Cohort Differences and Chronic Disease Profiles of Differential Disability Trajectories

Abstract

Objectives.

Research shows declining disability rates, but little is known about whether cohort differences are due to delayed onset, increased recovery, or reduced severity of impairment. Furthermore, disease is considered the proximate cause of disability yet chronic conditions rates are increasing, making it unclear whether the conditions predicting specific disability trajectories are changing.

Methods.

We use a latent class analysis of disability trajectories and corresponding mortality with three birth cohorts of the National Long-Term Care Survey to determine how long-term experiences of disablement differ by cohort and chronic conditions.

Results.

More recent cohorts were more likely to experience a decade free of disablement compared with all other disability trajectories. Sensory problems and hypertension correspond to trajectories of non-disablement, whereas hip fracture, stroke, arthritis, and diabetes predict more disabled experiences.

Discussion.

Later life disability is measured nonparametrically to distinguish patterns among long-term trajectories. Findings suggest that more recent cohorts are more likely to forego or delay disability over a decade rather than experience prolonged periods of mild to severe disablement. Serious health events such as stroke, along with diabetes, characterize trajectories of high impairment, warranting future research.

RESEARCH on disability among elders has shown that prevalence rates have fallen for at least 30 years (e.g., see Freedman, Schoeni, Martin, & Cornman, 2007). This finding is consistent with the expectations of the compression of morbidity, suggesting that poor health will be increasingly delayed toward the end of life (Fries, 1980). However, as Verbrugge (1989) observed, population rates can be obtained from different individual-level patterns, and these patterns may be complex. Indeed, falling prevalence rates may be obtained by (a) delaying the initial onset of disability until older ages, (b) increasing the probability of recovering from disability, and/or (c) reducing the severity of disability across time at the individual level. Determining which of these is occurring requires examining individual patterns of disability across later life, and the compression of morbidity hypothesis implies that these patterns of disability shift across birth cohorts. In this article, we address three questions. First, how many trajectories of disability exist in the older population? Second, how do these distinct experiences differ by cohort? Third, to what extent are these trajectories predicted by chronic conditions?

BACKGROUND

Studies beginning with Manton, Corder, and Stallard (1993) have shown steady and continuing declines in old age disability prevalence (e.g., Crimmins & Saito, 2000). However, there is evidence that gains may vary by the type of disability measure (Crimmins, Saito, & Reynolds, 1997) as varying measures of limitations tap into differential severity of impairment. Most demographic studies use cross-sectional and crude measures of disability in measuring prevalence. In particular, the proportion of the population experiencing at least one physical limitation as measured by activities of daily limitation (ADLs) and/or instrumental activities of daily living (IADLs) items. Such measurement is important from a macro-level policy perspective because it enables policymakers to determine the proportion of the population needing assistance and the associated costs. However, from a micro-level perspective, population-level measures of disability prevalence mask important heterogeneity in the disablement process experienced by individuals.

There are at least two reasons that understanding heterogeneity in the disablement process at the individual level is important. First, from a theoretical perspective, the disablement process is an individual-level phenomenon and should be studied as such. Population estimates can be the product of a number of different processes occurring at the individual level, and these individual processes have differing implications for theory and intervention (Verbrugge, 1989). Declining prevalence rates could imply that everyone in the population is experiencing a delay in the onset of (perhaps terminal) physical limitation. In contrast, declining prevalence rates could be a function of a slowdown in the increase of disability severity across age or an increase in the propensity for recovery from disability. Second, from a clinical view, differences in the experience of disability at the individual level imply different interventions. Understanding how to postpone disability or facilitate maintenance/recovery must be done at the individual level, and by focusing on disability at this level, we can gain a better understanding of the causal processes involved.

These arguments appear obvious, yet disability research often assumes, for example, that individuals follow a unidirectional path from healthy to ill, from ill to disabled, and from disabled to dead. Indeed, Fries’ (1980) compression of morbidity hypothesis implicitly makes this assumption. Future medical advances, therefore, can only postpone the onset of poor health until later in the life span, thus “compressing” poor health further and further against life span limits until everyone is essentially healthy up until the moment of death. Findings in support of compression do show declining disability prevalence rates, but, paradoxically, increases in the prevalence of chronic disease are also occurring (Freedman & Martin, 2000).

