Mind the Gap: An Analysis of Early Life Adversity and Race - - PDF document

Mind the Gap: An Analysis of Early Life Adversity and Race Disparities in Cognitive Aging in the U.S. Context Kyler J. Sherman-Wilkins, PhD Department of Sociology and Anthropology Missouri State University

2

Abstract One of the most enduring findings in health at older ages is that childhood matters. Research has explored whether racial inequalities in health over the life span result from racial differences in early life adversity. Two hypotheses have emerged to explain racial disparities in health: differential exposure (that racial differences in cognitive aging are due to differential exposure to early life adversity) differential vulnerability (which posits that group differences in cognitive aging are due to differential impacts of early life adversity across groups). Drawing on the life course perspective and data from eight waves of the Health and Retirement Study, this paper estimates growth curves of cognitive functioning and tests: a) whether blacks experience lower levels of cognitive decline over time than whites; b) whether exposure to early life adversity mediates the relationship between race and cognitive decline; and c) whether the relationship between early life adversity and cognitive aging varies by race. Findings show that race and gender are both significant predictors of episodic memory and mental status trajectory class membership. Blacks and men are more likely to belong in episodic memory trajectory classes characterized by low initial function with more rapid declines than their white and female counterparts. For mental status, blacks and women are more likely to be in lower initial functioning classes, but are advantaged with regards to rate of decline. Additionally, findings provide little evidence for the notion that early life adversity attenuates the relationship between race and/or gender and cognitive decline trajectory class membership, thereby no support for the differential exposure hypothesis was found. Lastly, lack of statistically significant interactions between both race and gender and early life adversity indicates that there is no support for the differential susceptibility hypothesis.

3 Differential Exposure and Susceptibility Perspectives on Cognitive Aging Disparities

Perhaps one of the most enduring findings in health over the lifespan is that childhood matters. Consequently, the literature examining the impacts of early life factors on adult health, broadly defined, is rapidly growing. With regards to cognitive aging/decline, previous research has investigated dimensions of early life adversity early life socioeconomic adversity (Zhang et al. 2008; Zhang et al. 2016). Recall that evidence suggest race and gender differences in cognitive aging across the life course. What is missing from this literature is an investigation of whether early life adversity at least partially explains the aforementioned black-white and male-female disparity in cognitive functioning. The closest study was conducted by Zhang and colleagues (2016) in which they find that early life factors perpetuate Black-White differences in cognitive

• impairment. What is not clear is how this shapes distinct trajectories of cognitive aging nor

whether early life adversity explains gender differences. Moreover, it is not clear whether race and/or gender moderate the relationship between early life adversity and cognitive decline Two hypotheses have emerged that connect racial/gender disparities, early life adversity, and poorer outcomes. The differential exposure hypothesis argues that the higher prevalence of early life adversity among blacks (Zhang et al. 2016) and women (Kessler and McLeod 1984; Matud 2004; Meyer et al. 2008; Turner and Lloyd 2004; Umberson et al. 1996) is what is responsible for disparities in cognitive aging. Consequently, the differential exposure hypothesis would predicts that exposure to early life adversity would at least partially explain any racial and/or gender differences in cognitive aging. Alternatively, the differential vulnerability hypothesis posits that there are not necessarily gender and racial differences in exposure to adversity (though this may be the case), but rather blacks and women react to adverse life exposures differently than their White and male counterparts, respectively. Interestingly, while women have been shown to be

4

more negatively influenced by early life adversity with outcomes such as obesity (Pudrovska et al., 2014), there is some evidence that blacks are actually less sensitive to early life adversity than Whites (Monnat et al., In Progress). It is not clear which hypothesis is more correct, but it should be noted that both hypotheses assert that gender socialization underlies the difference between men and women. Because women are socialized to be the nurturer and tend to the family, they are exposed to the more common stressors related to family and social relationships. Further, women are socialized to internalize and ruminate more than men, which would explain why they may be more adversely affected by stressful life events than men, who tend to be more detached (Rosenfield and Mouzon 2013). For blacks, their resistance to early life adversity may represent a type of resilience or hardiness perspective. The present analysis investigates the role of early life adversity in shaping later life cognitive

• functioning. To begin, I address the question as to whether early life adversity is indeed associated

with one’s cognitive aging trajectory in advanced ages and formulate the following hypothesis: Hypothesis I: early life adversity is associated with membership in cognitive decline classes characterized by low initial functioning After confirming the relationship between early life adversity and cognitive functioning in later life, I move to address the extent to which early life adversity explains the racial and gender differences discussed above. In this vein, I test my next hypothesis inspired by the differential exposure literature: Hypothesis II: the relationship between race and gender and later life cognition will be at least partially mediated by early life adversity As discussed above, though it could be the case that exposure to early life adversity explains the black-white and male-female gap in cognitive aging, alternatively, race and

5

gender could condition the effect of early life adversity on cognitive decline. Specifically, women may be more negatively impacted while blacks may be more resilient to the deleterious effects of early life adversity. In line with the differential susceptibility, I test the final hypothesis: Hypothesis IIIa: the relationship between early life adversity and later life cognitive functioning is conditioned by gender such that the relationship is more predictive of belonging to low functioning classes of cognition for women than men Hypothesis IIIb: the relationship between early life adversity and later life cognitive functioning is conditioned by race such that the relationship is less predictive of belonging to low functioning classes of cognition for blacks than whites.

Sample

Data for this study were drawn from eight waves (1998-2012) of the ongoing, nationally representative, longitudinal Health and Retirement Study (HRS). The HRS is sponsored by the National Institute on Aging (grant number NIA U01AG009740) and is conducted by the University

• f Michigan. Initiated in 1992, the HRS and its sister survey Study of Asset and Health Dynamics

Among the Oldest Old (AHEAD) were both conducted separately and biennially, before being integrated in 1998. The HRS collects measures on the health, employment, and familial conditions

• f non-institutionalized older adults aged 50+ in the United States via in-person interview or by

telephone.

