Physical Disability Trajectories in Older Americans With and Without Diabetes: The Role of Age, Gender, Race or ethnicity, and Education

Abstract

Purpose: This research combined cross-sectional and longitudinal data to characterize age-related trajectories in physical disability for adults with and without diabetes in the United States and to investigate if those patterns differ by age, gender, race or ethnicity, and education. Design and Methods: Data were examined on 20,433 adults aged 51 and older from the 1998 to 2006 Health and Retirement Study. Multilevel models and a cohort-sequential design were applied to quantitatively depict the age norm of physical disability after age 50. Results: Adults with diabetes not only experience greater levels of physical disability but also faster rates of deterioration over time. This pattern is net of attrition, time-invariant sociodemographic factors, and time-varying chronic disease conditions. Differences in physical disability between adults with and without diabetes were more pronounced in women, non-White, and those of lower education. The moderating effects of gender and education remained robust even after controlling for selected covariates in the model. Implications: This study highlighted the consistently greater development of disability over time in adults with diabetes and particularly in those who are women, non-White, or adults of lower education. Future studies are recommended to examine the mechanisms underlying the differential effects of diabetes on physical disability by gender and education.

The prevalence of diabetes is rapidly rising in the United States. According to the most recent available data, 7.8% of the total population (Centers for Disease Control and Prevention, 2007), a four- to eightfold increase over the last half-century, have diabetes or prediabetes. Trend data suggest that the burden will continue to increase (Engelgau et al., 2004; Wild & Forouhi, 2007) to 11.5% in 2011, 13.5% in 2021, and 14.5% in 2031 (Mainous et al., 2007). Diabetes is associated with many troublesome complications, such as eye, foot, heart, nerve, and kidney problems, as well as premature mortality (American Diabetes Association, 2008). Although the prevention of diabetes is certainly an important mandate for public health, understanding the long-term experience of adults living with diabetes is also critical for providing practical patient-centered clinical recommendations and, in turn, reducing the risks of diabetes complications.

Many complications of diabetes are associated with physical disability, one of the most relevant predictors of quality of life. For example, amputation and blindness may hinder from activities of daily living (ADLs; e.g., walking), nerve problems may put limitations on strengths and mobility activities (e.g., reach above the head), and the worsening cognitive function may impose limitations on instrumental activities of daily living (IADLs; e.g., managing money, using maps). Increasing studies have suggested that adults with diabetes experience higher levels of physical disability than do those without diabetes (Blaum, Ofstedal, Langa, & Wray, 2003; Bruce et al., 2003; De Rekeneire et al., 2003; Figaro et al., 2006; Gregg, Engelgau, & Narayan, 2002; Gregg, Mangione, et al., 2002; Gregg et al., 2000; Maty, Everson-Rose, Haan, Raghunathan, & Kaplan, 2005; Ryerson et al., 2003; Sinclair, Conroy, & Bayer, 2008; Wray, Ofstedal, Langa, & Blaum, 2005). However, because these studies were limited by cross-sectional data or conventional methods of analyzing longitudinal data (e.g., two-wave data or multiwave data with follow-up years as the time axis), it is not yet clear the role of diabetes in the context of physical disabilities over the life span development (LSD) processes.

The life course (Ben-Shlomo & Kuh, 2002; Kuh, Ben-Shlomo, Lynch, Hallqvist, & Power, 2003) or LSD perspective (Alwin & Wray, 2005) emphasizes that health status, such as physical function, is a continuum over the life span. Thus, health changes are transitions in later life, and the trajectories of these transitions are important research concerns (Elder, 1985; Riley, 1987). Based on this perspective, examining changes in physical function over the life span for adults with and without diabetes can identify both quantity and timing of change related to diabetes.

The disablement process model (Verbrugge & Jette, 1994) provides another theoretical perspective on the pathway between diabetes and physical disability and sheds light on contextual determinants of the course of disablement. According to this model, the disease and disability link may differ by sociodemographic (e.g., gender, race or ethnicity, education), intraindividual (e.g., changes in lifestyle), and extraindividual (e.g., medical care) factors. For example, people of lower SES may have lower access to medical care or inadequate self-care knowledge, thus leading to higher disability. Diabetes is a disease that requires especially high demanding daily based self-management, which requires competent health literacy, a skill essential in communicating with health providers, and navigating the health care system (Kim, Love, Quistberg, & Shea, 2004; Nath, 2007; Schillinger, 2004). Because there is increasing evidence showing significant sociodemographic differences in health literacy (Howard, Sentell, & Gazmararian, 2006; Schillinger, Barton, Karter, Wang, & Adler, 2006; Shire, 2002), it is of particular interest to know if sociodemographic characteristics, such as gender, race or ethnicity, and education, modify the diabetes and disability relationship.

