| HOME | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|---|
| ||||||||||||||||||||||||||||||||
RESEARCH ARTICLE |
1 Polisher Research Institute (formerly Philadelphia Geriatric Center), North Wales, Pennsylvania.
2 Department of Child and Family Development, University of Georgia, Athens.
3 National Institute on Aging, National Institutes of Health, Baltimore, Maryland.
Address correspondence to Adam Davey, Polisher Research Institute, Madlyn and Leonard Abramson Center for Jewish Life, 1425 Horsham Road, North Wales, PA 19454-1320. E-mail: adavey{at}abramsoncenter.org
 |
Abstract |
|---|
 TOP Abstract MethodsResultsDiscussionReferences |
|---|
The depression construct has a long history of either being treated as a state variable, responsive to stressful life events such as loss of a spouse or declining functional capacity (e.g., Gatz, Kasl-Godley, & Karel, 1996
), or as being conceptualized as a stable and enduring individual characteristic closely linked to negative affect and the personality dimension of Neuroticism (cf. Costa & McCrae, 1994). To characterize depressive symptoms as state-like obscures how the various aspects of depression are strongly linked to individual differences in normal personality dimensions. Similarly, to characterize depressive symptoms as trait-like may deflect interest in the changes that do occur in response to treatment. We set out to elaborate this distinction in order to provide a fuller evaluation of the nature of change in depressive symptoms longitudinally. Consistent with previous research in this area, reviewed in the paragraphs that follow, we focus on normative developmental changes in depressive symptoms, as opposed to changes in clinical diagnoses of major depression (e.g., Smyer & Qualls, 1999; Zarit & Zarit, 1998). This topic is important not only for the study of mental health and aging and developmental changes in emotion regulation, but also because the study of interindividual differences in intraindividual change lies at the heart of life-span developmental psychology.
Researchers have used both cross-sectional and longitudinal designs to study the relation between age and depressive symptoms. Whether this distinction matters depends primarily on (a) whether there are selection effects at the population level (i.e., if depressive symptoms pose a risk for mortality), (b) whether there are selection effects at the sample level (i.e., if depressive symptoms pose a risk for dropping out of a longitudinal study), or (c) whether there are cohort effects. Selection effects at the population level would render cross-sectional comparisons invalid, because those individuals who survive into old age exhibit better than average profiles of depressive symptoms and are therefore not directly comparable with groups of younger individuals. Under these circumstances, even if there were no change in depressive symptoms with age, those individuals surviving to old age would have lower scores, on average, than younger groups, making it appear as though depressive symptoms decline with age. Selection effects at the sample level would similarly bias inferences about the nature of depressive symptoms over time, resulting in, at a minimum, lower overall levels of depressive symptoms among individuals who return for multiple waves of a longitudinal study if individuals who are more depressed are more likely to drop out (cf. Barefoot, Mortensen, Helms, Avlund, & Schroll, 2001). The presence of cohort effects would invalidate cross-sectional comparisons and limit the generalizability of longitudinal findings drawn from a single cohort.
An examination of the literature on negative affect and depressive symptoms from these three perspectives suggests that all three confounds may exist. Cross-sectional research tends to report either that negative affect and depressive symptoms decrease with time or show a curvilinear association with age. Gatz and Hurwicz (1990) compared four age groups, ranging from 20–39 years to 70–98 years of age. These authors found that depressive symptoms decreased until late midlife and then increased in old age, with these differences largely driven by changes in the positive affect subscale of the Center for Epidemiological Studies–Depression Scale (CES-D; Radloff, 1977). Lewinsohn, Rohde, Seeley, and Fisher (1991) reported a similar cross-sectional trend in age differences, although age was only significantly associated with depressive symptoms in the largest (N = 2,730) of their representative community-based samples. Kessler, Foster, Webster, and House (1992) used data from two nationally representative samples to examine the association between age and depressive symptoms. These authors likewise found a curvilinear association between a person's age and depressive symptoms, with the lowest levels at midlife and increases into later life, beginning sometime after the person reaches about 60 years of age. They further separated out the depressed affect and somatic subscales of the CES-D, finding that both scales showed comparable age associations. Age differences in depressive symptoms, then, are not simply an artifact of increasing physical symptoms.