6

One feature of the HRS is its inclusion of various cohort samples to better represent the range of birth cohorts from the 1900s to later in the 21st century. In total, the HRS consists of six cohorts: the AHEAD cohort (born before 1923), the HRS cohort (1931-1941), the Children of the Great Depression (CODA) cohort (1924-1936), the War Babies cohort (1942-1947), the Early Baby Boomers (1948-1953), and the Mid Baby Boomers (1954-1959). While the AHEAD, HRS, CODA, and War Babies were all included in the integrated 1998 assessment, the Early and Mid Baby Boomers were not added until 2004 and 2010, respectively. The HRS employs a multi-stage national area probability sample design. During the first stage, U.S. Metropolitan Statistical Areas (MSAs) and non-MSA counties were selected using probability proportionate to size (PPS). Second, area segments were selected from the sampled primary sampling units (PSUs). Third, once a complete enumeration of all the housing units within the boundaries of the identified area segments is completed, housing units are selected

• systematically. The fourth and last stage consist of the selection of the specific household financial
• unit. Additionally, the HRS includes an oversampling of Blacks and Latinos as well as an
• versample of Floridians to ensure adequate numbers of members of these groups.

The data source is highly appropriate for the present analysis for several reasons. First, the coverage of older Americans across a wide age range and across several assessments allows me to assess individual changes in cognitive functioning over time. Second, the oversampling of blacks ensures that I will have adequate statistical power to make meaningful comparisons across

• race. Third, in addition to containing an extensive battery of cognitive tests that tap into key

domains of cognition, the HRS also a number of early life adversity measures. To extract the data, I relied on the RAND HRS Data file, an easy to use longitudinal data set based on the HRS data as well as data “fat files” found on the HRS website (http://hrsonline.isr.umich.edu/). My

7

analytic sample consists of respondents, aged 50+ who completed baseline assessment in 1998 and provided at least one data point for the cognitive measures. Because of my interest in black- white differences, I limit my analysis to non-Latino blacks and whites leaving me with an analytic sample of 16,558. To account for the oversampling of blacks, I used sampling weights to adjust point estimates and standard errors.

Measures

As discussed earlier, there are a number of domains of cognition. As such, the cognitive measures in the HRS are designed to tap into various domains. Grounded in theory and further justified by McArdle, Fisher, and Kadlec’s (2007) quantitative analyses of cognitive measures in the HRS from 1992-2004, the HRS codebook differentiates between the following three cognitive domains: (1) episodic memory, (2) mental status, and (3) vocabulary. RAND also constructed a total cognition score which includes all cognitive measures. Because my dissertation focuses on cognitive decline, I will focus only on memory and mental status, which both tap into more fluid intelligence/cognition and ignores more crystallized vocabulary. Below, I will summarize each of the cognitive measures broken down by the cognitive domain they seek to measure, then shift to describe the other measures used in my analyses.

Episodic Memory

The results from two tasks are used to tap into memory: the Immediate and Delayed Word

• Recall. From 1998-2012, the Immediate Word Recall task calls for the respondent to recall a list
• f 10 nouns immediately after hearing the full list of nouns from the interviewer. The interviewer

cycled between four lists which consisted of different words across subsequent waves in order to minimize recall bias. The Delayed Word Recall task is the same as the Immediate Word Recall except there is an delay of about 5 minutes between the reading of the list of words and calling for

8

the respondent to recall the words. Each task is scored from 0-10, with the score reflecting the number of words successfully recalled. The RAND HRS data file includes a word recall index with scores ranging from 0-20, which consists of summing the Immediate and Delayed Recall

• scores. Following the designation assigned to this domain by McArdle and colleagues (2007), this

measure is named episodic memory.

Mental Status

Three tasks (Serial 7’s Test, Backwards Count, and Naming) make up the category of mental status, which is conceptualized as an indicator of fluid cognition designated as mental status (McArdle et al. 2007). The Serial 7’s Test involves the interviewer asking the respondent to subtract the number 7 from 100 and to continue subtracting 7 from the resulting difference over a total of five trials. The respondent was responsible for remembering the value from the previous subtraction over the five trials. Scores ranged from 0-5 and reflected the number of correct

• subtractions. Ergo, higher scores reflect better performance. Backwards counting from 20 calls for

the respondent to count backwards from 20 as quickly as possible. Scores range from 0-2, with a score of ‘0’ denoting failure in completing the task over two attempts, ‘1’ if the respondent was correct in the second attempt, and ‘2’ if the respondent was correct on the first attempt. Again, higher sores reflect better performance. Last, the naming tasks has the respondent correctly identity the date (day, month, year, and day of the week (Monday-Sunday)), the name of the object used to cut paper, the name of the prickly plant that grows in the desert, the name of the current President and Vice President of the United States. Correct responses are given a value of ‘1’ while incorrect responses are coded as ‘0.’ Together, the naming task score ranges from 0-8, with higher values corresponding to better performance. Like for episodic memory RAND HRS data file includes an index consisting of the sum of the aforementioned tasks, with scores ranging from 0-15.

9

It is worth noting a couple of things about my treatment of the dependent variables described above. First, though discrete, I treat both episodic memory and mental status as continuous variables. Though technically not continuous outcomes, because I am not interested in the pure count of correct responses but rather tapping into the constructs of memory and mental status, it makes theoretical sense to treat each variable as continuous. Additionally, prior research concludes that when the number of discrete categories is large enough (6+), discrete variables can be treated as continuous without altering substantive conclusions and will indeed produce unbiased estimates (Johnson and Creech 1986; Lei 2009; Rhemtulla, Brosseau-Liard, and Savalei 2012). Past studies on cognitive functioning using HRS data has also dichotomized cognition by first summing episodic memory and mental status into a summary total cognition score and then designating one category as cognitively impaired and the other category cognitively unimpaired based on the distribution of scores within the sample (Freedman 2002; Reuser, Willekens, and Bonneux 2011; Zhang, et al. 2016). While there certainly is value in examining cognitive impairment per se, the lack of agreed upon clinical thresholds may raise questions to the appropriateness of dichotomizing cognition scores. Indeed in the reviewed studies that examined cognitive impairment, respondents with a total cognition score of 8 and below were classified as

• impaired. These researchers cited Herzog and colleagues’ (1997) paper that discussed cognitive

impairment thresholds in the HRS. However, the 8 point cutoff discussed by Herzog and colleagues (1997) was entirely based on estimated severe cognitive impairment prevalence rates

• f community members aged 70+, and may not be suitable for studies that include respondents

younger than 70. General agreement of impairment thresholds aside, my interest in examining change in functioning over time as opposed to clinical levels of impairment obviate the need for arbitrary dichotomization.