Based on the two conceptual frameworks, this study aimed to identify (a) age-related trajectories of physical disability from midlife to older adulthood for adults with and without diabetes and (b) if those trajectories differ by gender, race or ethnicity, or education. This study included cases of both Type 1 and Type 2 diabetes but not gestational diabetes. Instead of simply delineating change with observed scores by follow-up years and regressing data across those multiwave measures, as was done in previous studies, we use an age trajectory approach facilitated by multilevel models and an accelerated longitudinal design (aka., cohort-sequential design; Duncan, Duncan, & Hops, 1996; Miyazaki & Raudenbush, 2000; Schaie, 1965; Tonry, Ohlin, & Farrington, 1991) to piece together the longer term change pattern from the shorter term change pattern of people who are different ages at the beginning. Specifically, we can characterize age-related disability changes from age 51 to about 100 with data on physical disability obtained at five points from 1998 to 2006 in a representative sample of U.S. middle-aged and older adults in successive birth-year cohorts that spanned 55 years. In addition, the growth parameters that underlie disability status provide greater statistical power to detect small but consistent differences and isolate factors that produce those differences in the intra- and interindividual variations.

Methods

Data and Sample

Our data are drawn from the U.S. Health and Retirement Study (HRS, 1998–2006). The HRS is an ongoing nationally representative panel survey that collects a wide range of data on the physical and mental health, insurance coverage, financial status, family support systems, labor market status, and retirement planning of middle-aged and older Americans, including oversamples of Hispanic or Latino, Black or African American, and Florida residents (Juster & Suzman, 1995). It was funded by the National Institute on Aging and conducted by the Institute for Social Research at the University of Michigan. The initial wave of data collection, fielded in 1992, interviewed 12,543 individuals born 1931–1941 (thus, 51–61 years of age), along with their spouses or partners of any age. Reinterviews have occurred biennially and are ongoing. The 1998 survey not only reinterviewed those who had survived from the 1992 survey (N = 10,584) but also reinterviewed another nationally representative cohort of persons born in 1922 or earlier who were the survivors of the 1993 Study of Asset and Health Dynamics of the Oldest Old (N = 5,860; Soldo, Hurd, Rodgers, & Wallace, 1997) and added respondents from the 1942–1947 (N = 2,529) and 1923–1930 birth cohorts (N = 2,320) to round out the sample aged 51 and older. As a result, 21,384 adults (age 25–105 years) were interviewed in 1998. Among them, 20,443 were representative of the entire population of U.S. adults born in 1947 or earlier (aged 51 or older in 1998). The reinterview response rates for the biennial follow-up interviews for the 1998 HRS sample were 88.2%, 80.2%, 73.5%, and 67.0% in 2000, 2002, 2004, and 2006, respectively.

Given that the focus in this study was to examine changes in physical disability from midlife to older adulthood for adults with and without diabetes, this study utilized data on participants aged 51 or older in the 1998 HRS who answered questions on their diabetes status, yielding an eligible sample size of 20,433.

In order to draw on the advantages of the cohort-sequential design, which links adjacent and overlapping segments of available longitudinal data from different age cohorts, we categorized participants into eight cohorts based on their baseline age: 51–55, 56–60, 61–65, 66–70, 71–75, 76–80, 81–85, and 86 and older. Thus, with the 8-year follow-up, Cohort 1 contributes to the estimation of longitudinal change from ages 51 to 63; Cohort 2 contributes to the estimation from ages 56 to 68, and similarly, Cohort 3, ages 61–73; Cohort 4, ages 66–78; Cohort 5, ages 71–83; Cohort 6, ages 76–88; Cohort 7, ages 81–93; and Cohort 8, ages 86–109. The sample size for each cohort ranged from n = 1,031 for Cohort 1 to n = 3,991 for Cohort 7.

Variables and Measures

Physical disability was measured with self-reported data including six ADLs (bathing, dressing, eating, walking across a room, getting in or out of bed, and using a toilet independently; Katz, Ford, Moskowitz, Jackson, & Jaffe, 1963), five IADLs (preparing meals, shopping, managing money, using the telephone, and using maps; Lawton & Brody, 1969), and five strength or mobility activities (walking several blocks, climbing several flights of stairs, stooping or kneeling or crouching, reaching above the head, and lifting or carrying weights more than 10 pounds like a heavy bag of groceries; Nagi, 1964). Respondents who reported that they had difficulty or were unable to do the task or that they received help or use equipment when performing the task were coded as having difficulty with the task (1 = yes, 0 = no). These items were summed to obtain a physical disability score, with higher scores indicating more severe limitations (range 0–16). This composite measure allows us to assess a broad range of physical disabilities from early or “preclinical” disabilities to later personal care disabilities (Fried, Ettinger, Lind, Newman, & Gardin, 1994; Langlois et al., 1996; Siu, Reuben, & Hays, 1990; Wang, Sheu, & Protas, 2007; Wray et al., 2005). The average internal consistency of the 16 items across all waves was .90.