Cross-sectional research on age differences in the related construct of negative affect generally reports that negative affect decreases with advanced age, consistent with the idea that older adults become more effective at management of negative emotions over time (e.g., Carstensen & Charles, 1994). In a national sample of midlife adults, aged 25 to 75 years, Mroczek and Kolarz (1998) found that negative affect was lower and positive affect higher among older men than younger men. These findings were qualified by interactions with both trait-like variables (positive affect is more positively correlated with age among individuals who are more extraverted) and situational or sociodemographic variables (negative affect is more strongly negatively associated with age among married than nonmarried individuals). There were no age differences in negative affect for women. In a substantially older sample of adults aged 70 to 100 years, Smith and Baltes (1993) also reported a negative correlation between age and negative affect, with the association being stronger for men than for women. Whereas men generally report lower levels of negative affect than women, these gender differences diminish or disappear by later life.
Findings from longitudinal research are more difficult to interpret as a result of dropout. In their cross-sectional analyses, Stacey & Gatz (1991) found that negative affect was lower in the oldest age cohorts; their longitudinal analyses with those who were available for a follow-up also showed an age-associated decline in negative affect over time. These results should be qualified by findings from a study by Barefoot and colleagues (2001). Beginning with 739 individuals aged 50 years in 1964, they conducted follow-up interviews with 570 of these individuals in 1974, and 181 individuals in 1995, 30 of whom had not been present for the previous data-collection point. Several key points are illustrated by their findings. First, depressive symptoms were associated with both mortality and subsequent nonresponse. On the basis of the individuals who were present within each data-collection point, women had higher levels of depressive symptoms than men. Over time, depressive symptoms increased for men, but not for women, such that gender differences largely disappeared by age 80. Barefoot and colleagues also demonstrated that, although somatic symptoms increase for both men and women with age, somatic changes alone do not explain age- or gender-related changes in depressive symptoms.
Even within longitudinal studies, models implying different developmental processes have been applied to negative affect, psychological symptoms, and depressive symptoms. Consistent with a state-only (no trait) conception of depressive symptoms, Wetherell, Gatz, and Pedersen (2001) applied an autoregressive model to depressive symptoms over time in a sample of Swedish elders. This approach suggests that only the previous state levels predict current levels of depressive symptoms, and that more closely spaced time points correlate more strongly than more distal time points, eventually diminishing to zero. These authors concluded that depressive symptoms were somewhat less trait-like than anxiety, and that anxiety caused depressive symptoms rather than the reverse. Although their analysis was limited to only one longitudinal structure (autoregressive) and did not examine longitudinal mean changes, it does provide evidence of significant situational and stable variance over time.
Aldwin, Spiro, Levenson, and Bossé (1989) estimated two-stage growth curve models for depressive symptoms. This approach suggests that individuals have two trait-like components; the first identifies individual differences in overall level of depressive symptoms at a specific time point, and the second identifies individual differences in rates of change. Further, because individuals are expected to differ from one another in terms of how they change over time, these models predict that the correlations among observations at some time points may be positive, and, at others, negative (cf. Beach, Davey, & Fincham, 1999). Aldwin and associates (1989) found that a curvilinear function best accounted for psychological symptoms in their sample of older men from the Normative Aging Study, with symptoms decreasing until midlife and increasing again beginning when the men were around the age of 60. Likewise, using data from the Longitudinal Study of Generations, Charles, Reynolds, and Gatz (2001) fitted a linear growth curve model to negative affect data across up to 23 years for members of different generations. Although linear models were applied within each generation, subsequent analyses showed generational differences, such that negative affect was seen to decrease steadily until the person reached approximately the age of 60, at which point it essentially leveled off. In sum, a balance of evidence suggests that depressive symptoms may increase in later life, but that an accurate representation of their course must disentangle them from attrition and selective nonresponse.