10 Focal Covariates

Race, gender, education, and cumulative early life adversity serve as the key focal independent variables in the present analysis. Race and sex are binary variables coded ‘1’ for Blacks, ‘0’ for Whites and ‘1’ for Women, ‘0’ for Men, respectively. Years of education (0-17) is treated as a continuous measure denoting the respondent’s educational attainment. To capture early life adversity, I modeled Montez and Hayward’s (2014) and Zhang et al.’s (2016) construction of a cumulative childhood adversity measure. To construct this measure, I began by dichotomizing a number of indicators that tapped into early life adversity including whether the respondent’s mother reported less than 8 years of education (‘1’ if yes, ‘0’ if no), whether the respondent’s father completed less than 8 years of education (‘1’ if yes, ‘0’ if no), whether during childhood the family was forced to move due to financial difficulties (‘1’ if yes, ‘0’ if no), whether if during childhood, the family received aid from family after experiencing financial difficulty (‘1’ if yes, ‘0’ if no), whether the father worked in a blue collar job (‘1’ if yes, ‘0’ if no), and/or whether the respondent assessed their childhood family as poor (‘1’ if yes, ‘0’ if no). I collapsed the categories so that individuals could experience 1 to 5+ adverse childhood events (0-5). It is worth noting that the items used to construct the cumulative childhood adversity measure called for the respondent to recall experiences retrospectively, and is therefore subject to recall bias (Coughlin 1990; Hayward and Montez 2014; Zhang et al. 2016). Said bias could result in substantial measurement error and may bias estimates (Luo and Waite 2005). Though retrospective reporting of childhood events is not ideal, I argue that the cumulative index of childhood adversity that I construct for this study consists of major life events and/or salient features of childhood (e.g. moving due to financial difficulties, perceiving family as poor) and may, therefore minimize the amount of error due to recall bias. Additionally, prior research has

11

shown that retrospective indicators of socioeconomic conditions are fairly accurate (Batty et al. 2005).

Control Variables

A number of additional covariates were included in the model given there documented association with cognitive functioning. Number of chronic conditions in 1998 was treated as a continuous variable and represented the number of self-reported chronic diseases (arthritis, cancer, diabetes, heart disease, hypertension, lung disease, and/or stroke). Body mass index in 1998 (BMI) was calculated using respondent’s self-reported height and weight and the standard CDC formula 𝐶𝑁𝐽 =

𝑥𝑓𝑗𝑕ℎ𝑢 (𝑚𝑐)x 703 (ℎ𝑓𝑗𝑕ℎ𝑢 𝑗𝑜 𝑗𝑜𝑑ℎ𝑓𝑡)2. I then treated BMI as a categorical variable using CDC cutoffs

to indicate if a respondent was normal weight (BMI<25), overweight (BMI 25.0-29.9), or obese (BMI≥30) in 1998. These two measures tap into the overall health of the respondent. To capture health behaviors, I included smoking and drinking, both as categorical variables. Smoking status in 1998 was constructed using responses to inquiries into whether a respondent ever smoked or not and whether they are currently smoking. Based on these responses, I created three categories: those who never smoked were coded as ‘0,’ those who were former smokers in 1998 were coded ‘1,’ and those who were current smokers in 1998 were coded as ‘2.’ I constructed the variable drinking status in 1998 again by using two questions asked to respondents: whether respondents ever drank an alcoholic beverage, and the number of alcoholic drinks consumed daily. Based on these responses, I generated three categories: those who never drank, current light drinkers (those who drank less than one alcoholic beverages daily), and current heavy drinkers (those who drank more than one alcoholic beverage daily) examining. Age in 1998 is included (continuous, ranging from 51-105), as is marital status (married/cohabiting (reference), divorced/separated, widowed, never married). I included the proportion of waves present as a way to account for attrition.

12

Inclusion of this variable is consistent with a number of studies that have analyzed longitudinal data with significant attrition (Brown et al. 2015; Thomas 2011; Warner and Brown 2011). Lastly, a dichotomized indicator for whether the respondent had a proxy present in 1998 (0=no, 1=yes) was included to address any confounding due to respondents’ ability to conduct the full interview at that time.

Missing Data

As with all surveys, the HRS is not without missing values. I began by examining patterns

• f missing data across independent variables. The extent of missing data ranged from a low of

0.6% for smoking to a high of 19% for drinking. GGMM-LTC accommodates missing data using full information maximum likelihood, so there was no need to impute to estimate trajectories of both episodic memory and mental status. However, prior to the estimation of sample descriptive statistics and multinomial logistic regression models I addressed missing values on covariates via

• imputation. To do so, I generated 25 imputed datasets by way of multiple imputation using chained

equations (MICE), an improvement over multivariate normal imputation, in that it does not assume that all variables in the imputation model have a multivariate normal distribution (Lee and Carlin, 2010; Royston and White, 2011; White, Royston, and Wood, 2011). The data was imputed while in long form. It is often the case that the best predictor of a outcome at time X is the value of said

• utcome during previous waves and therefore the wide format should be used for imputation.

However, because I am not interested in imputing either episodic memory or mental status, the

• nly variables used in this project that wary over time, the long form is perfectly acceptable.