The time variable was determined by the respondent’s age at each interview wave and was represented as a continuous variable, centering on the grand mean or median age in the analyses (depending on the model). Eight cohorts based on respondents’ baseline age were categorized: 50–55, 56–60, 61–65, 66–70, 71–75, 76–80, 81–85, and 86 and older. Based on the 8-year longitudinal examination, cohorts 1–8 each contribute to longitudinal observations of ages 50–63, 56–68, 61–73, 66–78, 71–83, 76–88, 81–93, and 86–109, respectively. The sample size for each cohort ranged from n = 1,031 for Cohort 1 to n = 3,991 for Cohort 7.

Three time-invariant variables were included: gender, race or ethnicity, and education. Gender was represented as a dummy variable in the models, with men serving as the reference group. Race or ethnicity was categorized into three groups (non-Hispanic White, non-Hispanic Black, and Hispanic or other) with non-Hispanic White as the reference group. Education was categorized at three levels (low = less than high school, medium = GED or high school graduate, and high = some college and above) with high education as the reference group. We classified educational attainment by low, medium, and high levels because they capture greater amounts of disability differences than does a continuous measure of years of education (Ross & Wu, 1996). In addition, the classification will aid in the development of practical interventions by identifying key risk groups. The cutoff of years of education that resulted in the three categories was based on each category representing approximately one third of the entire sample.

Six diabetes-related chronic conditions, suggested by their significant association with self-reported diabetes in chi-square tests, and follow-up status were included as time-varying covariates. The chronic conditions were determined in each wave by asking respondents: “Has a doctor ever told you that you have…?” (a) high blood pressure, (b) cancer, (c) chronic lung disease, (d) heart attack, (e) stroke, and (f) arthritis.

Statistical Analyses

Multilevel models (MLM) using maximum likelihood estimation were fitted to the large longitudinal data using the SAS Proc Mixed procedure. MLM was employed for our analyses because it can analyze information about levels and rates of change in targeted variables over multiple time points, taking into account inter- and intraindividual variability in change. In addition, MLM has two great advantages over traditional latent growth curve models for studying longitudinal change. The first is its flexibility in handling irregularities in data collection (Osgood & Smith, 1995): MLM allows for the inclusion of individuals with incomplete data due to attrition or individuals who provide data at differing ages or unequally spaced measures. Second, MLM easily handles cross-level interactions in time-varying and time-independent covariates.

In the first stage, we determined the most accurate form of physical disability trajectory from midlife to older age. We began with a “full” model, incorporating age and cohort effects, to estimate the mean trajectories of physical disability for each cohort without assuming that all age cohorts follow a common trajectory. Thus, a separate mean trajectory, which bases the age trend strictly on how subjects changed over the 8-year period of our examination, is estimated for each cohort. A two-level model was used to define the mean trajectory for the sample (fixed effects) and to determine information about individual variation in the trajectory (random effects). At the first level of the model, each person’s observed physical function score is conceived of as a polynomial function of age (centering at the median age of each cohort) plus random error. At the second level of the model, these individual coefficients are assumed to vary as a function of the cohort plus person-specific random effects. We also determined if a “reduced” model, or a parsimonious model using a common trajectory for describing change in all cohorts (i.e., ignoring the cohort effect), was tenable. Following Miyazaki and Raudenbush (2000), a cohort convergence test based on the likelihood-ratio test was performed to compare the reduced model to the full model and to identify the best fitting model of age trajectory of physical disability from midlife to older age.

After the best fitting model was determined, we tested the heterogeneity associated with the mean trajectory. Five models were examined in this stage. Model 1 (the diabetes model) tested if the physical disability trajectories for adults with diabetes differed significantly from those without diabetes. This model was performed by adding baseline diabetes status in Level 2 into either the full or the reduced model, depending on the results in Stage 2. Then in Models 2–4 (the diabetes and covariates models), we tested the moderating effects of gender, race or ethnicity, and education by adding interaction terms at the Level 2 model. Finally in Model 5, we tested if diabetes, net of mortality-related attrition, time-invariant demographics, and time-varying chronic health conditions, had a unique effect on the physical disability trajectory by adding the attrition variable and time-invariant demographic covariates in Level 2, as well as time-varying chronic conditions at Level 1.

The HRS is structured by a complex sample design, which includes oversamples of Blacks, Mexican Americans, and Floridians, as well as poststratification adjustments to Current Population Survey estimates of the proportion of the household population in cells defined by race, gender, marital status, and age ranges (Heeringa & Connor, 1995). Many of the attributes (e.g., age, gender, race or ethnicity) upon which unequal selection probabilities were based were explicitly controlled in the multilevel modeling. When sampling weights are solely a function of independent variables in the model, unweighted estimates are preferred because they are unbiased, consistent, and have smaller standard errors than do weighted estimates (Winship & Radbill, 1994). In addition, we undertook separate analyses with and without weighting and obtained very similar results. Thus, we chose not to weight the data.

Throughout our analyses, the models were estimated with an unstructured covariance matrix for the random effects, and assumed a Gaussian distribution of the outcome. Although the composition of the sixteen items for assessing disability avoids the floor effect, the presence of skewness can pose a threat to the estimation of standard errors. We attended to this problem by estimating alternate models with square root transformations of the scores in physical disability. These additional analyses yielded the same substantive results. Accordingly, the results presented subsequently are from the model using the original physical disability scores.