Some recent longitudinal research has also shown increases in active life expectancy (Cai & Lubitz, 2007; Manton, Gu, & Lowrimore, 2008). Other studies have used parametric growth models to model disability (Maddox & Clark, 1992), allowing age-specific disability rates to vary by cohort (see Yang & Lee, 2009 for review). Although telling, these models generally fail to account for timing, severity, and recovery simultaneously. Instead, they tend to assume that individual delays in onset or declines in disability (i.e., recovery) reflect measurement error around a stable process of decline in functioning across age. Recent research shows that this approach may be biased in capturing important differences across demographic groups (Taylor, 2008).

Several social, economic, and health-related factors have been implicated in declining disability. Freedman and colleagues (2007) found that heart and circulatory conditions and vision limitations were likely the chronic conditions most responsible for disability declines. Although increasingly prevalent, these conditions have become less disabling, suggesting that the relationship between illness and disability states is changing. Verbrugge and Jette (1994) noted that disability occurs in a social context and is generally defined as an individual’s inability to act in the socially constructed roles they inhabit. Thus, pathology (chronic conditions) impairs specific body systems leading to general functional limitations that translate into the inability to function in ones daily roles. This pathway of disablement is likely impacted by a number of factors both internal and external to the individual including characteristics such as gender and education, individual health behaviors, and available medical interventions and assistive devices in their environments. Researchers studying disability declines note that the connection between pathology and socially defined disablement may be changing at multiple points along the disablement process (Schoeni, Freedman, & Martin, 2008). For example, higher levels of education among successive cohorts may increase income, access to care, and health knowledge, resulting in earlier diagnosis of disease at a time when it is not disabling. The increase of pharmaceuticals and surgical interventions may slow a disease’s impact on the body, whereas assistive devices may keep impairment or functional limitation from translating into disability by allowing individuals to perform daily tasks. In the present analysis, we do not attempt to parse out where or how chronic conditions affect disablement as defined by activities of daily living. Rather, we examine whether certain experiences of disability over time are associated with specific conditions, as suggested by prior research investigating individual heterogeneity in the experience of disability (Murray, Kendall, Boyd, & Sheikh, 2005).

A small body of research has focused on multiple experiences of disability at the end of life and their variation by disease. Lunney, Lynn, Foley, Lipson, and Guralnik (2003) grouped disability in the year before death into four trajectories based on primary diagnosis. Differential trajectories reflected those suffering from cancer, organ failure, frailty/dementia, and death with no disease. A fifth residual category emerged comprising roughly one quarter of the sample. Clipp, Pavalko, and Elder (1992) took on an alternate approach, placing individuals into self-rated health patterns determined by researchers. They found that five experiences generally captured health as individuals aged: constant good health, constant poor health, slow degeneration, rapid degeneration at the end of life, and alternating degeneration and recovery. More recent cohorts maintained good health longer than their earlier counterparts. These studies of multiple trajectories health rely on ad hoc methods of placing individuals into disability or health groups based on their primary condition or observed health, making the statistical testing of experiences and covariates difficult. In this research, we extend this body of literature by incorporating cohort in differentiating nonparametric model estimated trajectories of disability among older adults.

The literature guiding the present study leads us to formulate numerous hypotheses. From the literature on multiple trajectories of functioning and health, we expect that (a) the disability experiences of older adults over a decade will be represented by five general patterns of experience including non-disablement, high stable disablement, slow increase, delayed disability with a large increase thereafter, and a trajectory of alternating increase and recovery. Furthermore, we expect (b) trajectories marked by high levels of disability to be more associated with death over the observation period. (c) More recent cohorts will be more likely to remain non-disabled throughout the decade and will be less likely to die. And (d) individuals with long-term chronic conditions (arthritic and diabetes) will be more likely to be in stable high trajectories of disability, where acute health events will be associated with delayed disability with a sharp increase. Heart disease and sensory problems (vision and hearing) will be more associated with trajectories of non-disablement or slow degeneration.

METHOD

Data

We use the 1984–2004 publicly available National Long-Term Care Survey (NLTCS) and linked Medicare and Vital Statistics records obtained from the Center for Medicare and Medicaid Services. The NLTCS is a panel study with replenishment and was developed to examine health, disability, and service utilization among older individuals (aged 65 years and older). The study is nationally representative of both community-dwelling and institutional-dwelling older adults at each wave conducted in 1982, 1984, 1989, 1994, and 1999, and 2004. Replenishment at each wave has ensured representativeness across time, but this design also allows us to test cohorts at the same ages over time. For our analysis, we stacked three birth cohorts to create an age-based analysis to examine individuals over a decade (three waves) each. The cohorts represented are those aged 65–69 years in (a) 1984, (b) 1989, and (c) 1994 (i.e., the 1915–1919, 1920–1924, and the 1925–1929 birth cohorts).