Our goal in the current study was to characterize the developmental course of depressive symptoms in later life, and we propose a middle ground between the two longitudinal approaches to analysis that have previously been presented in the literature. Specifically, we propose that a latent trait–state model, drawing strongly on the work of Steyer and colleagues (Steyer, Ferring, & Schmitt, 1992; Steyer, Majcen, Schwenkmezger, & Buchner, 1989; Steyer, Schmitt, & Eid, 1999), would best account for depressive symptoms over time. An individual's score on any occasion ("state" score, in Steyer's terminology) consists of a trait component (common across occasions), a state-residual component (unique to each occasion and context), and measurement error. This model implies substantial stability of individual differences over time while allowing for situation-specific state residuals. As elaborated by Kenny and Zautra (2001; see also Davey, 2001), these models also allow for the possibility that more closely spaced observations will correlate more highly than observations made over longer time periods. The key feature of these models, however, is that they imply high correlations across observations spanning even very long time periods. We also test the possibility that methods variance may partially account for the stability we observe over time, building on recent developments in this area (Eid, 2000; Eid, Lischetzke, Trierweiler, & Nussbeck, 2003; Schermelleh-Engel, Keith, Moosbrugger, & Hodapp, 2004). On the basis of the findings from previous research, we were specifically interested in determining whether a state-only (autoregressive), trait–state (LTSM), or two-trait (growth curve; Tisak & Tisak, 2000) model was needed to account for depressive symptoms longitudinally, once attrition and measurement issues had been addressed.
The preceding review leads us to expect that depressive symptoms will increase over the age range we consider here; however, they do not inform us about changes that may occur earlier in the life span. We expect as well that longitudinal designs will show greater age-related increases in depressive symptoms than cross-sectional research, and that studies that attempt to correct for participant dropout and attrition will demonstrate higher levels of depressive symptoms than studies that do not. Because previous research has also suggested gender differences, with women having higher depressive symptoms than men, we include this variable in our model although we do not have a large enough sample size to perform analyses separately for men and women.
|
METHODS |
|---|
|
TOPAbstractMethodsResultsDiscussionReferences |
|---|
Measures
The primary measure we used in this study was the CES-D. The psychometric properties of the CES-D, including how many dimensions of depressive symptoms it measures and which items correspond with which dimensions, have been debated in the literature (e.g., Helmes & Warren, 1998), but we adopted an approach that has previously been applied within the longitudinal context (e.g., Dumenci & Windle, 1996), and with older adults (Hertzog, Van Alstine, Usala, Hultsch, & Dixon, 1990; Krause & Liang, 1992; McCallum, Mackinnon, Simons, & Simons, 1995). Specifically, we used Radloff's (1977) original four-factor structure, including dimensions of depressed affect (DA), positive affect (PA), somatic complaints (SO), and interpersonal symptoms (IN). A lack of variability on interpersonal symptoms (depending on the wave, between 86% and 91% of the sample had scores of 0 on this subscale) made this scale unsuitable for inclusion in subsequent analyses, so we focused only on the first three dimensions. An exploratory factor analysis of the CES-D by Krause, Liang, and Yatomi (1989) also yielded these three factors, and the interpersonal scale had the lowest mean and variance in the longitudinal study of depressive symptoms by Wetherell and colleagues (2001), even when they adjusted for the number of items in each scale. Scales were positively skewed; however, previous research in the longitudinal context using skewed neuroticism variables found that trait–state models were robust to this violation of normality because all variables were skewed in similar directions (Ormel & Rijsdijk, 2000). Internal consistency (alpha) of the depressed affect subscale ranged from 
=.74 to =.78 across the five occasions. For positive affect, internal consistency ranged from =.62 to =.70. For the somatic subscale, internal consistency ranged from =.57 to =.69.
|
RESULTS |
|---|
|
TOPAbstractMethodsResultsDiscussionReferences |
|---|
Participant Nonresponse and Attrition
First, it is inevitable that some participants will be lost from longitudinal research. The problem of loss is particularly acute when one is dealing with older adults. Second, there is also increasing awareness (e.g., Davey, Shanahan, & Shafer, 2001) that, with multiwave data, selective nonresponse may bias inferences about developmental processes, where generalizations are made regarding change in the original sample, and not just survivors.
In order for our analyses to be as statistically powerful as possible, we performed all final analyses by using FIML estimation (cf. Allison, 1987; Arbuckle, 1996; Little, 1995; Marini, Olsen, & Rubin, 1977; McArdle, 1994). This method makes use of all available information from each observation to estimate model parameters and their corresponding standard errors. There were no missing observations within waves, and, when a participant dropped out of the study, he or she was not recontacted to participate in subsequent waves; thus all individuals missing at a particular wave were also missing on subsequent occasions, creating a monotone nonresponse pattern. Logistic regression analyses (not presented) predicting patterns of missing data indicated that older age and greater depressive symptoms were associated with a greater likelihood of dropout at all subsequent waves. Every additional year of age was associated with a 7% greater likelihood of dropout at subsequent waves, and each additional point higher on the CES-D was associated with between a 4% and 10% greater probability of dropout, depending on the wave. Otherwise, there did not appear to be any systematic processes determining participant dropout. In particular, there were no gender differences in dropout rates, once we controlled for other variables. Unfortunately, we were not able to explore differences as a function of reasons for dropout and attrition (e.g., mortality or ill health).