Results

Sample Characteristics and Association with Physical Disability

Table 1 presents descriptive statistics of the sample at baseline and their bivariate associations with physical disability. Of 20,433 participants at baseline, the mean age of the sample was 66.8 years (SD = 10.4); more than half (56.9%) were women; and 82.5% were non-Hispanic White, 14.1% were non-Hispanic Black, and 3.4% were Hispanic or other. The educational levels of our participants are relatively high: 29% were less than high school, 35.3% were high school graduate or with GED, and 35.7% reported some college and above. Diabetes is present in 15.1% of our participants at baseline (n = 3,077). Older cohorts, women, non-Hispanic Black, those of lower education, and those who reported comorbidities had significantly higher physical disability scores than their counterparts.

Table 1.

Baseline Sample Characteristics (%) and Association With Physical Disability (means, SDs, and bivariate tests; N = 20,433)

		Association with physical disability
	Baseline characteristics %	Mean (SD)	95% confidence interval	p
(Age) cohort				<.001
1	3.3	1.48(2.37)	1.39–1.57
2		1.69(2.50)	1.61–1.77
3	10.4	1.85(2.60)	1.76–1.93
4	12.8	1.96(2.60)	1.86–2.05
5	14.1	2.25(2.89)	2.14–2.36
6	17.8	2.98(3.44)	2.83–3.12
7	19.5	4.05(3.99)	3.84–4.26
8	13.6	5.56(4.42)	5.29–5.83
Gender				<.001
Women	56.9	2.64 (3.25)	2.59–2.71
Men	43.1	1.80 (2.76)	1.75–1.86
Education				<.001
Low	29.0	3.43 (3.69)	3.34–3.52
Medium	35.3	2.10 (2.82)	2.04–2.17
High	35.7	1.54 (2.44)	1.48–1.59
Race or ethnicity				<.001
White	82.5	2.14 (2.94)	2.10–2.19
Black	14.1	2.99 (3.60)	2.86–3.12
Hispanic and others	3.4	2.67 (3.46)	2.42–2.94
Diabetes status				<.001
Yes	15.1	3.45 (3.60)	3.33–3.59
No	84.9	2.07 (2.92)	2.03–2.12
High blood pressure				<.001
Yes	49.1	2.86 (3.39)	2.80–2.93
No	50.9	1.72 (2.62)	1.67–1.77
Cancer				<.001
Yes	11.5	2.81 (3.22)	2.68–2.94
No	88.5	2.21 (3.05)	2.17–2.26
Lung disease				<.001
Yes	10.3	3.98 (3.61)	3.83–4.13
No	89.7	2.08 (2.95)	2.05–2.14
Heart disease				<.001
Yes	21.7	3.65 (3.69)	3.54–3.76
No	78.3	1.92 (2.76)	1.86–1.95
Stroke				<.001
Yes	7.69	5.57 (4.57)	5.34–5.79
No	92.3	2.01 (2.75)	1.97–2.05
Arthritis				<.001
Yes	55.5	3.11 (3.34)	3.05–3.18
No	44.5	1.24 (2.33)	1.20–1.29

The Best Fitting Model of Age Trajectory of Physical Disability from Midlife to Older Age

In the full model incorporating both age and cohort effects, the intercept-only (χ₍₃₎² = 382,585), linear (χ₍₂₂₎² = 366,692), quadratic (χ₍₃₄₎² = 365,632), and cubic (χ₍₄₃₎² = 365,608) growth curve models were examined. Based on the deviance tests, the results showed that a cubic model was preferred to a quadratic model (Δ χ₍₉₎² = 24, p < .001). Thus, a cubic function of age was selected to proceed to the next stage examining if a simpler (reduced) model without cohort effects fit the data as well as the full model. We counted 43 parameters in the full model, with a deviance of 366,304. The number of parameters in the reduced model was 11, and the deviance was 365,608. By subtraction, we obtained df = 32 for the likelihood-ratio test (Δ_χ(9)² = 696, p < .001), suggesting that it was not appropriate to use a single underlying change continuum to describe changes from midlife to older adulthood based on the present data. The cohort-based cubic function of age trajectory was therefore chosen to be the best fitting model to depict change in physical disability after 50. The first column of Table 2 (Model 0) presents the growth parameters (as shown in the fixed effects) and the variances (random effects) of the basic model. Nearly all the fixed effect coefficients in the basic model were significant, suggesting that age and cohort were meaningful predictors for describing change in physical disability from midlife to older adulthood. In addition, the covariation between random intercept and slope was significant and positive, indicating that on average, adults with higher levels of physical disability at baseline had more rapid increases in physical disability later in life. The significant variances of intercept, slope, and curvature, however, suggest that there were significant interindividual differences in the mean age cohort trajectory.

Table 2.

