Trends in the Prevalence and Disparity in Cognitive Limitations of Americans 55–69 Years Old

Abstract

Objectives

To determine whether the prevalence of cognitive limitation (CL) among Americans ages 55 to 69 years changed between 1998 and 2014, and to assess the trends in socioeconomic disparities in CL among groups defined by race/ethnicity, education, income, and wealth.

Method

Logistic regression using 1998–2014 data from the biennial Health and Retirement Study, a nationally representative data set. CL is defined as a score of 0–11 on a 27-point cognitive battery of items focused on memory. Socioeconomic status (SES) measures are classified as quartiles.

Results

In models controlling for age, gender, and previous cognitive testing, we find no significant change over time in the overall prevalence of CL, widening disparities in limitation by income and, in some cases, wealth, and improvements among non-Hispanic whites but not other racial/ethnic groups.

Discussion

Among people 55–69, rates of CL are many times higher for groups with lower SES than those with higher SES, and recent trends show little indication that the gaps are narrowing.

Nearly all research examining U.S. national trends over time in dementia or cognitive limitation (CL) focuses on the population 70 and older (for recent reviews, see Langa, 2015 and Larson et al., 2013). Instead, we examine trends in CL for the population 55–69. We do so for four reasons. First, CL is fairly common within this age group. Using the approach developed by Langa and Weir (Crimmins et al., 2011), we estimate that the prevalence of CL for people 55–59, 60–64, and 65–69 in 2010 based on data from the Health and Retirement Study (HRS) is 10.7%, 11.9%, and 14.1%, respectively.

Second, individuals who do not have dementia but do have CL (sometimes designated as cognitive impairment—no dementia [CIND] or mild cognitive impairment) make up the majority of people with CL in this age range. Those with CIND are much more likely to subsequently experience dementia than those with no such limitation (for a review, see Petersen, 2004).

The third reason is that in recent decades, the prevalences of physical and activity limitations among Americans in their 50s and early 60s have remained stable or even increased (Martin, Freedman, Schoeni, & Andreski, 2010; Martin & Schoeni, 2014; Weir, 2007). Furthermore, one study found that morbidity and mortality increased between 1997/1999 and 2011/2013 for non-Hispanic whites ages 45 to 54 years (Case & Deaton, 2015). Previous research has generally concluded that the prevalence of dementia is declining for the population roughly 70 and older in the United States (Langa, 2015). However, it could be that CL, like activity limitation, is not declining for younger age groups who have experienced changes in factors associated with cognitive health such as education, cardiovascular risk factors (including diabetes, smoking, obesity, and physical inactivity), and depression (Beydoun et al., 2014; Deckers et al., 2015). Fourth, the birth cohorts that are ages 55 to 69 during the period we analyze are particularly large because they include almost all of the Baby Boom generation, which was born between 1946 and 1964. It is therefore important to assess trends in cognitive health for these large birth cohorts as they begin to approach ages at which the incidence of dementia increases.

Another major motivation for this study is the fact that CL is strongly associated with socioeconomic factors, in particular education (Fitzpatrick et al., 2004; Kukull et al., 2002; Plassman et al., 2008; Stern, 2012; Sando et al., 2008). Moreover, risk factors for CL such as hypertension, smoking, and diabetes differ substantially by socioeconomic status (SES; Centers for Disease Control and Prevention, 2013). During the time period we examine, 1998–2014, economic inequality as measured by income (Stone et al., 2016) and wealth (e.g., Pfeffer et al., 2016) increased substantially. Furthermore, research has demonstrated that socioeconomic disparities in activity limitations widened for the older adult population in the 1980s and 1990s (Schoeni et al., 2005), and socioeconomic disparities in life expectancy widened between 1990 and 2010 (Bound et al., 2015). In this study, we investigate whether socioeconomic disparities in CL for the 55–69 group are also on the rise.

Our research has two objectives. For the population 55–69, we determine whether the prevalence of CL has changed between 1998 and 2014. We then determine for the same 16-year period, during which economic inequality increased, whether socioeconomic disparities in the prevalence of CL also widened for this age group. We examine disparities by race/ethnicity, education, income, and wealth, which represent the most salient dimensions of SES. These four dimensions are related, but they are distinct conceptually (Cutler, Lleras-Muney, and Vogl, 2011), empirically, and in terms of policy implications.