Of the 18,491 individuals originally sampled who were 65–69 years in each of these three waves, 16,264 cases comprise the analytic sample; 12% are missing. Most of the missingness (∼10%) is due to individuals having no report of disability over the observation window. A small number of cases (N = 241) were missing on the chronic conditions measures because the linkage with Medicare files was incomplete or the Medicare data themselves was not coded properly. Because the percentage of these cases was so small (∼1.5%), these individuals were assumed not to have conditions and were imputed as such. Sensitivity analyses revealed no substantial differences in substantive findings compared with listwise deletion of these cases. Other missingness was handled in the analyses using a full information maximum likelihood estimator, which computes the likelihood function of each individual given all available information. Therefore, individuals were allowed attrition due to mortality, other sources of attrition, or nonresponse while still contributing to the analyses. Because these analyses were performed on stacked (pooled) data of three cohorts matriculating into the NLTCS in different survey years, all analyses are unweighted. The majority of variables used for sample selection criteria (age, cohort, sex, and race) were included as control variables in the models, suggesting that bias in the coefficients should be reduced (Winship & Radbill, 1994). However, it was not possible to include the multiple sampling strata in analyses, which may compress standard errors. Thus, we recognize this as a limitation of the current study.

The NLTCS is uniquely suited for this analysis because of its representativeness. Its main weakness, however, is that the long form of the questionnaire (including most social and health variables) was only administered to individuals if they reported disability on a screener questionnaire. Thus, although measures of disability, age, sex, and race are available for all individuals, the information on socioeconomic status, chronic conditions, etc. is missing for those non-disabled at each wave. For chronic condition measures, we used the Medicare files linked to the NLTCS (1982–2004). Unlike many surveys, the NLTCS boasts linked files for all individuals. Although self-reports of health and disease are the “gold standard” in life course research (Ferraro & Farmer, 1999), obtaining conditions from Medicare files has advantages over self-reports. For example, not all older individuals—especially those with comorbidity or cognitive impairment—may remember all their diagnoses. In addition, because the Medicare records were available annually, we were able to retain information on individuals who died or dropped out of the survey at any wave.

Measures

Disability.—

The most widely accepted and utilized measures of disability in the elderly are Katz’s Activities of Daily Living and IADLs (Katz, Ford, Moskowitz, Jackson, & Jaffe, 1963; Lawton & Brody, 1969). The NLTCS includes a six-item ADL index asking the respondent whether they had any problem eating, getting in and out of bed, walking around inside, dressing, bathing, and getting to the bathroom or using the toilet without help. The 10-item NLTCS IADL index asks whether the respondent can do heavy housework, light housework, laundry, prepare meals, shop for groceries, get around outside, travel, manage money, take medicine, and make telephone calls without help.

We combine ADL and IADL measures using Manton’s index based on how individuals report items by age in the NLTCS (see Manton & Gu, 2001). The index ranges from 1 to 6, where 1 = no disability, 2 = any IADL limitation, 3 = 1–2 ADL limitations, 4 = 3–4 ADL limitations, 5 = 5+ limitations, and 6 = institutionalized. Here, institutionalized individuals are assumed to be the most disabled. We choose this strategy because, in the NLTCS, institutionalized individuals reported ADL limitations only under the assumption that they did not provide the basic care for themselves for successful completion of tasks measured via IADL items. We note here that differences in wording across surveys can alter prevalences of disability, and estimates from the NLTCS show higher proportions disabled compared with some other national data sources (see Freedman, Martin, & Schoeni, 2002). We argue that these differences are generally small, and we have chosen a broad index that should decrease differences between data sources.

Mortality.—

Mortality was taken from the Vital Statistics files linked to the NLTCS and was coded as a binary variable indicating whether or not the individual died over the observation window (starting at ages 65–69 for the following decade). Although this binary indicator is a crude measure of death and does not capture individuals after the observation window, we were not able to include the timing of death because our models were already so computationally complex.