Growth Models of Physical Disability Trajectories From Midlife to Older Adulthood (numbers are fixed effects coefficient estimates, random effects variances, and fit indices)

	Model 0a	Model 1b	Model 2c	Model 3d	Model 4e	Model 5f
Fixed effects
Intercept	1.784***	1.614***	1.208***	1.477***	1.232***	0.0839
Age	0.0786***	0.072***	0.068***	0.068***	0.070***	0.026*
Age2	0.0007	0.0007	0.0007	0.0007	0.0008	0.001
Age3	−0.0002	−0.0003	−0.0003	−0.0003	−0.0003	−0.0002
Cohort 2	0.161*	0.111	0.162*	0.112	−0.005	−0.057
Cohort 3	0.353***	0.235**	0.281**	0.236**	0.080	−0.119
Cohort 4	0.708***	0.595***	0.649***	0.637***	0.432***	0.066
Cohort 5	1.171***	1.052***	1.100***	1.113***	0.827***	0.271***
Cohort 6	2.230***	2.105***	2.105***	2.149***	1.841***	0.886***
Cohort 7	4.097***	4.014***	3.980***	4.047***	3.640***	2.276***
Cohort 8	8.293***	8.268***	8.164***	8.288***	7.770***	5.942***
Age × Cohort 2	−0.0128	−0.015	0.645	−0.015	−0.016	−0.008
Age × Cohort 3	0.0288*	0.025	0.461	0.025	0.023	0.032
Age × Cohort 4	0.0675***	0.064***	0.242***	0.064***	0.061***	0.061*
Age × Cohort 5	0.1567***	0.153***	0.152***	0.154***	0.151***	0.149***
Age × Cohort 6	0.2491***	0.243***	0.064***	0.244***	0.241***	0.226***
Age × Cohort 7	0.4669***	0.462***	0.025***	0.462***	0.457***	0.436***
Age × Cohort 8	0.6538***	0.653***	−0.014***	0.651***	0.635***	0.593***
Age2 × Cohort 2	0.0008	0.0008	0.0007	0.0007	0.0007	−0.0007
Age2 × Cohort 3	0.0059**	0.006**	0.006**	0.006**	0.005**	0.004*
Age2 × Cohort 4	0.0093***	0.009***	0.009***	0.009***	0.008**	0.006**
Age2 × Cohort 5	000156***	0.015***	0.016***	0.015***	0.015***	0.015***
Age2 × Cohort 6	0.0255***	0.025***	0.025***	0.025***	0.025***	0.025***
Age2 × Cohort 7	0.0202***	0.019***	0.019***	0.195***	0.019***	0.020***
Age2 × Cohort 8	0.0072	0.007	0.007	0.007	0.006	0.002
Age3 × Cohort 2	0.0004	0.0004	0.0004	0.0004	0.0004	0.0003
Age3 × Cohort 3	−0.0000	−0.00002	0.000	−0.00002	−0.00001	−0.0001
Age3 × Cohort 4	0.0014*	0.001*	0.001*	0.001*	0.001**	0.001**
Age3 × Cohort 5	0.0008	0.0007	0.0007	0.0007	0.0007	0.0005
Age3 × Cohort 6	0.0020**	0.002**	0.002**	0.002*	0.002**	0.002***
Age3 × Cohort 7	0.0003	0.0004	0.0003	0.0003	0.0003	0.0003
Age3 × Cohort 8	−0.0007	−0.0007	−0.0007	−0.0007	−0.0006	−0.0008
Diabetes		1.699***	1.370***	1.528***	1.396***	0.652***
Diabetes × Age		0.073***	0.076***	0.075***	0.046**	0.047**
Women			0.657***			0.718***
Women × Age			0.005			0.006
Women × Diabetes			0.648***			0.389***
Women × Diabetes × Age			−0.006			−0.005
Black				0.766***		0.413***
Hispanic				0.628***		0.427***
Black × Age				0.012		0.013
Hispanic × Age				0.025		0.029
Black × Diabetes				0.287*		0.233
Hispanic × Diabetes				0.213		0.206
Black × Age × Diabetes				−0.002		−0.009
Hispanic × Age × Diabetes				−0.042		−0.048
Education_low					1.389***	1.007***
Education_med					0.515***	0.332***
Education_low × Age					0.006	0.011
Education_med × Age					0.004	0.004
Edu_low × Diabetes					0.349*	0.199
Edu_med × Diabetes					0.022	0.011
Edu_low × Age × Diabetes					0.059**	0.059**
Edu_med × Age × Diabetes					0.015	0.014
Deceased or dropout						1.379***
High blood pressure						0.162***
Cancer						0.188***
Lung disease						0.735***
Heart disease						0.524***
Stroke						1.824***
Arthritis						0.691***
Random effects
Variance
Intercept (11)	8.211***	7.849***	7.681***	7.750***	7.503***	5.438***
Slope (22)	0.070***	0.069***	0.069***	0.069***	0.069***	0.065***
Curvature (33)	0.001***	0.001***	0.001***	0.001***	0.001***	0.001***
Covariance
Intercept, slope (21)	0.204***	0.185***	0.183***	0.184***	0.178***	0.122***
Intercept, curvature (31)	−0.034***	−0.035***	−0.035***	−0.035***	−0.035***	−0.030***
Slope, curvature (32)	0.002***	0.002***	0.002***	0.002***	0.002***	0.002***
Residual	2.047***	2.048***	2.048***	2.048***	2.049***	2.034***
Goodness of fit
AIC	365,686	364,835	364,433	364,484	363,515	355,564
BIC	365,995	365,160	364,789	364,872	363,904	356,103
−2LL	365,608	364,753	364,343	364,386	363,417	355,428
Parameters	43	47	59	65	65	109
LR test: Δχ2 (degrees of freedom)	—	855 (4)***	410 (12)***	367 (18)***	1,336 (18)***	9,325 (62)***