Data and Methods

Main Analyses

We use data from the 1998–2014 HRS to examine a nationally representative sample of individuals 55–69 in each year. Our main analyses exclude the 5% of sample members whose interviews were completed by proxy respondents (and thus did not take cognitive tests) or were not living in the community (sample weights that incorporate the institutionalized population are not available in 1998). Supplemental analyses assess the sensitivity of the results to these two sample restrictions.

Our indicator of CL is based on a measure developed by Langa and Weir (Crimmins et al., 2011). They use the items from the modified Telephone Interview for Cognitive Status, excluding those having to do with orientation since they are not asked of younger respondents in the HRS (Ofstedal, Fisher, & Herzog, 2005). The total score of cognitive functioning ranges from 0 to 27 points and represents the sum of: immediate word recall (0–10 points); delayed word recall (0–10 points); serial 7s (0–5 points); and backwards counting from 20 (0–2 points). A greater number of points reflect better cognitive functioning. In the Langa–Weir specification, a total score of 0–6 points is labeled as “demented,” 7–11 as “cognitively impaired but not demented (CIND),” and 12–27 as “normal.” We create an indicator of what we call “cognitive limitation” by assigning 1 if the total score is 0–11, and 0 otherwise. Too few individuals are in the 0–6 point group (1.5% among individuals 55–69 in 1998) to examine it separately.

Langa–Weir’s grouping choices are based on comparisons for people ages 71 and older of their HRS responses with clinical and neuropsychological assessments from the HRS substudy Aging Demographics and Memory Study (ADAMS). We recognize that scoring poorly on the HRS cognitive questions in and of itself does not constitute a clinical diagnosis of dementia. It is also possible that for the younger group we study, the Langa–Weir groupings are not appropriate. But we do not have evidence to suggest that a different cut-point (e.g., 12 or 13, instead of 11) would be more appropriate because neuropsychological assessments were not conducted among the younger sample members, so we conduct sensitivity analysis using a continuous specification of the total score (0–27). However, we do find that the dichotomous classification we use for people 55–69 for our main analyses is strongly associated with a subsequent very low score. Individuals classified as having CL (0–11 score) in 1998 are nearly 10 times more likely to have a score of 0–6 in 2014 than those scoring 12 or higher in 1998: 30.2% versus 3.2% (not shown in tables).

We use multivariate logistic regressions to estimate change over time in the dichotomous measure of CL. Model 1, which is used to answer the first research question, adjusts for age (5-year age categories), gender, and whether the respondent has ever taken a cognition test in a prior HRS wave. To estimate the trend over time, we include in Model 1 a continuous variable, i.e., the trend variable, that takes the value of the calendar year associated with the HRS wave, for example, 1998 for the 1998 wave. We estimate the adjusted annual percent change of CL using adjusted risk ratios (ARR), as suggested by Norton, Miller, and Kleinman (2013). In our application, the ARR is the ratio of the predicted probability of the outcome for 2014 to the predicted probability of the outcome for 1998. These predicted probabilities are calculated using the estimated parameters from models based on the data for all survey waves from 1998 to 2014 and the means of the explanatory factors (except trend) across all years combined. The calculation of the predicted probability for 2014 sets the trend variable to 2014, and that of 1998 uses 1998. The estimated annual percentage change is 100 * ln(ARR_{2014 vs 1998})/16.

We examine three racial/ethnic groups: non-Hispanic whites, non-Hispanic blacks, and Hispanics. We use information on years of education and the highest degree attained. Income is the total for the household in the calendar year prior to interview and is adjusted for household size. Household wealth is similarly adjusted and includes the total of financial assets (e.g., cash, savings accounts, stock holdings) and nonfinancial assets (e.g., homes, property, vehicles) minus the value of debts and liabilities (e.g., mortgages, credit card debt). Wealth is a particularly important dimension of SES for the individuals we examine because they are beginning to transition to retirement, during which they likely will depend heavily on their accumulated assets. Wealth is distributed much more unequally than income; some studies demonstrate that wealth inequality is 10 times greater than income inequality in the United States. Moreover, these dimensions of SES are empirically distinct. In our sample, the correlation is far from perfect between education and income (0.18), education and wealth (0.17), and income and wealth (0.33).

