as the number of test exposures and separately modeling cognitive changes as a function of both age and practice effects. One possibility is simply to reconsider retest effects as an intrinsic process associated with change in processes of interest, which as such cannot be disentangled from ‘‘true change’’ associated with aging. Such an approach permits ‘‘relative’’ comparison with individuals who have undergone similar experimental conditions (e.g., number of assessments, age of assessment). In studies of early life development, investigators are unlikely to consider separating learning (i.e., retest) from growth and intellectual development. A similar approach should also pertain to gerontological studies. More promising are designs that sample people across different time scales, both short term and long term, and that permit conceptual and empirical distinction of retest or practice gains from long-term aging-related changes. In sum, the comparative design-based analyses in the H70 study indicated a trend toward retest effects on two out of five cognitive tests. Direct estimates of retest gains were permitted by the unique sampling design of the H70 study, in which the population of interest was identified at the beginning of the study. Although intraindividual-level effects of retest, aging, and their interaction cannot be statistically estimated, populationlevel average effects can be usefully evaluated. Individual effects of aging can always be evaluated within a study and relative to individuals with common study experiences such as test exposure or to other samples and designs where retest effects are considered to be minimal. Age-homogeneous designs permit assessment of aging-related change without the confound resulting from population mortality selection that is present in age-heterogeneous sampling designs and featured in recent statistical approaches suggested for estimating retest effects. Given that all longitudinal cognitive measurements can involve a retest or learning component, interpretation of aging-related changes arguably must be made in terms of dynamic constructs conditional on test exposure. Such conditional estimates require intensive measurement designs to estimate long-term changes that permit consideration of morbidity, mortality, and agingrelated changes in within-person learning effects. ACKNOWLEDGMENTS The H70 study is supported by grants from the Swedish Medical Research Council, the Delegation for Social Research within the Ministry of Health and Social Affairs, the Bank of Sweden Tercentenary Foundation, The Go¨teborg Medical Services and Social Services Administrations, Stiftelsen So¨derstro¨m-Ko¨nigska sjukhemmet, Konung Gustaf V:s och Drottning Victorias Stiftelse, Stiftelsen fo¨r Gamla Tja¨narinnor, and Handlanden Hjalmar Svenssons Forskningsfond. Work on this particular article was supported by grants from the Swedish Research Council, Swedish Brain Power, National Institute on Aging, and Wilhelm and Martina Lundgrens Vetenskapsfond. Address correspondence to Boo Johansson, Department of Psychology, Go¨teborg University, Box 500, SE- 405 30 Go¨teborg, Sweden. E-mail:
[email protected]
P353
REFERENCES American Psychiatric Association. (1987). Diagnostic and statistical manual of mental disorders (3rd ed., rev.). Washington, DC: Author. Baltes, P. B. (1968). Longitudinal and cross-sectional sequences in the study of age and generation effects. Human Development, 11, 145–171. Berg, S. (1980). Psychological functioning in 70- and 75-years-old people. A study in an industrialized city. Acta Psychiatrica Scandinavica 62 (Suppl. 