Notes: AIC = Akaike’s information criterion; BIC = Bayesian information criteria; LR = likelihood ratio; −2LL = −2 log-likelihood.

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

^aModel 0: Age and cohort disability trajectory model.

^bModel 1: Diabetes disability trajectory model.

^cModel 2: Diabetes and gender disability trajectory model.

^dModel 3: Diabetes and race or ethnicity disability trajectory model.

^eModel 4: Diabetes and education disability trajectory model.

^f Model 5: Diabetes and fully controlled disability trajectory model.

Age Trajectories of Physical Disability for Adults with and without Diabetes

Model 1, which added diabetes and diabetes × age, significantly improved the model fit from Model 0 ($\Delta {\chi }_{\left(4\right)}^2$ = 855, p < .001) and accounted for 4.4% ([8.211 − 7.849] / 8.211) and 1.4% ([0.070 − 0.069] / 0.070) of the variances in mean disability score and change over time, respectively. The main effect of diabetes was positive and significant, meaning that adults with diabetes had trajectories characterized by higher levels of physical disability. The positive and significant diabetes by linear age change interaction indicated that physical disability increased at a faster rate in adults with diabetes than in those without diabetes. For example, for adults from age 57 (the medium age of Cohort 1) to age 67, although nondiabetic individuals add an average of 0.72 (0.072 × 10) disabilities, individuals with diabetes add an average of 1.45 ([0.072 + 0.073] × 10) disabilities. The trajectories for adults with and without diabetes are shown in Figure 1.

Figure 1.

Physical disability trajectories in adults with and without diabetes (HRS 1998–2006). The solid lines are estimated for adults with diabetes; the dashed lines are estimated for adults without diabetes.

The Role of Gender, Race or ethnicity, and Education

Models 2–4 demonstrated the effects of diabetes on disability trajectories, which were associated with gender, race or ethnicity, and education, respectively. These models all significantly improved the model fit from Model 1, indicating that gender, race or ethnicity, and education were significant predictors for the interindividual variation in the last diabetes disability trajectory model. First, all the main effects of gender, race or ethnicity, and education were significant. Being a woman, member of a minority group (both Black and Hispanic or other), or reporting less education were associated with higher levels of physical disability compared to men or their nonminority or more highly educated counterparts. In addition, despite controlling for the significant main effect of each of the sociodemographic factors, the effect of diabetes on levels of and rate of change in physical disability remained highly significant. Most importantly, the positively significant women × diabetes (0.648, p < .001), Black × diabetes (0.287, p < .05), and education_low × diabetes (0.349, p < .05) effects in each of the three models further imply that higher levels of physical disability shown in adults with diabetes than those without diabetes was especially evident in women, Blacks, and adults with lower education. It should be noted, however, that the effects of diabetes on linear change in physical disability over time were only moderated by education (β_{(edu_low × age × diabetes)} = 0.059, p < .01), a significant three-way interaction not commonly examined in previous research. These results revealed that adults with diabetes and lower levels of education not only had higher levels of physical disability but also followed faster rates of deterioration in functioning over time than did adults with diabetes with higher education, implying that education has a stronger moderating effect on the diabetes–disability relationship than does gender and race or ethnicity.

Model 5 tested the independent effect of diabetes on physical disability trajectories when controlling for all the prior predictors plus time-varying covariates of chronic health conditions and follow-up status. Although the diabetes effects on the intercept and linear change were diminished by 61.6% ([1.699 − 0.652] / 1.699) and 35.6% ([0.073 − 0.047] / 0.073) in this model, these effects remain significant and positive. Diabetes had independent effects on both mean level (β_diabetes = 0.652, p < .001) and change in disability over time (β_{diabetes × age} = 0.047, p < .01). Furthermore, there were a significant diabetes by gender interaction (β_{diabetes × gender} = 0.389, p < .001) and a three-way interaction of diabetes by age and education (β_{edu_low × age × diabetes} = 0.059, p < .01) in this full controlled model, suggesting that physical disability differences between adults with and without diabetes were more pronounced in women than in men and that lower education may worsen the effects of diabetes on physical disability over time. The two effects cannot be explained by other sociodemographic or health factors in the model. Figures 2 and 3 provide the mean growth trajectories of physical disability in adults with and without diabetes by gender and educational level.