A goal of our study is to determine whether socioeconomic disparities in the prevalence of CL are changing over time. Specifically, we want to determine whether the relationship between the prevalence of CL and one’s relative rank within the population based on, say, education, is changing. By doing so, we reduce the influence on trends of socioeconomic groups becoming more—or less—selective over time (Bound et al., 2015; Dowd and Hamoudi, 2014). For example, the proportion of the population with less than a high school degree has declined substantially during the period we analyze, and therefore high school dropouts in 2014 are a more vulnerable population relative to their better educated peers than were high school dropouts relative to their peers in 1998. Determining relative rank for income and wealth, for example, placement within a quintile or quartile within the distribution, is straightforward because income and wealth are continuous variables with relatively little heaping on specific dollar values. However, a substantial portion of the population has the same level of schooling, in particular 12 years. To address this challenge, we rank sample persons within each survey wave in the order of the years of education, the highest degree attained, and a random number, and then classify sample persons into each quartile based on that ranking. In other words, in the case in which a percentile break point falls within the combination of single year of education and a specific highest degree, the individuals within that group are randomly assigned to a quartile either above or below the break point and in proportion to the amount needed to achieve equally sized groups. To illustrate, consider the case of 1998 in which 23.3% have less than 12 years schooling, 35.3% have 12 years, 16.3% have 13–14, 3.7% have 15, 10.1% have 16, and 11.5% have 17 or more. To construct evenly sized quartiles, 4.9% of those with 12 years (1.7% of overall sample) are assigned to the lowest quartile, 70.9% of those with 12 years (25.0% of overall sample) to the second lowest quartile, and 24.2% of those with 12 years (8.5% of overall sample) to the third quartile based on additional information on the highest degree and random number. All quartiles including income, wealth, and education are determined after a sample weight adjustment within each year. Household income and wealth are adjusted for household size by dividing by the square root of the number of people in the household.

In our multivariate analysis of disparities in trends, in addition to the basic covariates and indicators for SES, for example, race/ethnic group or income quartile, we include interactions between survey year and the SES indicator. Using the estimates from the model, we calculate the ARR and implied annual percent change for each SES group. We test for differences in trends across SES groups using the significance of the interaction terms in models with different SES reference groups. This approach is used to assess disparities in trends for racial/ethnic groups (Model 2), education quartiles (Model 3), income quartiles (Model 4), and wealth quartiles (Model 5).

We exclude from our analysis four cases with missing values for gender, and 238 cases (0.3%) missing education values. For the 2,819 cases (3.7%) with missing information on at least one of the cognition test items that make up the total score, we use the imputed cognitive measure provided by the HRS (Fisher et al., 2015). For missing values of income and wealth, we use the imputed values calculated by RAND (Hurd et al., 2016).

In all analyses, which we conduct using STATA 14 software, we take into account the complex survey design of the HRS and use the cross-sectional weights for each survey wave that allow us to analyze the data as a time series of cross-sections of a nationally representative sample.

Supplemental Analyses

We undertake five sensitivity analyses that focus on: inclusion of proxy reports, inclusion of institutionalized population, the role of learning through prior testing, specification of the outcome variable, and trends in absolute differences in CL across socioeconomic groups.

Proxy Reports

HRS sample members for whom interviews were completed by proxy respondents are incorporated in supplemental analyses. The share with proxy interviews is small and declines from 7% in 1998 to 3% in 2014, which might bias estimates of trends. We do not include data reported by proxies in the main analyses but examine sensitivity of trends to inclusion of information about sample persons with proxies by using three different approaches to measuring CL among them: (a) the proxy reports the person has ever been diagnosed with a memory condition, dementia, or senility, (b) the proxy assesses the person’s memory as fair or poor, the bottom two categories on a 5-point scale, and (c) the approach used by Langa and Weir (Crimmins et al., 2011). This last estimate of CL is based on the sum of three variables: proxy’s assessment of memory ranging from excellent (0) to poor (4; score 0–4), number of five instrumental activities of daily living that the sample person cannot do or has difficulty doing (score 0–5), and the interviewer’s assessment of difficulty completing the interview because of the sample person’s CLs (score of 0 = none, 1 = some, and 2 = prevents completion). A summary score of 0–2 is classified as normal cognition, 3–11 as CL. The third approach cannot be used for 1998 because the interviewer’s assessment was not collected; therefore, supplemental analysis using this approach is based on data from 2000 to 2014.