Following the estimation of the imputation model, the imputed datasets were pooled to arrive at more precise parameter estimates in all descriptive and multinomial logistic regression analyses.

13 Analytic Strategy

The analysis proceeded in several steps. First, I construct classes of trajectories using general growth mixture modeling. Second, I predict membership into said classes.

Analysis: Stage I

Before estimating the more complex general growth mixture models with latent trajectory classes (GGMM-LTC), I assessed the added value of utilizing GGMM-LTC over a more parsimonious unconditioned latent growth curve model. To do this, I started by estimating an unconditioned growth curve model using random-effects. I then estimated another growth curve model, adding indicators for class. I next estimated another growth curve model including class as well as interactions between age and class. The interactions were included to test the added benefit of allowing age-trajectories in cognitive functioning to vary across class, something that the GGMM- LTC would allow for. With the latent growth curve models estimated, I then compared each model using Wald tests. Based on Wald test comparisons, I concluded that GGMM-LTC was an appropriate modeling choice. Indeed, the addition of both class indicators and interactions between age and class yielded a better fitting model than the less complex unconditioned growth curve analysis. With the results of the Wald tests justifying my decision to use GGMM-LTC, I next I estimated a series GGMM-LTC models using a Stata plugin for the SAS operation ‘PROC TRAJ’ (Jones and Nagin 2013). While latent growth curve (LGC) analyses have shown to be an effective method for depicting patterns of change, this particular modeling strategy posits a mean trajectory for a single population and may not reflect reality for unobserved populations (Muthen and Muthen 2000; Muthen 2004). Below, I present the equations for a linear general growth mixture model consisting of k latent trajectory classes where Class ki equals:

14

(1) 𝑧𝑗𝑢 = 𝜃0𝑗 + 𝜃1𝑗 𝛽𝑙𝑢 + 𝜁1𝑢 (2) 𝜃0𝑗 = 𝛽0𝑙 + 𝛿0𝑙 𝑑𝑗 + 𝜂0𝑗 (3) 𝜃1𝑗 = 𝛽1𝑙 + 𝛿1𝑙 𝑑𝑗 + 𝜂1𝑗 where i indicates that parameters are specific to the ith person, t indicates the particular time point, while k denotes class membership. Equation 1 provides estimates for within-individual change across time, while both Equations 2 and 3 provide estimates for between-individual change across

• time. For Equation 1, 𝑧𝑗𝑢 represents the dependent variable—cognitive functioning—while the

intercept, slope, and covariate i are represented by 𝜃0𝑗, 𝜃1𝑗, and , respectively. Note that both the αk and γk parameters are allowed to vary across class membership, and reflect the allowed variation in trajectories (as is the case with α) as well as the variation in the influence of covariates on the specific growth factors (as is the case with γ) across classes. Lastly, 𝜁1𝑢, 𝜂0𝑗 and 𝜂1𝑗 represent error

• terms. To estimate trajectories, PROC TRAJ requires that a respondent has no missing values on

the dependent variables t baseline. Two sets of trajectory estimations were performed: one for episodic memory, and one for mental status. To decide on the appropriate number of trajectory classes, each model of k class was tested iteratively, and I made a decision based on model fit indices and substantive meaning of the classes. With regards to model fit, the class specification that yielded the smallest absolute value Bayesian information criterion (BIC), sample-size adjusted BIC, and Akaike information criterion (AIC) was deemed to be the best fitting model. In addition to relying on fit statistics values, I also considered classes that were substantively meaningful and which consisted of at least 5% of the total sample, a practice common in previous research utilizing the same method (Hill et al. 2016; Olson and Crosnoe 2017;Thomas 2011; Thomas 2012).

15

Once I ascertained the number of classes, I moved on to determine the functional form of the trajectories. Though linear models are the most parsimonious, processes of aging are rarely linear, with growth and declines occurring at differing rates across time. To account for this possibility, I tested quadratic and cubic specifications of the trajectories and compared them to each other. I first compared linear, quadratic, and cubic functional forms across fit indices and distribution of members across class. To visualize the shape of the identified trajectories, I next estimated a series of random- effects regression models using the xtreg command in Stata. The models, which separately regressed both episodic memory and mental status on age, age2, and age3, provided parameter estimates that I then used to generate graphical displays of each of the class trajectories using

• Excel. The advantage of this procedure is that it allowed me to detect statistically significant

differences in trajectory intercepts and slopes by class, thus noting meaningful differences across classes of trajectories. Figures 4.1 and 4.2 provide graphical representations of the identified trajectories for episodic memory and mental status, respectively.

Analysis: Stage II

To start, I generated weighted descriptive statistics and tested race and gender differences in exposure to early life adversity. Table 6.1 displays these estimates and provides an initial test for the hypothesis that blacks and whites, and men and women have different levels of exposure to early life adversity. Next, I examine the hypothesis that increased exposure to early life adversity is associated with belonging to trajectory classes with lower level functioning across time. Table 6.2-6.3

16

presents weighted multinomial logistic regression models predicting episodic memory and mental status class membership, respectively. For both sets of models, Model I will include early life adversity exposure and Model II will introduce model covariates to see if the relationship between early life adversity and cognitive trajectory class membership holds up. Moving on to the assessment of the role of early life adversity in explaining cognitive aging disparities in race and gender, Table 6.4 presents parameter estimates predicting membership into various episodic memory trajectory classes, while Table 6.5 does the same for mental status trajectory classes. For both Tables 6.4 and 6.5, Model I tests the hypothesis that both race and gender predict membership into cognitive aging trajectory classes. Model II then introduces early life adversity into the model and tests the differential exposure hypothesis. Full and/or partial mediation of the effects of race and/or gender (as indicated by either nonsignificant relationships

• r significant reductions in the relationship to membership to cognitive aging trajectory classes)

would provide support for the differential exposure hypothesis. Model III provides additional evidence for the differential exposure hypothesis by including additional controls with known associations to cognitive aging. And last, Model IV includes two interaction terms: race by early life adversity and gender by early life adversity and control variables to test the differential susceptibility hypothesis. Statistically significant interaction terms would provide support for the differential susceptibility hypothesis.