Figure 2.

Diabetes by gender interaction (HRS 1998–2006). The left panel presents patterns in women; the right panel presents patterns in men.

Figure 3.

Diabetes by education interaction (HRS 1998–2006). From left to right are trajectories estimated in adults of low, medium, and high educational levels.

Discussion

The skyrocketing rate of diabetes worldwide over the past few decades has fostered increasing attention in the health consequences of diabetes. Building upon the conceptual framework of the LSD perspective (Alwin & Wray, 2005) and the disablement process model (Verbrugge & Jette, 1994), this study discerned age-based disability trajectories from midlife to older age for adults with and without diabetes and examined potential moderating factors in the diabetes–disability link, adding importantly to the literature on the consequences of diabetes over the life span.

The finding of higher levels of physical disability in adults with diabetes compared to those without diabetes existed in all age cohorts of our examination. This finding provides cross-sectional evidence that diabetes was associated with more physical disability and supports previous studies using different covariates and methodologies (Bruce et al., 2003; Chou & Chi, 2005; Wray et al., 2005). The greater slopes shown in adults with diabetes than in those without diabetes provides further longitudinal evidence that changes in disability was more rapid in adults with diabetes than in those without diabetes. As illustrated in Figure 1, our results answered the query about the similarity of diabetes effects on physical disability in midlife versus older adulthood. Based on the cross-sectional evidence, in which we compared the differences between adults with and without diabetes across paralleled age cohorts, physical disability differences between adults with and without diabetes were less distinct in older ages. This result complements the results from previous research comparing physical disability between adults with and without diabetes in different age groups (Gregg, Mangione, et al., 2002). However, the observation may be simply due to differential survivorship: Those who have diabetes and are more fragile are more likely to be lost to follow-up than are their healthier counterparts. By comparing within-individual change trajectories between the two equivalent age cohorts in adults with and without diabetes, a comparison based on longitudinal evidence that accounts for attrition, the data showed that differences between adults with and without diabetes increase with age.

It is interesting to note that although estimations of physical disability based on different cohorts of adults without diabetes at parallel ages were relatively close, the estimations based on younger cohorts were always higher than those based on older cohorts in adults with diabetes. This finding provided evidence of a “dose-response effect”—duration of diabetes—of diabetes on physical disability. This is because at any given age, physical disability was estimated with a longer follow-up period for younger age cohort adults than that for older cohort adults (e.g., where age 63 is the median age 57 plus 6 years of follow-up in Cohort 1 but median age 62 plus 1 year of follow-up for Cohort 2). Given the fact that adults in different age cohorts reported random years of diabetes duration at baseline, the longer follow-up period may actually represent the accumulated years of living with diabetes.

Findings from this study add to the literature of the sociodemographic disparities in different health outcomes in adults with diabetes (Gary, McGuire, McCauley, & Brancati, 2004; Heisler et al., 2007; Hertz, Unger, & Ferrario, 2006; LeMaster, Chanetsa, Kapp, & Waterman, 2006; Quandt et al., 2005) showing that the effects of diabetes on the prevalence and growth of physical disability were not the same across social groups. As illustrated in Figure 2, the independent effect of the diabetes by gender interaction indicates that the physical disability differences between adults with and without diabetes were more pronounced in women than they were in men. The gender differences across all age cohorts provide cross-sectional evidence of the moderating effect of gender in the diabetes–disability relationship across the life span. In contrast, the moderating effect of sociodemographic factors in the diabetes–disability relationship longitudinally was only evident for education. As shown in Figure 3, the three-way interaction of education by diabetes and age suggests that lower educational levels predicted faster rates of increase in physical disability over time.

These findings support our conceptual framework and have important implications for clinicians who work with diabetes patients. First, because the adverse effects of diabetes start from midlife, health promotion programs that prevent or reduce physical disability should start from early age. In addition, our finding that the effect of diabetes increases with age may simply represent a cumulative dose-response effect of diabetes. This finding also echoes the LSD perspective, suggesting that increasing heterogeneity over the life span is expected due to the accumulation of risk or protective factors over time (Alwin & Wray, 2005; Ben-Shlomo & Kuh, 2002; Elder, 1985; Kuh et al., 2003; Riley, 1987) and highlights the importance of taking a life span approach in diabetes care. Specifically, effective diabetes care should focus not only on the proximal or concurrent conditions of the patients but also taking account of distal factors across the life span. Third, our finding that women, Blacks, and adults with lower levels of education suffered more adverse effects of diabetes than did their counterparts supports the need to develop diabetes care that specifically targets the needs of women, Blacks, and adults with lower education in the United States. Our result that comorbidities and morbidity-related attrition may mediate the racial or ethnic disparity of diabetes on disability implies that helping prevent comorbidities may be a beneficial component in guarding against a subsequent race or ethnic disparity of diabetes effects on physical disability. In contrast, comorbidities and attrition did not explain the gender differences in the diabetes–disability link observed across all age cohorts and the moderating effect of low education in that link over time. Further efforts are encouraged to disentangle the patterns found. For example, previous research has suggested that the link between lower education and poorer health outcome may be explained by factors such as health literacy (Schillinger et al., 2006; Tang, Pang, Chan, Yeung, & Yeung, 2008) and sense of personal control (Mirowsky & Ross, 2003). To reduce education-related disparities, it may be helpful for the diabetes care system to tailor health education instruction to appropriate literacy levels and design activities to promote a greater sense of personal control in adults with diabetes. As indicated in the disablement process model (Verbrugge & Jette, 1994), factors such as differential health care access or utilization across differential social groups may also contribute to the observed gender, racial or ethnic, and social class differences in the disablement process.