Several features were of interest in this process. First, although baseline scores differed across the groups as a function of whether and when they dropped out of the study, the change within each pattern was parallel and increasing over time. Likewise, the proportion of individuals within each pattern of missing data who met clinical cutoffs on the CES-D (a score of 16 or higher) differed in a striking fashion. Whereas only 2.5% of individuals who provided five waves of data were at this level or above it, 16.7% of those who provided only one wave of data met the cutoff. Again, however, the proportion of individuals meeting clinical cutoffs increased in each pattern of missing data over time, suggesting that our findings are not biased by differences between clinical and nonclinical levels of depressive symptoms but rather appear to operate in a similar fashion across a wide continuum of depressive symptomatology. For individuals providing all five waves of data, the proportion meeting clinical cutoffs 8 years after recruitment into the study more than doubled to 5.8%. For this reason, we further examined change in depressive symptoms by excluding individuals who initially met the clinical cutoff; levels of depressive symptoms, although obviously lower overall in this subsample, showed parallel longitudinal changes to the full sample. Change was also parallel in the depressed affect and somatic subscales, suggesting that the change was not being driven solely by changes in somatic symptoms (Berry, Storandt, & Coyne, 1984). Likewise, when dropout between the first and second waves was predicted, only the depressed affect subscale of the CES-D predicted nonresponse. A plot of mean depressive symptoms by chronological age is presented in Figure 1.
|
) has also suggested the possibility of cohort differences in neuroticism, at least in cohorts born after World War II, and we were able to examine this possibility with the current data. Unlike participants in many other longitudinal studies, not all participants began this study at the same time. In fact, entry into the study spanned a period of nearly 10 years across participants. We observed multiple cohorts at comparable ages, and we use this to advantage in the current study. Setting individuals born prior to 1910 as the reference group, we were able to partition out variance according to birth cohorts (i.e., 1910–1919, 1920–1929, and 1930 or later). Relatively few differences emerged as a function of cohort membership: Time 1 and Time 2 levels of somatic symptoms are higher for cohorts born in 1910 or after; Time 2 levels of positive affect are lower and somatic symptoms higher for the 1910 and 1920 cohorts than in the others; and Time 2 levels of depressed affect are higher in the 1930s cohort than the others. Although there were cohort differences (which might affect inferences from cross-sectional data), the results we report here were unaffected by whether variability caused by cohort was included in the models. Our findings are robust in this regard. Thus, we included age, sex, race, and birth cohort as covariates in all of our analyses (cf. Horn & McArdle, 1992; Schafer & Graham, 2002).
Longitudinal Measurement Properties of the CES-D
Measurement models
We began with a baseline measurement model in which we permitted all three indicators of depressive symptoms (depressed affect, positive affect, and somatic symptoms) to load on the latent variable at each time of measurement; that is, we used the multistate model of Steyer and associates (1992), allowing residuals to autocorrelate. We permitted the latent variables to correlate freely over time; see Table 1, Model 1: 
2(50) = 81.3, p <.05, root mean square error of approximation (RMSEA) =.03, normative fit index (NFI) =.99, and Tucker–Lewis Index (TLI ) =.98. Correlations among the latent variables over time are shown in Table 2. Model 1, which demonstrates that the same scales load on the latent variables at all occasions, is what Horn and McArdle (1992) term configural invariance. We next tested whether factor loadings could be held equivalent across occasions as a step toward demonstrating metric invariance, that is, that the latent variables had comparable scales across occasions. These additional constraints did not significantly reduce model fit; see Model 2: 