Chronic conditions.—

From the linked Medicare files, we created variables for each of nine conditions: arthritis/rheumatism, diabetes, hearing conditions, heart attack, hip fracture, hypertension, respiratory conditions, stroke, and vision problems which were constructed from ICD-9 diagnosis code (World Health Organization, 1977) brackets found in Appendix 1. Originally, cancer and Alzheimer’s disease were also included in models but were not significant in predicting disability. Although we initially expected Alzheimer’s disease to be a substantial predictor of disability, the proportion of those diagnosed with the disease during this age and time frame in the Medicare records was very low (<1%). Lack of statistical power is likely at work for the absence of significant findings for this disabling disease in our analyses.

Medicare records were examined from inpatient and outpatient procedures, hospice, home health agencies, skilled nursing facilities, and physicians (if individuals were enrolled in Part B) for each year the individual contributed to these analyses (e.g., records from years 1984–1994 for those aged 64–69 years in 1984). Because our interest is in condition profiles corresponding to disability trajectories over the decade rather than predicting disability in any given year and because of the level of computational complexity needed for our modeling strategy, we chose to include nine time-invariant binary variables for whether the individual had any report of each condition during the observation window. Individuals could have more than one condition as they were coded, thus controlling for comorbidity.

Covariates.—

Age (continuous 65–69), sex (female = 1), and race (black = 1) were included in analyses along with dichotomous indicators of successive cohorts (1920–1924, 1925–1929).

Analytic Strategy

Previous research has often used strategies for binary disability outcomes (disabled and non-disabled) either cross sectionally in trend studies or longitudinally using event history models. The latter are telling in examining the timing of shifts into and out of disability, and some studies have used this approach to study transitions within disabled states. Although these strategies are useful for research questions surrounding transitions and have the benefit of addressing the timing of outcomes and predictors well, they generally do not address individual trajectories over time as outcomes (with possible movement into, out of, and within disabled states within one individual’s experience) but rather each transition as an outcome. Previous research on trajectories of disability has predominantly used growth curve methods, which estimate individual trajectories of health and their variation from an “average” trajectory (Lynch, 2003). Although this approach may be appropriate in many situations, it assumes a unilateral increase in disability when applied to older ages and cannot investigate whether there are significant numbers of the population that follow distinct trajectories (delayed onset, impairment, and recovery). Previous studies attempting to categorize distinct trajectories of impairment have generally done so with ad hoc strategies examining specific chronic conditions (Lunney et al., 2003). Although this approach is useful for examining different experiences by primary causes of death, it does not allow statistical tests of multiple conditions, cohort, and other covariates.

In this study, we use latent class analysis (LCA) because it captures trajectories over long periods of time and can model timing, onset, recovery, and growth efficiently in one model. LCA, developed by Lazarsfeld and Henry (1968), offers a nonparametric complement to growth modeling by assuming that deviations from one average trajectory may be substantively important and may not simply reflect measurement error or other random deviations from a linear or curvilinear pattern of increasing disability.

LCA uses a multinomial modeling strategy and, as applied here, does not impose a parametric shape across time. The method assumes that time-specific continuous measures of disability may represent qualitatively different trajectories that exist in a population (Land, 2001; Nagin, 1999). The best model (number and shape of trajectories) is chosen through an examination of overall and component fit statistics (including the Bayesian Information Criterion [BIC] and Akaike Information Criterion [AIC] statistic), amount of classification error (Entropy), and residuals. Generally, models with lower BIC values provide a better fit to the data, and Entropy statistics near 1 (above .8) convey a model with well-separated classes (Celeux & Soromenho, 1996). Predicted classes of disability trajectories are also checked with plots to assure classes represent within-individual trajectories rather than variation at each time point.

Because we are interested in whether individuals are becoming healthier in more recent cohorts, we incorporate mortality into our analyses. We do this by treating a binary mortality variable as a distal outcome both of the disability trajectories and of the covariates. Therefore, we may simultaneously examine which disability experiences and chronic conditions over the decade are most associated with death during that time. Although a true test of morbidity compression would examine both years spent disabled before death and the variability of disability timing and death in the population, the inclusion of mortality as a distal outcome does allow us to examine whether more recent cohorts of older adults are significantly less likely to die over the decade net of their level of disablement. All LCAs were performed using Mplus 4 Software (Muthén & Muthén, 2004).