Institutionalized Population

The sample weights for institutionalized respondents are not available for 1998, so we do not include them in our main analysis of trends from 1998 to 2014. However, given the small size of the institutionalized population for the age group we examine (e.g., 0.6% in 2000 for individuals 55–69), we do not expect this exclusion to affect our results. Nevertheless, we conduct sensitivity analyses that include the institutionalized population, 2000–2014.

Prior Testing

Cognitive test scores may improve through repeated testing across survey waves (Rodgers et al., 2003). Research tends to conclude that if test takers improve their performance over time because they learn how to take the test, the majority of learning occurs between the first and second test, with little or no additional learning in subsequent tests. These findings motivate the inclusion in all our main analyses of a dichotomous indicator of ever having taken in a previous HRS interview any of the cognitive tests used in constructing CL. However, the reasons sample members have not taken tests in prior waves of the HRS may include having CLs, which may make it challenging to complete the interview. Alternatively, knowing from completing prior interviews with HRS that the survey asks them to recall past events and complete cognitive tests, sample members may refuse subsequent invitations to be interviewed due to the stigma associated with the difficulty of completing the survey. For these and likely other reasons, the variable representing prior testing may be endogenous, and its inclusion could bias the coefficient estimates for all other covariates, including the variable representing time trends. To investigate this issue further, we take advantage of the fact that in 1998, 2004, and 2010, the HRS added a new sample of individuals 51–56. Because it is the first HRS interview of these respondents, no HRS-specific learning has taken place. For people 51–54 in this new sample of people added to the HRS in these three years, we conducted identical analyses to those described above for people 55–69 except that in the multivariate analyses we do not include the control variable for learning (prior test) and age is specified as a continuous variable. We also estimate identically specified models using data for these three years for all respondents 51–54, not just members of the new cohort, and for all respondents 55–69.(In every wave of the HRS, there are individuals in the study who are 51–56 years old. However, unless they are part of the explicit sample additions in 1998, 2004, and 2010, they are members of the study because they are married to someone who is in the HRS age range but are not representative of all individuals 51–56 living in the United States. Our supplemental analysis of 51- to 54-year-old people also includes individuals who are part of the new cohort in the given year and are representative of persons 51–54 who are single or married to someone in the same age range or younger. Persons 51–54 who are married to someone older than 54 are already represented in the HRS in the year the new cohort is added.)

Specification of Outcome

The focus of our main analyses is CL, as indicated by using a cut-point of 11 on the 27-point battery of items focused on memory, as suggested by Langa and Weir for the 71-and-over population (Crimmins et al., 2011). We do not have clinical assessments for the younger HRS sample we analyze, so we are not able to ascertain if 11 is the best cut-point, but we have no evidence to suggest that an alternative cut-point would be better. However, to assess whether cognitive functioning more generally is changing over time, in supplemental analyses we examine trends in the total score for cognitive functioning, ranging from 0 to 27, using ordinary least squares regression analysis.

Absolute Differences

For our main analyses in which the outcome variable is dichotomous, we use logistic models, which generate estimates of trends in relative differences of CL among groups. However, given the large differences in prevalence rates of CL across socioeconomic groups, absolute differences may change substantially even though relative differences might be unchanged. For example, as we discuss below, the increases in CL between 1998 and 2014 for the second lowest and the highest education quartiles are similar, +37% and +30%, respectively, yet the absolute increases are much greater for the less educated group, +3.3 versus +0.8 percentage points. To investigate changes in absolute differences, we estimated linear probability models (LPMs) in supplemental analyses.

Results

Main Analyses

The first row of Table 1 shows unadjusted weighted prevalence of CL in each year for all individuals 55–69. The prevalence in 2014 (11.2%) was similar to that in 1998 (11.3%). The estimate of the trend over time from the multivariate logistic model for the total sample (Model 1) is reported in the first row of Table 2. The estimated annual percent increase in CL of 0.45% is not statistically significant (p = .316).