2(8) = 9.0, ns. We then tested whether the latent intercepts (not shown) could be held equivalent across occasions, thus imposing a common frame of reference on observed scores for all occasions and permitting us to estimate the latent means of depressive symptoms over time, relative to the first occasion of measurement (scalar invariance, after Vandenberg & Lance, 2000). Because latent variables have indeterminate scaling, latent intercepts must first be demonstrated to be proportional in order to make comparisons of latent means meaningful. This is analogous to ensuring that two temperatures are measured on the same scale, such as Fahrenheit, before attempting to determine whether the temperature is increasing or decreasing, a finding that does not depend on the arbitrary scale selected. Again, these assumptions did not significantly reduce model fit; see Model 3: 2(8) = 8.7, ns. This model clearly showed that depressive symptoms increased across occasions, in a manner comparable with that shown in Figure 1. What is important to note about this model, in particular, is that it clearly demonstrates that change is proportional across all of the CES-D subscales and is not due to changes in only one scale. This further elaborates previous findings that suggested that change in depressive symptoms was driven by the somatic scale alone (e.g., Berry et al., 1984). Although such changes likely do occur, neither they nor differential item functioning (e.g., Grayson, Mackinnon, Jorm, Creasey, & Broe, 2000) are sufficient to account for the present findings, because the most relevant covariates such as gender and race were included in the model. Our final measurement model, then, provided an excellent fit to the data; see Model 3: 2(66) = 93.2, p <.05, RMSEA =.03, NFI =.97, and TLI =.98. Having identified the measurement properties of the CES-D over time, we next moved to modeling the trait-like and occasion-specific variability in depressive symptoms longitudinally.
|
|
. We placed no further equality constraints on either the stability coefficients or the variance of the uniquenesses. Even this relatively unrestricted model did a poor job of accounting for the time-dependent variance in depressive symptoms; see Model 4: 2(6) = 60.4, p <.0001. The long-term correlations were much higher in our data than this model would predict, consistent with our expectations that at least one trait component would be needed to model the data. Next, we fitted a standard latent trait–state model, similar to that applied previously to scores of the CES-D in adolescents (e.g., Davey, 2001; Dumenci & Windle, 1996). This model did an excellent job of accounting for variability in depressive symptoms over time; see Model 5: 2(5) = 6.6, ns. To test whether the trait-like variability was equally associated with true scores at each occasion, that is, whether interindividual differences remained constant over time, we constrained loadings of true scores on the trait factor to be equal at all times. Imposing these constraints did not decrease model fit; see Model 6: 2(4) = 7.4, ns. Thus, in addition to maintaining rank order, individuals were neither fanning out nor converging across occasions on trait-like variability. Individuals who were one unit above the mean on trait depressive symptoms initially were still one unit above the mean at Wave 5, despite the fact that trait scores had increased for all individuals.
We were next interested in determining whether the state-residual variability was constant across occasions. Constraining the state-residual variability to be equal across occasions resulted in a decrease in model fit; see Model 7: 2(4) = 10.5, p =.03. Specifically, the proportion of true score variability that was situation dependent (i.e., the latent state–trait coefficients of Steyer et al., 1992) decreased across occasions from roughly 36% at the initial wave of data collection (when participants were an average of 73 years old) to 22% 8 years later. Our final model (Model 6) provided an excellent fit to the data, 2(75) = 107.0, p <.05, RMSEA =.03, NFI =.96, and TLI =.98, and it is shown in Figure 2.
|
; Sayer & Cumsille, 2001; Tisak & Tisak, 2000). Adding a random slope variable to the model did not significantly improve model fit, 2(2) = 3.7, ns, and so the single trait–state model is preferable by virtue of being more parsimonious. Individuals tend to increase at similar rates to one another over time. Overall, depressive symptoms increase over time in a similar way for all individuals, and state-residual variability diminishes with age, suggesting that individuals become more similar to themselves over time. In other words, the stable and enduring differences between individuals remain constant over time, whereas the situational and transient differences between individuals' state levels and trait levels diminish over time. An individual's depressive symptoms are likely to deviate less from his or her overall trait levels on a particular occasion with advancing age.