A parametric model of f(y,λ) may be assumed for disability, where y = (y₁,y₂ … y_T) is the longitudinal sequence of observed levels of these constructs across T periods. If it is assumed that subgroups exist and differ in parameter values, the model may be rewritten as:

(1)

In Equation 1, $p_k$ is the probability of belonging to class k with corresponding parameter(s) ${\lambda }_k$, and ${\lambda }_k$ is dependent on time. First, models were run without covariates to establish the best-fitting model (and thus the number and shape of the trajectories). Second, cohort was examined (controlling for age, gender, and race) to replicate previous cross-sectional studies showing decreases in disability with successive cohorts. Finally, we included chronic condition predictors in the LCA models to establish disease profiles for long-term disability trajectory experiences while taking comorbidity into account.

RESULTS

Descriptive statistics for the three cohorts in the sample are presented in Table 1. Demographic statistics are similar for the three cohorts. Consistent with Freedman and colleagues (2007), the chronic conditions generally increased with more recent cohorts, although this could be due in part to changing diagnosis rates and coverage under Medicare. Increases were greatest among conditions experienced at earlier ages in later life (e.g., hypertension), which correspond to the observation window. On average, there is some decrease in disability and mortality risk with more recent cohorts, but it is impossible to tell from descriptive statistics alone whether these decreases are statistically significant net of demographic factors.

Table 1.

Means (SEs) and Percentages of Descriptive Statistics: Three Cohorts Aged 65–69 Years, National Long-Term Care Survey

	1915–1919 Cohort	1920–1924 Cohort	1925–1929 Cohort
N	7,711	4,701	3,852
Covariates
Age	66.74 (1.33)	67.12 (1.19)	67.01 (1.40)
Black (%)	6.60	6.74	7.29
Female (%)	55.54	56.22	55.53
Arthritis (%)	49.73	65.71	73.42
Diabetes (%)	24.12	31.36	38.19
Hearing problems (%)	7.66	13.17	16.12
Heart attack (%)	19.32	24.89	24.07
Hip fracture (%)	4.67	5.57	5.48
Hypertension (%)	57.89	70.98	77.91
Respiratory problems (%)	35.82	46.69	49.66
Stroke (%)	19.06	27.97	31.10
Vision problems (%)	49.16	64.18	69.44
Disability Wave 1	1.31 (0.95)	1.23 (1.22)	1.23 (0.82)
Disability Wave 2	1.63 (1.31)	1.51 (1.25)	1.42 (1.11)
Disability Wave 3	1.59 (1.31)	1.56 (1.25)	1.53 (1.22)
Mortality (Waves 1–3; %)	28.69	27.31	26.53

The means and standard errors of the best fitting observed classes of disability trajectories are presented in graphic form for the stacked cohorts in Figure 1, and the simultaneously estimated binary outcome for mortality over the observation period is presented in terms of estimated probabilities of death for each trajectory in Figure 2. Examination of BIC, Entropy, and other statistics suggests that a six class model reveals the best fit (estimated both with and without the mortality), but we chose to present the five class model because the sixth class was extremely small (∼1% of the sample) and was substantively very similar to another class. Therefore, we chose the more parsimonious model. The overall model fit statistics are available in Appendix 2. As a reminder, the unadjusted trajectories and mortality probabilities are presented here, meaning that there are no covariates included. Because the model presented predicts trajectory class membership using the covariates (and not variation within the classes), the means and variances of the trajectories and the sample percentages in each of the trajectories remain the same when covariate effects are included. However, with the inclusion of covariates, the relationships between the trajectories and mortality do change because the covariates also predict mortality independent of disability, thus the probability of death corresponding to each disability trajectory is estimated at zero in the final model.

Figure 1.

Five classes of estimated disability trajectories (means and 95% confidence intervals).

Figure 2.

Ten-year mortality probabilities for corresponding disability trajectories.

The class trajectories generally support those hypothesized. Five classes emerged from the data. The majority of the sample (78%) experienced a non-disabled trajectory throughout the decade, and 11% experienced what could be termed a delayed/increase trajectory because they did not experience IADLs or ADLs before ages 70–74 and then went on to become highly disabled by ages 75–80. Less than 10% of the sample experienced a mild (3%) or moderate (4%) trajectory, where individuals aged into later life with either IADL level disablement (mild) or low levels of ADL disability and remained relatively stable throughout the decade. Only 3% of the sample experienced a high stable disability experience, and although some recovery is observed in this trajectory, it is important to note that it is relatively small (less than one unit on the index) and there is no way to tell in these models whether this change over time is significant. In addition, sensitivity analyses suggest that the observed decrease in means is likely due to mortality selection (because individuals in this trajectory were most likely to die) and transition out of institutions rather than true functional improvement.