Table 1.

Unadjusted Prevalence (Weighted %) of Cognitive Limitation, Ages 55–69

		Unweighted N	All years	1998	2000	2002	2004	2006	2008	2010	2012	2014
	Overall	76,972	11.71	11.34	11.81	10.72	11.48	11.92	12.07	11.96	12.51	11.20
Race/ethnicity	non-Hispanic white	52,412	7.58	7.72	8.16	7.20	7.77	8.25	8.22	7.37	7.77	6.13
	non-Hispanic black	13,601	29.20	30.85	30.43	28.76	28.47	27.85	27.18	29.18	29.98	30.10
	Hispanic	8,835	26.64	26.12	26.85	23.33	27.01	26.28	28.06	27.31	28.27	25.23
Education	Bottom 25%	22,655	27.54	29.10	29.80	26.23	27.68	27.75	28.38	26.88	27.90	25.49
	25–50%	19,543	10.35	7.62	8.78	8.78	10.49	10.95	11.00	10.44	11.95	11.09
	50–75%	18,320	6.36	6.13	6.36	5.98	6.18	6.53	6.08	7.34	7.06	5.41
	Top 25%	16,216	2.58	2.53	2.31	1.89	1.56	2.59	2.79	3.22	3.04	2.68
Income	Bottom 25%	22,788	26.62	24.79	25.73	23.66	25.70	26.90	26.79	26.81	29.11	27.69
	25–50%	20,421	10.80	10.45	11.33	10.90	10.86	10.62	11.83	10.75	10.82	9.92
	50–75%	17,755	6.18	6.50	7.05	5.79	6.68	5.99	5.88	7.23	6.90	4.20
	Top 25%	16,008	3.23	3.62	3.14	2.53	2.70	4.18	3.73	3.05	3.17	2.96
Wealth	Bottom 25%	22,336	24.17	24.78	24.88	24.04	24.50	23.34	24.76	25.08	24.57	22.24
	25–50%	19,864	12.48	11.24	12.35	8.75	11.30	13.98	12.17	12.88	14.81	12.89
	50–75%	18,028	6.54	6.25	6.10	6.69	6.14	5.96	7.74	6.47	7.42	5.82
	Top 25%	16,744	3.64	3.10	3.87	3.41	3.98	4.39	3.58	3.40	3.20	3.81

Table 2.

Multivariate logistic estimates of annual percent change in the prevalence of cognitive limitation, overall and by SES

Model	SES group	Unweighted N	ARR	p value	Annual % change	p value testing difference in trend between groups
Model 1	Overall	76,968	1.075	.316	0.452	—	—	—
Model 2, race/ethnicity	non-Hispanic white	52,411	0.879	.045	−0.806	(ref)	—	—
	non-Hispanic black	13,600	1.012	.901	0.075	.300	(ref)	—
	Hispanic	8,833	1.033	.783	0.203	.332	.896	—
Model 3, education	Bottom 25%	22,653	0.930	.270	−0.454	(ref)	—	—
	25–50%	19,543	1.408	.000	2.139	.000	(ref)	—
	50–75%	18,318	1.033	.780	0.203	.327	.016	(ref)
	Top 25%	16,216	1.523	.016	2.629	.005	.801	.072
Model 4, income	Bottom 25%	22,786	1.175	.034	1.008	(ref)	—	—
	25–50%	20,420	0.958	.581	−0.268	.016	(ref)	—
	50–75%	17,754	0.850	.124	−1.016	.006	.356	(ref)
	Top 25%	16,008	0.983	.908	−0.107	.182	.862	.362
Model 5, wealth	Bottom 25%	22,334	0.971	.681	−0.184	(ref)	—	—
	25–50%	19,863	1.306	.006	1.669	.006	(ref)	—
	50–75%	18,027	1.085	.459	0.510	.329	.159	(ref)
	Top 25%	16,744	0.985	.918	−0.094	.900	.065	.590

Note. p value is from t-test of the null hypothesis of the trend estimate being equal to the reference group. For example, p = .300 for the equality test between non-Hispanic white and non-Hispanic black, and p = .332 for the equality test between non-Hispanic white and Hispanic. ARR = adjusted risk ratio; SES = socioeconomic status.