Results: Trajectories

Turning first to the shapes of the trajectories of cognitive functioning, Figure 4.1 presents a graphical display of classes of episodic memory trajectories by age. Class I and II have

17

trajectories that are the most disadvantaged in terms of initial functioning. Classes III and IV both have intermediate initial functioning, while Class V is the most advantaged with regards to functioning at age 51. In considering changes in episodic memory over time, all classes experience decline with age, with Classes I and II seeing the largest decline over the 39 year age range. Classes III, IV, and V are all much more similar in their rates of decline. I found that the patterns of disparity in episodic memory trajectories were different depending on the classes being compared. Whereas the differences in episodic memory Classes I and II, Classes III and IV, and Class V persisted over time, differences between Class I and Class

18

II, and the differences between Class III and IV converged over time. Indeed, the figure indicates some crossover, with Classes II and IV exhibiting higher functioning relative to Classes I and III respectively at the beginning of the age range, and exhibit slightly lower functioning at the end. The trajectories for mental status are presented in Figure 3. Like for episodic memory, results indicate a clear ordering of initial functioning at age 51, with Class I showing the lowest level of mental status and Class IV showing the highest. In terms of the patterns of change over time for mental status, Class I is the most disadvantaged in that it experiences the most rapid decline after remaining flat from ages 51-60. Class II experiences a bit of an increase in mental status performance from ages 51-60 before declining at a gradually accelerating pace beginning at age 65. I found this same pattern of increasing in functioning before declining for both Class III and IV, with the subsequent rate of decline being the slowest for Class IV.

19

Relative to each other, the disparities in mental status changed over time. In fact, I found that regardless of the classes being compared, class-specific scores in mental status diverged with increasing age. The biggest divergence was between Class I and Class IV, and is due to the relatively rapid decline experienced by Class I compared to the slow decline of Class IV. In sum, while for episodic memory disparities in functioning converge or persist depending on the classes being compared, the disparities between classes in functioning are get larger with age for mental status.

Results: Early Life Adversity and Cognitive Trajectory Class Membership

The first models test the prediction made by Hypothesis I, namely that early life adversity exposure is associated with membership in cognitive decline trajectories characterized by low initial

• functioning. Next, Hypothesis II which maintains that the relationship between race, gender, and

cognitive decline trajectories is at least partially mediated by early life adversity exposure (differential exposure hypothesis) is tested. The final set of models assess the differential vulnerability hypotheses. Hypothesis IIIa predicts that the effect of early life adversity is conditioned by race such that early life adversity is less predictive of membership in cognitive aging trajectories with low initial functioning for blacks relative to whites. Turning to the conditioning effects of gender, Hypothesis IIIb predicts that the effect of early life adversity is more predictive of belonging to low initial functioning cognitive trajectories for women relative to their male counterparts.

Descriptive Statistics

Table 6.1 presents the weighted mean number of early life adverse events respondents reported at baseline and examines differences by race (Panel A) and gender (Panel B). Considering race first,

20

blacks reported a higher number of adverse early life events than their white counterparts, a finding consistent with previous literature. With regards to gender, though women reported statistically significant higher number of early life adverse events, the magnitude of the difference is quite

• small. Though Table 6.1 does not provide a formal test of any hypothesis formulated for this

chapter, the results are consistent with previous literature regarding the distribution of blacks and women across episodic memory and mental status performance.

Multivariate Analyses

For the next set of results, I begin by examining the relationship between early life adversity and cognitive functioning trajectory class membership. Then I move on to present multinomial logistic regression estimates examining race and gender difference and how said differences are impacted by levels of exposure to early life adversity. Effects of Early Life Adversity In Table 6.2, parameter estimates for a series of multinomial logistic regressions are displayed. For both episodic memory (Panel A) and mental status (Panel B), Model I examines the effect of early life adversity while Model II introduces controls to examine the robustness of the early life adversity-trajectory membership connection. Combined, Table 6.2 provide a test of Hypothesis I. Consistent with the prediction made by Hypothesis VIII, for episodic memory, early life adversity is associated with belonging to classes characterized by lower initial functioning relative to classes

Table 6.1 Weighted Descripitve Statistics Stratified by Race, Health and Retirement Study, 1998 (N= 16,558) Panel A: Race Distribution Blacks Whites (n = 2,447) (n = 14,111) Cumulative Early Life Adversity (0-5) 2.66 (0.04) 1.94 (0.01) *** Panel B: Gender Distribution Women Men (n = 9,542) (n = 7,016) Cumulative Early Life Adversity (0-5) 2.03 (0.02) 1.98 (0.02) * Note: Standard Errors in parentheses *p <0.05; **p <0.01; ***p <0.001

21

with mid-level and higher levels of initial functioning. Indeed, for each additional early life adverse event exposure, the relative risk of belonging to Class I or Class II compared to Class III is higher by about 32% and 16%, respectively. Conversely, each additional early life adverse event a respondent is exposed to decreases the odds of belonging to Class IV over Class III (12% reduced

• dds) and belonging to Class V over Class III (24% reduced odds).