It is important to note that there was no one observed disability trajectory marked by alternating periods of decline and recovery as suggested by previous research. For this trajectory to emerge from the data, individuals would have had to experience these changes at roughly the same time and/or in a similar pattern. It is also possible that 5-year periods between waves were too long to observe the instability found in research using a shorter time frame (Lunney et al., 2003). To examine recovery further, we outputted the modal class trajectory for each individual and calculated proportions of individuals recovering (defined as any decrease in disability on our index between waves) by trajectory (models not shown). Overall, we found that roughly 3% of those surviving the decade experienced some recovery and that three trajectories (mild, moderate, and high) had individuals recovering within them. In other words, there was no one “recovery” trajectory, rather most disability experiences included at least some individuals who recovered.

The probability of death over the decade corresponding to each disability trajectory can be found in Figure 2. Death was least likely among the non-disabled and most likely in the high disability trajectory, as hypothesized. It is interesting to note that the delay/increase trajectory had a lower probability of death than either the mild or moderate trajectories, suggesting that any delays in disability are beneficial for mortality risk even though individuals in this trajectory ultimately experienced higher levels of disablement by ages 75–79.

The effects of cohort along with demographic controls are presented in Table 2 for both the simultaneously estimated odds of disability trajectory membership (with the non-disabled trajectory as the referent) and the odds of dying over the decade (with survival as the referent). Those in more recent cohorts had 25%–44% lower odds of experiencing the three stable disability trajectories (high, moderate, or mild) compared with the non-disabled. However, there was no significant difference by cohort in experiencing the delayed/increase disability experience compared with the non-disabled. This suggests that after controlling for gender and race, more recent cohorts are more likely to age into late life free of disability and to delay disability onset for at least five years. We should note here that although cohort differences are significant, the difference in the percentage of individuals in each trajectory for each cohort was small in magnitude. We conducted a post hoc analysis of the percentage of each cohort in the different classes, finding that the percentage in the non-disabled trajectory increases roughly 2.5%–3% across these cohorts. Although these differences are small in magnitude, we argue that they are still significant and substantial in projecting these differences beyond the 10-year cohort difference.

Table 2.

Odds Ratios for Cohort Effects Among Differential Disability Trajectories and Mortality

Cohort effects, N = 16,264
Classes	High	Moderate	Mild	Increasing	Mortality
Sample %	2.742	4.808	3.191	10.613	—
Covariates
Age	1.258***	1.219***	1.140***	1.182***	1.115***
Black	2.726***	2.581***	2.694***	1.812***	1.335***
Female	1.149	1.415***	1.310***	1.290***	0.486***
1920–1924 cohort	0.679***	0.653***	0.603***	1.050	0.941
1925–1929 cohort	0.627***	0.753***	0.632***	0.975	0.908***
Loglikelihood (number of parameters)	−28,525.773 (49)
BIC	57,526.684
Entropy	0.926

Notes: BIC = Bayesian Information Criterion. Referents for independent variables: non-Black, male, 1915–1919 cohort. Referent for disability trajectories: non-disabled trajectory. Referent for mortality: survival.

*p < .05; **p < .01; ***p < .001.

The effects of cohort on mortality over the decade can be interpreted as mortality risk independent of effects on disability trajectories because the models assume that covariates work both independently on the distal outcome of mortality and also through the disablement trajectories. The youngest cohort was significantly less likely to die during the decade (odds ratio [OR] = 0.91) suggesting compression of disability for the 1920–1924 cohort (because they were no less likely to die) and either compression or dynamic equilibrium for the youngest (1925–1929) cohort.

After controlling for cohort effects, each year of age increased the odds of experiencing any trajectory characterized by disability over the decade and increased the odds of death over the decade. Black individuals were also at considerably higher risk of both disability and death. Notably, Black individuals had more than 2.5 times the odds of experiencing disability throughout the decade (high, moderate, and mild trajectories) and had 81% higher odds of the delayed/increase trajectory compared with remaining non-disabled. They also had 34% higher odds of death over the decade. The findings for gender in this table support the gender paradox, with women more likely to experience trajectories characterized by disability (with the exception of high stable) compared with men but less likely to die (OR = 0.49) over the decade.