The disparities by SES are very large for all four measures of SES. The unadjusted rates of CL in 1998 for non-Hispanic whites, non-Hispanic blacks, and Hispanics were 7.7%, 30.9%, and 26.1%, respectively (second panel of Table 1). The results of Model 2 in Table 2 imply that prevalence of CL was unchanged for Hispanics and non-Hispanic blacks, but declined for non-Hispanic whites by 0.81% per year (p = .045). However, p values for testing differences in trend across groups indicate statistically insignificant differences for non-Hispanic whites and non-Hispanic blacks (p = .300), non-Hispanic whites and Hispanics (p = .332), and non-Hispanic blacks and Hispanics (.896).

The unadjusted rates of CL in 1998 for each quartile of education, from lowest to highest, were 29.1%, 7.6%, 6.1%, and 2.5%, respectively. The prevalence of CL increased for the second lowest quartile (2.14% per year; p < .001) and the highest quartile (2.63% per year; p = .016). The difference in the trends between the lowest and second lowest (p < .001), lowest and highest (p = .005), second lowest and second highest (p = .016), and second highest and highest quartiles (p = .072) are statistically significant.

The unadjusted rates of CL in 1998 for each quartile of income, from lowest to highest, were 24.8%, 10.5%, 6.5%, and 3.6%, respectively. The only group to experience a statistically significant change was the lowest quartile, with an annual increase of 1.01% (p = .034). This change is significantly different from the changes for the second lowest (p = .016) and second highest quartiles (p = .006).

For each quartile of wealth, from lowest to highest, the unadjusted prevalences of CL in 1998 were 24.8%, 11.2%, 6.3%, and 3.1%, respectively. The second lowest quartile experienced an annual increase of 1.67% (p = .006), whereas the trends for the other three wealth groups were not statistically significant. The estimated increase for the second lowest quartile was statistically significantly different than the change for the lowest quartile (p = .006) and the highest quartile (p = .065).

Supplemental Analyses

Proxy Responses

Estimates of the proportion of proxy responses in 2,000 indicating CL (not shown) are 5.6% based on report of diagnosis of memory-related disease (Approach #1), 20.6% based on proxy’s assessment of respondent’s memory as fair or poor on a 5-part scale (Approach #2), and 29.1% based on the Langa–Weir multifactorial indicator (Approach #3). Estimates from the multivariate logistic Model 1 that include both self-reports and proxy reports using each of the three approaches are reported in Table 3. Regardless of the approach used to determine CL status among proxies, there is no statistically significant trend in the prevalence of CL. Results from Models 2–5 that focus on SES subgroups and that incorporate proxies are reported in Supplementary Table A1. Although there are some exceptions, conclusions about disparities in trends when proxies are included are for the most part similar to those in the main analysis excluding proxies. In particular, non-Hispanic whites but not other groups experienced a decline in CL (except for Approach 1). There are also less favorable trends for the second lowest education group relative to the lowest education group; less favorable trends for the lowest income group relative to the next two higher income groups; less favorable trends for the second lowest wealth group relative to the lowest and highest wealth groups.

Table 3.

Multivariable Logistic Estimates of Annual Percent Change in the Prevalence of Cognitive Limitation, Including Proxy Responses by Three Different Approaches

Sample	Approach for proxies	Years	Unweighted N	ARR	p value	Annual % change
Self-reports		1998–2014	76,968	1.075	.316	0.452
Self- and proxy reports	Approach #1	1998–2014	81,503	1.111	.142	0.658
	Approach #2	1998–2014	81,511	1.033	.629	0.203
	Approach #3	2000–2014	71,047	0.961	.549	−0.284

Note. #1—the proxy report of the person having ever been diagnosed with a memory condition, dementia, or senility. #2—the proxy assessment of the person’s memory as fair or poor, the bottom two categories on a 5-point scale. #3—the approach used by Langa and Weir (Crimmins et al., 2011) based on three variables: proxy’s assessment of memory ranging from excellent (0) to poor (4; score 0–4); number of five instrumental activities of daily living that the sample person cannot do or has difficulty doing (score 0–5); and the interviewer’s assessment of difficulty completing the interview because of the sample person’s cognitive limitations (score of 0 = none, 1 = some, and 2 = prevents completion); A summary score of 0–2 is classified as normal cognition, 3–11 as CL. ARR = adjusted risk ratio; CL = cognitive limitation.