At the suggestion of one insightful reviewer with particular expertise in latent state–trait models, we also estimated models that included methods variance associated with specific CES-D subscales in lieu of autocorrelated residuals, following recent developments for estimating multitrait, multimethod models (e.g., Eid, 2000; Eid et al., 2003; Schermelleh-Engel et al., 2004). We estimated two models. In the first, Model 8, 2(98) = 162.5, p <.0001, we replaced autocorrelated residuals in Model 6 with latent variables predicting the positive affect and somatic symptoms (see Schermelleh-Engel et al., 2004). We permitted these methods factors to covary with one another, but not with the other latent variables representing trait and state residuals. A second model, Model 9, 2(84) = 135.4, p <.0001, predicted a general trait variable (i.e., a single factor with loadings to all observed indicators) with separate situation-specific latent variables predicting each of Occasions 2 through 5. We again used correlated latent variables representing positive affect and somatic subscales in place of autocorrelated residuals. Although neither of these models fit as well as the selected model (Model 6), they are important for two reasons. First, they did suggest the possibility of method-specific variance (to the extent that different subscales scored in the same fashion can be construed as reflecting distinct "methods") with low correlations between the two methods (–.19 in each case). Second, none of the substantive conclusions were affected by inclusion of methods factors in our models. However, future researchers should attempt to replicate these findings by using methods beyond self-report.
|
DISCUSSION |
|---|
|
TOPAbstractMethodsResultsDiscussionReferences |
|---|
), others (e.g., Barefoot et al., 2001) report an increase; still others show a curvilinear association decreasing early in adulthood and increasing in later life (e.g., Gatz & Hurwicz, 1990; Kessler et al., 1992), although previous methodological approaches render these conclusions tentative. We were able to use multiwave, multicohort data to investigate these changes in later life. We found that individuals with initially higher levels of depressive symptoms were precisely those individuals who were least likely to return at subsequent occasions of measurement, consistent with prior research (Barefoot et al., 2001). One potential reconciliation of the contradictory findings of previous research is that selection processes (into subsequent waves of longitudinal studies as well as within the population at large) in cross-sectional designs (e.g., Lindenberger et al., 1999) obscure the true course of longitudinal change in depressive symptoms when one wishes to generalize longitudinally.
Our results on the longitudinal measurement properties of the CES-D corroborate and significantly extend previous research using this measure with older adults. Like Krause and colleagues (1989), we found that there was little evidence that the interpersonal symptoms subscale of the CES-D had value in a community-dwelling population of older adults. In our sample, there was simply insufficient variability to permit inclusion of this scale in our analyses. Consistent with what Krause and colleagues (1989) and Wetherell and colleagues (2001) found previously, the structure of depressive symptoms is comparable beyond the 18.7-month and 6-year periods used in these authors' studies, to the 8-year period considered here.
Similarly, our findings indicated that true score variability in depressive symptoms is mostly due to trait-like, stable interindividual differences that endure across the 8-year period of the current study (covering an average period of adulthood from 73 to 81 years of age). Accompanying the finding of stable trait-like variance is the finding that there was also significant state-residual variability at each occasion of measurement in this time period. Further, the proportion of state-residual variability decreased over time, relative to the trait-like component, from approximately one third state-residual variance initially to only one fifth state-residual variability by the last occasion of measurement. Whether this reflects a gradual linear decline or a decrease only for the oldest-old in this study will have to be addressed in future research. Individuals differ from one another in levels of depressive symptoms, and these differences appear to increase in a similar fashion for all individuals. These same individuals, however, appeared to differ less from themselves (i.e., average occasion-specific deviations from their trait levels) with increasing age. This finding is highly consistent with Fleeson's (2001) recent assertion that, when one is considering traits and states in personality, one should also consider moments other than the mean, particularly variance.
To further support our findings, we also replicated our findings by using the published Swedish twin data from Wetherell and colleagues (2001). We found that a latent trait–state model also provided a very good fit to their data; for Sample 1, 2(53) = 94.1, RMSEA =.04, NFI =.97, and TLI =.98; for Sample 2, 2(56) = 85.6, RMSEA =.03, NFI =.97, and TLI =.99. Both of their samples showed significant state-residual and trait-like variance, and an increase in depressive symptoms from the first to last occasion of measurement, which was a period of 6 years. The only difference between our analyses and the replication of our analyses using the data of Wetherell and associates was that, for their Sample 2, it was possible to hold state variance constant across occasions at 37%, and mean changes were found after 3 years, again increasing in a linear fashion. This is strong evidence that our findings are reasonably robust across data sets, age groups, and even cross-nationally, although state-residual variability may not change in the same way across all samples.