When introducing sociodemographic characteristics and health/health behavior covariates, the effects of early life adversity are greatly reduced. For Class I and Class II membership over Class III membership, the effect of early life adversity is completely attenuated as evidenced by coefficients losing statistical significance. Though the parameter estimates are still significant for predicting Class IV and Class V membership relative to Class III, the percent reduction in odds is lessened by about two-thirds for Class IV and by about half for Class V. These results suggest that, while Hypothesis VIII is supported, the relationship between early life adversity and membership in the lower trajectory classes (relative to Class III) is completely explained by adulthood socioeconomic and health profiles Patterns are roughly similar when looking at mental status. Early life adversity increases the odds for belonging to classes characterized by lower initial functioning than the reference

Table 6.2 Bivariate Associations Between Early Life Adversity and Cognitive Change Trajectory Class Membership Health and Retirement Study, 1998 (N=16,558) Panel A: Episodic Memory Class I Class II Class IV Class V Class I Class II Class IV Class V Early Life Adversity 0.28 *** 0.15 *** -0.13 ***

• 0.27 ***

0.02 0.03

• 0.04 *
• 0.12 **

(0.03) (0.02) (0.02) (0.04) (0.04) (0.02) (0.02) (0.04) Panel B: Mental Status Class I Class II Class III Class I Class II Class III Early Life Adversity 0.50 *** 0.33 *** 0.22 *** 0.12 ** 0.12 *** 0.03 (0.03) (0.02) (0.02) (0.04) (0.02) (0.02) *p <0.05; **p <0.01; ***p <0.001 Note: Standard Errors in parentheses. Model II controls for education, number of chronic conditions, smoking, drinking, BMI, age, marital status, proportion

• f waves completed, and whether a proxy was present.

Model I Model II Model I Model II

22

category, providing support for Hypothesis VIII as it pertains to mental status. For Class I relative to Class IV, there is 65% higher odds. Class II relative to Class IV membership is increased by 39% for each additional early life adverse event experienced, while it is 25% higher odds for Class III over Class IV membership. The inclusion of covariates dramatically reduces the effects of early life adversity of membership in trajectory classes. The relative risk for Class I and Class II membership over Class IV membership falls from 65% and 39% higher odds to 13% higher odds. As for the effect of early life adversity on predicting Class III membership over Class IV membership, it is no longer significant after accounting for controls. As is the case for episodic memory, the relationship between early life adversity and mental status can be largely explained by adulthood characteristics. Models Predicting Episodic Memory Trajectories Table 6.3 presents weighted estimates generated from multinomial logistic regression equations predicting episodic memory trajectory class membership. Model I shows that both race and gender predict episodic class membership, with blacks having increased risk of belonging to trajectory classes marked by low levels of functioning at age 51 compared to whites. Conversely, for gender, women have a greater relative risk of belonging to trajectory classes characterized with high levels of initial functioning. In Model II, early life adversity is included and assesses how the nature of the race and gender effects on episodic memory change once exposure to early life adversity is accounted for. Model II of Table 6.3 , then, provides a test of Hypothesis II (differential exposure hypothesis) as it pertains to episodic memory. First and foremost, early life adversity is predictive of episodic memory class membership, with higher levels of exposure linked to belonging to classes

23

characterized by lower initial functioning relative to those with higher functioning at onset. Results also show slight attenuation for the effects of race once early life adversity is controlled for. Indeed, whereas blacks have a 362% higher odds of belonging to Class I relative to Class III, when controlling for early life adversity, the increased odds drop to 300%. Such reductions in the effects are the same across comparisons. The associations between race and class membership for episodic memory trajectories remain virtually unchanged from Model I to Model II. In sum though there is some reduction in the race effect, given the small magnitude of these changes and the fact that the race effect remains strongly significant, there appears to be no support for the idea that early life adversity mediates the relationship between race and episodic memory class membership. Similarly, there is no evidence that the gender relationship is attenuated at all with the inclusion of early life adversity. Because there is no attenuation in the relationship between the race, gender, and episodic memory cognitive decline class membership, Hypothesis II does not receive any support for this particular cognitive domain.

Table 6.3 Weighted Estimates from Multinomial Logistic Regressions Predicting Episodic Memory Trajectory Class Membership Health and Retirement Study, 1998 (N= 16,558) Class I Class II Class IV Class V Class I Class II Class IV Class V Class I Class II Class IV Class V Blacks 1.53 *** 0.94 *** -0.78 *** -1.42 *** 1.39 *** 0.87 *** -0.71 *** -1.28 *** 1.02 *** 0.70 *** -0.63 *** -1.20 *** (0.10) (0.06) (0.08) (0.19) (0.10) (0.06) (0.54) (0.19) (0.11) (0.07) (0.07) (0.20) Women

• 0.63 *** -0.47 *** 0.52 ***

1.12 ***

• 0.65 ***
• 0.48 *** 0.53 ***

1.15 ***

• 0.82 *** -0.61 *** 0.66 ***

1.43 *** (0.09) (0.05) (0.05) (0.10) (0.09) (0.05) (0.05) (0.10) (0.10) (0.06) (0.05) (0.11) Early Life Adversity 0.23 *** 0.13 *** -0.13 *** -0.27 ***

• 0.002

0.008

• 0.02

0.08 * (0.03) (0.02) (0.02) (0.04) (0.04) (0.02) (0.02) (0.04) Number of Chronic Conditions 0.08 * 0.05 *

• 0.07 **
• 0.11 *

(0.04) (0.02) (0.02) (0.04) Smoking Behavior Current Smokers 0.18 0.14 †

• 0.28 ***
• 0.03

(0.14) (0.07) (0.08) (0.14) Former Smokers

• 0.16

0.04

• 0.04
• 0.03

(0.12) (0.06) (0.05) (0.10) Drinking Behavior Current Heavy Drinker

• 0.003
• 0.08

0.06 0.01 (0.20) (0.10) (0.09) (0.16) Current Light Drinker

• 0.68 *** -0.32 ***

0.08 0.20 † (0.14) (0.07) (0.06) (0.10) Body Mass Index Overweight

• 0.15
• 0.09

0.02 0.10 (0.11) (0.06) (0.05) (0.10) Obese 0.02

• 0.03
• 0.04
• 0.09

(0.13) (0.07) (0.06) (0.12) Years of Education

• 0.30 *** -0.16 *** 0.17 ***

0.32 *** (0.02) (0.01) (0.01) (0.02) Age in 1998

• 0.04 *** 0.02 ***

0.02 *** 0.04 *** (0.006) (0.003) (0.003) (0.006) Marital Status Divorced/Separated 0.16