Institutionalized Population (Not Reported in Tables)

The estimated trend from a sample including the institutionalized population is similar to that from a sample excluding the institutionalized population. When estimating Model 1 for the waves for which sample weights are available for the institutionalized population (2000–2012) and including them with the community-dwelling population (total N = 58,617), the estimated annual change in the prevalence of CL is an insignificant +0.48% (p = .301). Excluding the institutionalized population for these same years (N = 58,516) yields a statistically insignificant estimate of +0.52% (p = .265).

Prior Testing (Not Reported in Tables)

The estimated annual change for individuals 51–54 who are members of the cohorts added to the HRS in 1998, 2004, and 2010 (N = 5,706) is +1.23% and not statistically significant (p = .242) (Model 1). When the same model is estimated for the same years (1998, 2004, and 2010) but including all sample members ages 51–54 (N = 7,089), the estimated annual change is +1.08% and again insignificant (p = .299). Among individuals 55–69 for these three years (N = 26,554), the estimated annual change is a statistically insignificant +0.48% (p = .268).

Specification of Outcome (Not Reported in Tables)

For individuals 55–69 in 1998, the average score for cognitive functioning was 16.9 points out of 27 possible. Ordinary least squares models using the continuous specification (0–27) imply that cognitive functioning worsened by 0.0186 points per year (p = .021; N = 76,968). Relative to the level of cognitive functioning in 1998, this represents an estimated annual change of −-0.11%.

Absolute Differences (Not Reported in Tables)

To estimate trends in the absolute differences of CL across SES groups, the models reported in Table 2 (Models 1–5) were re-estimated using LPM instead of logistic models. All estimates of time trends that are statistically significant based on the logistic models are also statistically significant in the corresponding LPM, and vice versa. Furthermore, the ranking of estimated trends across groups within each dimension of SES is the same for LPM and logistic models with one exception; for education, the LPM implies an unfavorable trend for the second lowest quartile that is larger than that for the highest quartile both qualitatively (0.0217 vs 0.0088 percentage point increase per 10 year increase in time, respectively) and statistically (p = .039 for test of difference). There are additional differences in tests of significance for subgroup disparities in trends between LPM and the logistic model. In the LPM, the trends differ between the highest and second lowest quartiles of education (p = .039) in favor of the highest group; trends differ between the highest and lowest income quartile (p = .052) in favor of the highest group; and the trends differ between the second highest and second lowest wealth quartile (p = .054) in favor of the higher group. These differences across groups were not found in the logistic models. One significant difference found in the logistic model (highest vs second highest education group [p = .072] in favor of the highest) is not found in the LPM model (p = .170). In sum, estimates based on LPM tend to provide stronger evidence of widening disparities than estimates based on logistic models which is expected given the very larger differences in the prevalence of CL by SES at the beginning of the period.

Summary and Discussion

Previous research indicates that the prevalence of dementia among the 70 and older population has declined in recent years (for reviews, see Langa, 2015, and Wu et al., 2017). We do not find such a favorable trend in CL for persons 55–69 between 1998 and 2014. Our main analysis, which is based on a dichotomous indicator of CL, yields no statistically significant change, and our supplemental analysis that used a continuous specification of total score on the cognitive tests finds a worsening of −0.11% per year.

The finding of no trend in CL prevalence for the 55–69 group using the dichotomous indicator is robust to inclusion of proxy respondents and the institutionalized population, and is supported by supplemental analysis that further controls for learning through prior testing by focusing solely on first-time respondents.

Our analyses by SES group (using the dichotomous measure and examining both relative and absolute differences) find that CL declined for non-Hispanic whites but not for Hispanics or non-Hispanic blacks. However, there are no statistically significant differences in trends among the racial/ethnic groups. Thus, threefold and fourfold disparities in the prevalence of CL at ages 55–69 between non-Hispanic whites and the other two groups persisted throughout the 14-year period.