Limitations of the study must be acknowledged. First, our study combined “difficulty” performing a task and “unable to do” the task into a single response category, given the relatively small number of participants indicating they were unable to do a task. According to the disablement process model, however, difficulty indicates disability, whereas unable to do represents dependence. Thus, the disability measure in the present study may be confounded with dependence. Second, although self-reported disability identifies a broad range of disability in older age (Langlois et al., 1996), previous research has shown that performance-based and self-reported measures of disability may not measure the same construct (Hoeymans, Feskens, van den Bos, & Kromhout, 1996). Thus, the use of performance-based measures may provide different results. Third, few studies have explored if the validity of self-reported disability varies by gender, education, or race or ethnicity. Thus, inferences from the present study may change if the tendency for reporting disability differs by social status. In addition, this study investigated possible important moderating effects related to social status on the physical disability relationship. We found that education has a prospective effect in changing the relationship of diabetes on disability, with lower levels of education associated with faster rates of deterioration in physical function over time. It should be noted, however, that other social status factors (e.g., occupation, income, assets) may involve different mechanisms of action and have differential effects on disability (Braveman et al., 2005). Thus, this study is not conclusive in reporting a comprehensive effect of socioeconomic status in the diabetes–disability link (Feinglass et al., 2007). Future studies examining other potential moderating variables will further contribute to knowledge in this area. Similarly, because Type 1 and Type 2 diabetes have significantly different etiological mechanisms and recommended treatment or self-management regimens, factors moderating the diabetes–disability link may be different by type of diabetes (e.g., Type 1 diabetes outcomes may be more relevant to the SES of a patient’s parents in earlier life). However, limited by the HRS core interview data from 1998 to 2006, we are unable to examine any potential differences. Future studies are needed to clarify if the moderating effects of gender, race or ethnicity, and education on the relationship between diabetes and disability differ for adults living with Type 1 or Type 2 diabetes.

This longitudinal study has several strengths. To our knowledge, this is the first empirical study to examine the effects of diabetes on physical disability successively from midlife to older adulthood and to quantitatively characterize those effects with trajectories. The study’s research design and analytical methods, which take advantage of the accelerated longitudinal design of large population-based data to approximate long-term change in physical disability from midlife to older adulthood, enable us to delineate the disability trajectory for people from 51 to 105 without having to follow a sample of age 51 for 54 years. In addition, with age (longitudinal) and age cohort (cross-sectional) effects being modeled as major within-individual changes in physical function over time, we teased out effects confounded with natural aging processes and facilitated simultaneous tests of cross-sectional and longitudinal relationships between diabetes and physical disability. Our results fill an empirical void by describing the dynamic effect of diabetes on physical function and may provide a point of comparison with other populations. In addition, our study acknowledged that women suffer substantially more physical disability than do men in general, and based on that understanding, we controlled for the main effect of gender when investigating the effect of diabetes on physical disability. The interesting finding of the present study is that at any give age, differences in physical disability between adults with and without diabetes were larger in women than in men, a finding not confounded with life expectancy, highlighting the understudied gender differences in the effect of diabetes on physical disability.

In conclusion, by estimating individual growth curves for more than 20,000 older Americans, this study provided longitudinal evidence that diabetes is not only associated with higher levels of physical disability but also accelerates other aging-related processes in developing disability from midlife to older adulthood. This study also demonstrated that the diabetes–disability relationship may differ by gender, race or ethnicity, and education. Specifically, the diabetes–disability relationship is more deleterious for women, Blacks, and adults with lower education. In addition, the finding that adjustment for comorbid conditions (e.g., high blood pressure, cancer, lung disease, heart disease, stroke, and arthritis) reduced the moderating effect of race or ethnicity to nonsignificance suggests that racial or ethnic differences between adults with and without diabetes in the physical disability trajectories are largely due to the existence of comorbidities. Thus, health care designed to prevent the incidence of comorbidities in members of minority groups who are living with diabetes may promote greater racial or ethnic equality in health outcomes. Finally, the robust diabetes by gender and diabetes by age and education effects in our fully controlled model support the need for future studies to further evaluate other unexamined factors (e.g., behavioral and psychological) that may underlie the strong moderating effects of gender and education in the diabetes–physical disability relationship.

Funding

This study was supported by a Penn State Center on Population Health and Aging Level 2 grant (parent grant, National Institute on Aging, P30-AG024395).