These results are consistent with recent work by Carstensen and colleagues (Carstensen, Pasupathi, Mayr, & Nesselroade, 2000) showing that negative affective states have shorter durations for older adults than for younger adults. These findings imply that the more quickly individuals are able to regulate their negative affect, the closer they should be, at any given time point, to their trait-like levels of depressive symptoms. Such an emotion regulation approach can thus explain the decreasing state-residual variability we find in the current study and its replications using the data from Wetherell and colleagues (2001). Clearly, however, any regulatory process is insufficient to prevent normative increases in overall mean levels of depressive symptoms with age.
We tested three alternative developmental models of depressive symptoms over time. A state-only autoregressive model of depressive symptoms fit our data poorly. A latent trait–state model allowing occasion-to-occasion carryover effects fit well and the autoregressive effects were nonsignificant, suggesting that, at least over the 2-year period between observations in this study, there were no such residual effects; state-residual variability was independent across occasions (Davey, 2001; Kenny & Zautra, 2001).
We also extended our latent trait–state model to become a two-trait linear growth curve model with latent variables (Aldwin et al., 1989; Charles et al., 2001; Tisak & Tisak, 2000). The substantive conclusion of this model was that there were no interindividual differences in intraindividual rates of change. This finding is particularly important, because some recent studies (e.g., Helson, Jones, & Kwan, 1999; Klohnen, 1999) have assumed that constructs that are theoretically expected to behave in a trait-like fashion (e.g., depressive symptoms and personality variables) may adequately be fit by using growth curve models, and this is predicated on the assumption of interindividual differences in intraindividual change. We find no evidence that this is the case in the present study. It is important to note that each of these longitudinal models implies a different covariance structure among the constructs over time. For this reason, researchers should carefully evaluate their theoretical expectations regarding the structure of the construct over time when they are selecting an appropriate longitudinal model.
The present study contributes to the literature on aging and depressive symptoms in several important ways. Foremost among these is that it provides strong support for the trait-like conception of depressive symptoms, affirming what personality researchers have long suspected (e.g., Costa & McCrae, 1994, 1997). Several key points about the nature of change in depressive symptoms in later life are also worth stating here.
Our research clearly supports the idea that there are mean-level increases in depressive symptoms with age, and they are due to proportional changes in each subscale of the CES-D, rather than just one aspect of depressive symptoms. These increases are small in magnitude, accounting for approximately a 1-point increase per decade. Nonetheless, they are also highly reliable and replicable across subsets of our and others' data and measures, and they persist even once variability that is due to sex, race, and cohort is considered. Identifying the source of these small but reliable changes is an important topic for future research.
It is also clear that previous theory and research in this area do not appear to have overestimated the trait-like nature of depressive symptoms. Indeed, if anything, these results suggest that the opposite may in fact be true. Trait-like variability accounts for the lion's share of the reliable variance in the CES-D, and this proportion increases with age. Although individuals become more similar to themselves with age, the magnitude of rank-order differences between individuals remains constant over time, with people becoming on average neither more or less divergent from one another over time. Finally, our research also informs the debate on interindividual differences in intraindividual change with regard to depressive symptoms and aging. We did not find any evidence that individuals change in a different fashion from one another over time, although there may still be substantial occasion-specific effects at any given time point.
Rather than deny differences in depressive symptoms as a function of life events and experiences, the current findings clearly point toward the need for additional research to carefully investigate how experiences such as changing functional limitations or the loss of a spouse moderate the state-residual component of depressive symptoms. The absence of carryover effects in the state-residual component of depressive symptoms suggests that these influences subside in less than this time period, unless of course they endure for at least the 8-year period of this study (Lucas, Clark, Georgellis, & Diener, 2003).
|
Acknowledgments |
|---|
|
Footnotes |
|---|
Received for publication November 19, 2003. Accepted for publication June 17, 2004.
|
References |
|---|
|
TOPAbstractMethodsResultsDiscussionReferences |
|---|
This article has been cited by other articles:
 |
|
 |
|
  R. S. Wilson, S. E. Arnold, T. L. Beck, J. L. Bienias, and D. A. Bennett Change in Depressive Symptoms During the Prodromal Phase of Alzheimer Disease Arch Gen Psychiatry, April 1, 2008; 65(4): 439 - 445. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 S.-T. Cheng and A. C. M. Chan Filial piety and psychological well-being in well older chinese. J. Gerontol. B. Psychol. Sci. Soc. Sci., September 1, 2006; 61(5): P262 - P269. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||
| HOME | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|---|