For education, CL prevalence increased for the second lowest and the highest quartiles. In contrast, in earlier work that focused on trends in a variety of health outcomes for 55- to 69-year-old people from 1998 to 2012, Choi and colleagues (2016) found that CL significantly increased for all four of the educational groups they examined: less than high school, high school graduation or GED, some college, and BA or more. There were no statistically significant differences in changes in CL across the education groups in that study. In the current study, our focus has been on education quartiles and thus relative rank by education—as well as income and wealth—with the goal of reducing the influence on trends of changes in the selectivity associated with different levels of education. We find that CL for the second but not the first lowest group increased, but by 2014 there remained substantial disparities in CL between the two lowest groups. The highest education group also lost ground in terms of CL in comparison to the second highest group, but at the end of the period rates of CL remained much lower for the highest group.

CL increased for the lowest income group but was unchanged for the top 75% of the population as measured by income, and the differences between the CL trend for the lowest group versus each of the next two groups are statistically significant. Perhaps relatedly, between 1998 and 2014, median inflation-adjusted income (not shown) increased within each of the top three quartiles (by 10%, 18%, and 19% for the second lowest, second highest, and highest quartiles, respectively), but fell within the lowest quartile (by 6%).

For wealth, the prevalence of CL increased significantly within the second lowest quartile and was unchanged for the other three groups. There were statistically significant differences in trends in CL between the two lowest groups, with the advantage for the lowest, and between the highest and the second lowest group, with the advantage for the highest. Thus, the trend in CL prevalence for the second lowest wealth quartile is uniformly poor. The median wealth (inflation-adjusted) of this quartile declined by 30% ($27,578) between 1998 and 2014 (not shown). In contrast, there was a 4% decline ($9,959) for the second highest group and a 10% increase ($69,638) for the highest group. The median wealth of the lowest quartile declined 93% from $9,200 to $620 over the period, but to the extent that wealth is associated with CL, the already low wealth level of the lowest quartile in 1998 may not have a materially different association with CL from that of the extremely low wealth level for the lowest quartile in 2014.

In sum, underlying the overall lack of significant change in CL among those ages 55 to 69 are subgroup trends in CL that generally favor higher over lower income and wealth groups, but these results are not uniform. Non-Hispanic whites experienced improvements but non-Hispanic blacks and Hispanics did not.

Education is one of the most commonly used indicators of SES to examine health disparities. The results of the education subgroup analyses are mixed, which may be an artifact of the challenge of using heaped data on educational attainment and highest grade to allocate people into education quartiles for each wave. In 1998, 37% of the sample had to be randomly assigned to an education quartile, and this share increased to 47% in 2014. Despite the advantages of examining relative rank instead of the level of SES (Bound et al., 2015; Dowd and Hamoudi, 2014), implementing this approach for education is challenging and an area for future research. Nevertheless, the massive disparities across all SES variables that existed at the beginning of the period persisted.

Future research should monitor trends in cognitive function for these birth cohorts as they reach older ages at which dementia becomes more common. Examination of possible explanations of the disparities in trends that we find is beyond the scope of this descriptive paper. Especially fruitful in future work may be fitting trend models that include both SES variables and some of the conditions and behaviors associated with CL in the cross section, namely, cardiovascular disease, diabetes, depression, sensory limitation, obesity, physical inactivity, and smoking. Also, informative may be an exploration of SES disparities in trends in the prevalence of these health indicators, some of which may have been differentially affected by the economic downturn that occurred during the study period. Moreover, besides the SES variables considered here, labor force indicators might well be investigated. For example, occupation, besides reflecting economic resources, is associated with both physical activity and complexity of cognitive tasks, both of which may influence cognitive function (National Academies, 2017).

It will also be important to pursue a better understanding of how the HRS cognitive test scores for this age group relate to contemporaneous and future clinical and neuropsychological assessments of cognitive function. Another productive extension might be to analyze separately the different domains of cognition as measured by the HRS. In addition, our sensitivity analysis found that there was a worsening in average cognitive score, although the main analysis found no statistically significant change in CL using a dichotomous variable. Considering alternative cut-points for CL, as well as assessing SES disparities in trends in continuous cognition scores might yield greater insight into the results presented here.

Supplementary Material

Supplementary data is available at The Journals of Gerontology, Series B: Psychological Sciences and Social Sciences online.

Funding

The research was supported by the National Institute on Aging, grant P30 AG012846.

Conflict of Interest

None reported.