288), 1–47. Dureman, I., & Sa¨lde, H. (1959). Psykometriska och experimentalpsykologiska metoder fo¨r klinisk tilla¨mpning [Psychometric and experimental psychological methods for clinical application]. Uppsala, Sweden: Almquist & Wiksell. Ferrer, E., Salthouse, T. A., McArdle, J. J., Stewart, W. F., & Schwartz, B. S. (2005). Multivariate modeling of age and retest in longitudinal studies of cognitive abilities. Psychology and Aging, 20, 412–422. Ferrer, E., Salthouse, T. A., Stewart, W. F., & Schwartz, B. S. (2004). Modeling age and retest processes in longitudinal studies of cognitive abilities. Psychology and Aging, 19, 243–259. Hofer, S. M., Hoffman, L., Sliwinski, M. J., & Piccinin, A. M. (2005, October). Short and long-term cognitive change in aging individuals. Paper presented at the annual meeting of the Society for Multivariate Experimental Psychology, Lake Tahoe, CA. Hofer, S. M., & Sliwinski, M. J. (2006). Design and analysis of longitudinal studies of aging. In J. E. Birren & K. W. Schaie (Eds.), Handbook of the psychology of aging (6th ed., pp. 15–37). San Diego, CA: Academic Press. MacDonald, S. W. S., Hultsch, D. F., Strauss, E., & Dixon, R. A. (2003). Age-related slowing of digit symbol substitution revisited: What do longitudinal age changes reflect? Journal of Gerontology: Psychological Sciences, 58B, P187–P194. McArdle, J. J., Ferrer-Caja E., Hamagami, F., & Woodcock, R. W. (2002). Comparative longitudinal structural analyses of growth and decline of multiple intellectual abilities over life span. Developmental Psychology, 38, 115–142. McArdle, J. J., & Woodcock, R. W. (1997). Expanding test–retest designs to include developmental time-lag components. Psychological Methods, 2, 403–435. Naumova, E. N., Must, A., & Laird, N. M. (2001). Tutorial in biostatistics: Evaluating the impact of ‘critical periods’ in longitudinal studies of growth using piecewise mixed effects models. International Journal of Epidemiology, 30, 1332–1341. Rabbitt, P., Diggle, P., Holland, F., & McInnes, L. (2004). Practice and drop-out effects during a 17-year longitudinal study of cognitive aging. Journal of Gerontology: Psychological Sciences, 59B, P84–P97. Rabbitt, P., Diggle, P., Smith, D., Holland, F., & McInnes, L. (2001). Identifying and separating the effects of practice and of cognitive ageing during a large longitudinal study of elderly community residents. Neuropsychologia, 39, 532–543. Rinder, L., Roupe, S., Steen, B., & Svanborg, A. (1975). Seventy-year-old people in Gothenburg. A population study in an industrialized Swedish city. I. General presentation of the study. Acta Medica Scandinavica, 198, 397–407. Schaie, K. W. (1965). A general model for the study of developmental problems. Psychological Bulletin, 64, 92–107. Schaie, K. W. (1988). Internal validity threats in studies of adult cognitive development. In M. L. Howe & C. J. Brainard (Eds.), Cognitive development in adulthood: Progress in cognitive development research (pp. 241–272). New York: Springer-Verlag. Sliwinski, M., & Buschke, H. (1999). Cross-sectional and longitudinal relationships among age, memory and processing speed. Psychology and Aging, 14, 18–33. Svanborg, A. (1977). Seventy-year-old people in Gothenburg. A population study in an industrialized city. II. General presentation of social and medical conditions. Acta Medica Scandinavica, 611 (Suppl.), 3–37. Received December 14, 2005 Accepted April 17, 2006 Decision Editor: Thomas M. Hess, PhD
P354
THORVALDSSON ET AL.
APPENDIX Estimates of Random Effects on the Final Models Cognitive Test
Model Term
Estimate
SE
Synonym
Intercept Pre-age 85 Post-age 85
47.32** — 0.46*
4.489 — 0.098
Block Design
Intercept Pre-age 85 Post-age 85
38.75*** 0.103 0.42*
3.841 0.038 0.101
Figure Identification
Intercept Pre-age 85 Post-age 85
30.35** 0.26 0.19
3.412 0.070 0.070
Digit Span Forward
Intercept Pre-age 85 Post-age 85
0.76** 0.01 —
0.104 0.003 —
Digit Span Backward
Intercept Pre-age 85 Post-age 85
0.67** 0.0003 0.01
0.100 0.011 0.003
Notes: SE ¼ standard error. The intercept is centered at age 85; Pre-age 85 ¼ linear age-based change slope before age 85; Post-age 85 ¼ linear age-based change slope after age 85. The em dash refers to a nonsignificant model parameter that has been constrained to zero, determined by a full maximum likelihood function estimate. *p , .05; **p , .01; ***p , .001.