Chronic condition effects for the disability trajectories and mortality are examined in Table 3, with each trajectory characterized by disability hosting its own disease profile, net of comorbidity. The high disability trajectory was characterized by diabetes as hypothesized but was characterized primarily by hip fracture and stroke. These three conditions also characterized the mild and moderate trajectories compared with remaining non-disabled but to a lesser extent. The moderate disability trajectory was also characterized by heart attack and respiratory disease, somewhat consistent with the hypothesis that heart disease would predict a slow degenerative experience. Heart attack was not significant in predicting the mild disability trajectory but respiratory disease was, and the magnitude was greater than for moderate. The increasing/delayed trajectory was the only one significantly characterized by arthritis and heart attack, again suggesting partial support for the hypothesis that acute health events (heart attack) would be more predictive of this experience. As hypothesized, sensory diseases (vision and hearing) and hypertension were significantly less likely to characterize disability trajectories overall, and thus, they significantly characterize an experience of non-disablement over the decade. Net of their impact on disability, respiratory disease, and heart attack were the conditions most associated with mortality over the decade followed by hip fracture and diabetes.

Table 3.

Odds Ratios for Chronic Condition Effects Among Differential Disability Trajectories and Mortality

Chronic disease effects, N = 16,264
Classes	High	Moderate	Mild	Increasing	Mortality
Sample %	2.742	4.808	3.191	10.676	—
Covariates
Age	1.263***	1.218***	1.142***	1.170***	1.154***
Black	2.368***	2.409***	2.538***	1.623***	1.267***
Female	1.453***	1.572***	1.414***	1.391***	0.628***
1920–1924 cohort	0.822	0.611***	0.57***	0.877	1.246***
1925–1929 cohort	0.867	0.695***	0.589***	0.794***	1.406***
Arthritis/rheumatism	0.633***	1.091	1.035	1.279***	0.418***
Diabetes	2.481***	1.929***	1.614***	1.922***	1.346***
Hearing problems	0.316***	0.743***	0.759	0.728***	0.496***
Heart attack	0.941	1.313***	1.201	1.246***	1.647***
Hip fracture	6.044***	3.736***	2.923***	5.363***	1.443***
Hypertension	0.536***	0.738***	0.92	1.057	0.614***
Respiratory	1.223	1.563***	1.792***	1.494***	1.754***
Stroke	2.506***	2.038***	1.319***	2.803***	1.311***
Eye problems	0.259***	0.523***	0.617***	0.597***	0.295***
Loglikelihood (number of parameters)	−26,736.838 (94)
BIC	54,385.166
Entropy	0.932

Notes: BIC = Bayesian Information Criterion. Referents for independent variables: non-Black, male, 1915–1919 cohort. Referent for disability trajectories: non-disabled trajectory. Referent for mortality: survival.

*p < .05; **p < .01; ***p < .001.

Once chronic conditions were included in the model, the impact of cohort on disability trajectories was diminished but remained significant. Although temporal order makes statements of causation impossible for these analyses (because disease, disability, and death are all examined over the same time frame with no time varying effects), the fact that more recent cohorts were still less likely to experience disability throughout the decade suggests that factors other than the nine diseases we include here are at work in explaining the trend of decreasing disability among successive cohorts. Compared with Table 2, the effects of age remained consistent. The race differentials decreased in magnitude but remained significant. Interestingly, once chronic disease was included, the gender differentials increased in magnitude for disability but decreased for mortality.

DISCUSSION

This study attempted to examine cohort differences in the long-term experience of disability for older Americans. Five average trajectories generally characterize the disability experiences of three cohorts of older adults over the decade. These patterns are in line with findings by Clipp and colleagues (1992), except that no one trajectory emerged characterized by a pattern of alternating increase and recovery. Overall, more recent cohorts were more likely to experience trajectories free from disability and were as likely to delay disability for five years compared with their counterparts in the earliest cohort. In establishing disease profiles for disability trajectories, we find that remaining disability free over the decade was associated with sensory problems (vision and hearing) and hypertension. Experiences of slow degeneration (mild and moderate trajectories) were characterized by heart attack and respiratory problems. Stroke and hip fracture, along with diabetes, were associated with a high stable trajectory.