• 0.03

0.06 0.22 (0.16) (0.09) (0.10) (0.14) Widowed 0.05 0.11 0.003 *

• 0.06

(0.13) (0.07) (0.07) (0.12) Never Married 0.35 0.41 ** 0.15 † 0.19 (0.25) (0.16) (0.15) (0.28) Proportion of Waves Present

• 1.18 *** -0.52 *** 0.66 ***

1.48 *** (0.16) (0.09) (0.10) (0.20) Proxy Present 0.55 ** 0.24 *

• 0.46 **
• 1.81 **

(0.18) (0.12) (0.16) (0.55) *p <0.05; **p <0.01; ***p <0.001; † p <0.10 Note: Standard Errors in parentheses. Model I Model II Model III

In Model III, results from fully adjusted models are presented. Addition of additional covariates reduces the effect for race which shows some evidence for attenuation. Again, these changes are quite small, and the coefficients remain highly significant suggesting that race effects are robust to inclusion of key focal covariates. Interestingly, the effects of gender, though virtually unchanged from Model I to Model II, actually increase in magnitude, albeit slightly, when introducing additional sociodemographic and health controls. Models Predicting Mental Status Trajectories Results for parameter estimates predicting mental status trajectory classes are presented in Table 6.4. Main effects for race are strong and highly statistically significant. Compared to whites, blacks have 20 times higher odds of belonging to Class I relative to Class IV, and a 6.8 times higher odds of belonging to Class II over Class IV. Though blacks also have higher odds of belonging to Class III over Class IV, it’s higher by a more moderate factor of about 2.5. Gender is also highly predictive of mental status trajectory class membership. Women are found to have a much higher odds of belonging to Class I relative to Class IV. With regards to belonging to Class II over Class IV, women have higher relative risk. Last, women have significantly higher

• dds of belonging to Class III of mental status trajectories relative to Class IV.

Table 6.4 Weighted Estimates from Multinomial Logistic Regressions Predicting Mental Status Trajectory Class Membership Health and Retirement Study, 1998 (N= 16,558) Class I Class II Class III Class I Class II Class III Class I Class II Class III Blacks 3.00 *** 1.91 *** 0.94 *** 2.79 *** 1.77 *** 0.83 *** 2.50 *** 1.61 *** 0.89 *** (0.10) (0.08) (0.08) (0.10) (0.08) (0.08) (0.11) (0.09) (0.08) Women 0.32 *** 0.44 *** 0.50 *** 0.33 *** 0.44 *** 0.50 *** 0.40 ** 0.42 *** 0.33 *** (0.09) (0.05) (0.06) (0.09) (0.05) (0.05) (0.11) (0.05) (0.05) Early Life Adversity 0.40 *** 0.29 *** 0.20 *** 0.05 0.08** 0.02 (0.04) (0.02) (0.02) (0.04) (0.02) (0.02) Number of Chronic Conditions 0.17 *** 0.16 *** 0.10 *** (0.04) (0.02) (0.02) Smoking Behavior Current Smokers 0.21 0.25 ** 0.24 ** (0.13) (0.06) (0.07) Former Smokers

• 0.18

0.03

• 0.09

(0.12) (0.06) (0.05) Drinking Behavior Current Heavy Drinker

• 0.30
• 0.29 *
• 0.34 ***

(0.22) (0.09) (0.09) Current Light Drinker

• 0.67 *** -0.37 *** -0.21 ***

(0.14) (0.07) (0.06) Body Mass Index Overweight

• 0.17
• 0.16 *
• 0.05

(0.11) (0.06) (0.05) Obese

• 0.24
• 0.07
• 0.004

(0.13) (0.07) (0.06) Years of Education

• 0.51 *** -0.30 *** -0.16 ***

(0.02) (0.01) (0.01) Age in 1998

• 0.07 *** 0.03 ***

0.04 *** (0.006) (0.003) (0.004) Marital Status Divorced/Separated 0.08 0.09

• 0.06

(0.14) (0.09) (0.09) Widowed 0.13 0.008 0.04 (0.13) (0.08) (0.06) Never Married 0.19 0.13

• 0.10

(0.30) (0.17) (0.15) Proportion of Waves Present

• 0.83 *** -0.16 *** 1.07 ***

(0.17) (0.11) (0.09) Proxy Present 1.18 *** 0.73 *** 0.53 *** (0.20) (0.15) (0.13) *p <0.05; **p <0.01; ***p <0.001 Note: Standard Errors in parentheses. Model I Model II Model III

27

In Model II, early life adversity is included into the model. As was the case for episodic memory, each additional exposure to an early life adverse event increased the relative risk of belonging to a mental status trajectory class with a higher level of initial functioning. In terms of how the inclusion of early life adversity impacted the main effects of race and gender, there was some slight attenuation of the main effects of race, though hardly large enough to provide any evidence for mediation. Further, for gender, inclusion of the early life adversity indicators left the gender effect unchanged from Model I. This pattern of findings is quite consistent with that found for episodic memory trajectory class membership. Once additional sociodemographic and health/health behavior variables are included in the equation (Model III), there is additional attenuation of the race main effect. Again, the magnitude

• f the change is not large and the results remain highly statistically significant. Whereas a

suppression effect was found for the relationship between gender and episodic memory trajectory class membership, the results for mental status trajectory class membership show the effects of gender decreasing, slightly when estimated with the fully adjusted model. Interaction Effects for Episodic Memory and Mental Status The final models provided a formal test of Hypothesis IIIa and IIIb for both episodic memory and mental status. To test for the moderating effects of race and gender on the relationship between early life adversity and episodic memory class membership, I tested for an interaction between race and early life adversity and an interaction between gender and early life adversity (Model IV of Table 6.3). Neither interaction terms were found to be significant and are therefore not displayed in the table nor discussed further. Model IV of Table 6.4 (results not shown) includes an interaction between race and early life adversity as well as an interaction between gender and early life adversity to predict mental