Although this study extends previous research on trends and trajectories of disability and chronic disease, it has limitations worth noting here. We chose to examine chronic conditions because these have been focused on in previous research. We do, however, note that the pathway between conditions and disability as defined by Verbrugge and colleagues is complex and mediating factors such as assistive technology have likely played a large role in how disabling conditions are and how this has changed over time. Although the NLTCS data are a nationally representative data source, none of these analyses are weighted because the original sample weights and strata differ by survey year and we study a relatively small age range (64–69) from each replenishment sample. Although the cohorts were very similar on demographic measures, it is impossible to derive trends from these analyses.

We chose to use Manton’s scale to make the best use of the available measures; however, it has the limitation of assuming institutionalized individuals have the highest level of disability. Extensive sensitivity analyses with this and another data source show that this is a valid assumption (Taylor, 2005), however, analyses utilizing other measures of disability (ADLs only) would likely generate trajectories varying in number and shape from those presented here. Finally, we chose to use linked Medicare records in order to gain chronic condition information for the individuals in the NLTCS. However, self-report remains the gold standard in life course research (see Ferraro & Farmer, 1999; Freedman et al., 2007). Furthermore, our measures of chronic conditions were not time varying because of the complexity of our models, making statements about causation inadvisable. Like the condition measures, our measure of mortality was crude and did not incorporate timing. This was also because the model complexity would not allow this level of detail. Future research should examine timing of death in more detail.

Overall, we find support for trend studies showing delays in disability among more recent cohorts. We also find, consistent with Fries (1980), limited support for compression because the most recent cohort was less likely to die over the observation period but was marked by delayed disability. Our findings show that no one chronic disease predicts a specific disablement experience and that individuals with similar chronic condition profiles could have different disability experiences. We find no difference in the disability trajectory experienced between substantial health events (stroke) and more chronic conditions (diabetes), but this may be due to the large time periods between our waves.

Understanding these differences among the population is paramount for intervention strategies and policy. Support for compression means that, at the policy level, the necessity for care at the end of life is decreasing in duration, although some scholars have noted that this is at least in part due to the utilization of assistive technologies that allow individuals to function with limitation rather than individuals being free of limitation. Our findings also point to the importance of interventions or medical advances to delay particularly disabling health events such as stroke and heart attack but also more chronic conditions such as diabetes and respiratory disease because these are moderately to severely disabling for substantial amounts of time in later life. These diseases are also important due to changing patterns of smoking and obesity in the elderly. Future research stemming from the findings presented here also includes a closer look at the intersection of cohort and condition in explaining differential disability trajectories. Differences by race and/or gender were not explicitly explored here and should be in the future because it is clear that disability declines have not been universal across the older adult population and trends in mortality contribute to disparities. Specifically, the race differentials were substantial in magnitude, and likely, there are interactions between race/ethnicity and cohort in estimating both disability experiences and mortality in successive cohorts.

FUNDING

M. G. Taylor was supported by the National Institute on Aging (grants no. F32AG026926 and K99AG030471).

Appendix 1. Chronic condition codes

Condition/event	ICD codes
Vision problems	365.xx, 366.xx, 369.9
Hearing loss	389.xx
Diabetes	250.0–250.9
Chronic respiratory diseases	490.xx–496.xx
Arthritis (general, based on NADW)	95.6, 95.7, 98.5, 99.3, 136.1, 274.xx, 277.2, 287.0, 334.6, 353.0, 354.0, 355.5, 357.1, 390, 391, 437.4, 443.0, 446.xx, 447.6, 696.0, 710.xx, 716.xx, 719.0, 719.2–719.9, 720.xx 721.xx, 725.xx–727.xx, 728.0–728.3, 728.6 728.9, 729.0-729.1, and 729.4
Hypertension	401.xx–403.xx
Heart attack	410.xx
Stroke	433.xx, 434.xx, 436.xx
Hip fracture	820.xx

Note. xx refers to any number after the decimal point.

Appendix 2. Overall Fit Statistics for Latent Class Analysis Model Choice

	Loglikelihood	Number of parameters	AIC	BIC	Entropy
2-Cluster	−37,334.112	10	74,668.224	74,765.19	0.998
3-Cluster	−29,985.836	14	59,999.671	60,107.43	1
4-Cluster	−25,145.729	18	50,327.458	50,466	0.953
5-Cluster	−19,502.993	22	39,049.985	39,149.4	0.952
6-Cluster	−12,221.96	26	24,495.92	24,696.04	0.957

Note. AIC = Akaike Information Criterion; BIC = Bayesian Information Criterion.