28

status trajectory class membership. As was the case for episodic memory, neither interaction was found to be statistically significant, suggesting that the effects of early life adversity do not vary across race or gender when predicting membership to particular classes of mental status. The lack

• f statistically significant interactions translates to a lack of support for both Hypotheses IIIa and
• IIIb. Stated another way, there is no evidence that blacks and women have differential vulnerability

to the effects of early life adversity relative to their white and male peers, respectively. Discussion The results of this chapter tell the story of whether and how early life adversity matters for later life cognitive functioning. Additionally, results herein show the extent to which early life adversity mediates and/or moderates the relationship between race/gender and cognitive decline trajectory class membership. Findings show that blacks and women are exposed to more early life adversity than their white and male counterparts. Turning to formulated hypotheses, Hypothesis I, which predicts that early life adversity is predictive of membership in both episodic memory and mental status trajectory classes characterized by low initial function, receives some support. Indeed, while highly associated with low initial status for episodic memory class membership in unadjusted models, early life adversity has an effect net of model covariates only when considering membership in the higher level categories relative to membership in the mid-level category (Class III). For mental status, the effect of early life adversity was robust for predicting relative risk of Classes I and II membership relative to Class IV, but was not predictive of relative risk of belonging to Class III over Class IV. In sum, for some class distinctions, the effects of early life adversity are explained by compositional differences across sociodemographic factors and health/health behaviors.

29

Recall that Hypothesis II predicts that any racial and gender differences in cognitive decline trajectory class membership is largely due to differential exposure to early adversity. Though black-white, and men-women differences in cognitive decline trajectory membership was found for both cognitive domains examined, inclusion of early life adversity in subsequent models did not even partially mediate either the race or gender effect. Consequently, no evidence was found to support Hypothesis II (the differential exposure hypothesis). Finally, Hypotheses IIIa and IIIb draws on the differential susceptibility hypothesis and predicts that gender and race differences in cognitive decline are due to increased susceptibility of certain racial groups and women to the stressors associated with early life adversity. Given that no statistically significant interactions were found, this hypothesis also failed to receive support. To conclude, while findings suggest that early life adversity predicts cognitive trajectory class membership, and though race and gender differences in cognitive functioning trajectory class membership are present, the higher levels of exposure to early life adversity does not seem to explain either the black-white or men-women difference in cognitive functioning trajectory class. Further race and gender do not condition the effect that early life adversity has on class membership across cognitive domain. For cognitive decline at least, race and gender differences in exposure and susceptibility to early life adversity do not account for observed racial and gender differences in cognitive decline.

Study Contributions

As the first study to use GGMM-LTC to estimate distinct trajectories of cognitive decline using the nationally representative HRS, this study serves to introduce novel modeling strategies to age-related cognitive change over time. This study, then, adds cognitive aging to other outcomes suitable for an examination using an integrated person-centered and variable-centered approach.

30

Not only do I find evidence for distinct classes of both episodic memory and cognitive decline, but I also find evidence that such a classification provides valuable insight above and beyond standard growth curve models following a more variable-centered approach. In applying this method, I engage directly with prior research attempting to predict cognitive change using standard growth models and provide added insight into the nature of trajectories of cognitive decline within the U.S. context, specifically with regards to the types of characteristics of those who belong in specific classes of cognitive decline trajectory. Secondly, this paper contributes to the growing area of research that posits the long reach

• f early life factors on later life outcomes. I explore the effects of early life adversity on cognitive

aging trajectory class membership and find that early life adversity predicts low functioning trajectories in both episodic memory and mental status. I also found significant black-white and male-female differences in cognitive trajectory class membership, differences that were robust to the inclusion of model covariates. With regards to support for the differential exposure and differential susceptibility hypotheses, this study did not find any evidence that early life adversity mediates the race/gender effect nor do race/gender conditioned the relationship between early life adversity and later life cognitive trajectory class membership. These null findings provide reasons to question the applicability of said theories to explain race and gender differences in cognitive functioning.

Study Limitations

Collectively, the results herein provide a fuller picture of cognitive aging trajectories and the determinants of those trajectories within the U.S. context. Though this dissertation increases

• ur understanding of cognitive aging within the U.S., there are a number of limitations that should

be considered. First, though I was able to examine two particular domains of cognition (episodic

31

memory and mental status), there are a number of important cognitive domains that would benefit from increased investigation. For instance, processing speed is a cognitive domain that may be directly tied to practical, everyday activities such as driving. To the extent that processing speed and other cognitive domains impede or diminish quality of life, they should receive the focus of researchers (IOM 2015). Future research should tap into more domains of cognition (e.g. attention, learning, processing speed, and spatial orientation) and examine their social patterning as well. Another significant limitation worth mentioning is that this study does not examine Latino

• populations. Ethnic disparities have been documented in previous research examining cognitive

functioning, with Latinos displaying levels of cognitive function than non-Latinos (Glymour and Manly 2008; Masel and Peek 2009). Moreover, given the rapid increase of the older adult Latino population, it is becoming increasingly important to examine the cognitive decline trajectories of Latino groups (Angel and Hogan 2004; IOM 2015). Future research should examine Latinos as well, and take care to distinguish between Latinos from various countries of origin. Next, though it is crucial to consider early life characteristics when examining health

• utcomes from a life course perspective (Ben-Shlomo and Kuh 2002), longitudinal studies of aging

are often limited in their ability to adequately tap into said characteristics. This study is no

• different. Though I rely on the same early life adversity indicators used by previous users of the

HRS (Montez and Hayward 2014; Zhang 2016), use of these measures is problematic as they are subject to recall bias (a problem that may be overstated as discussed in Chapter 3). Asking older respondents to recount events that occurred during childhood with high accuracy is most certainly an issue, so caution should be taken when assessing the true extent to which early life adversity influences later life cognitive functioning. Another issue with the early life indicators is that they do not allow for the investigation of timing and sequencing of adverse events. Critical and sensitive

32

models assert that the timing, duration, and ordering of early life events could matter for later life

• utcomes (Ben-Shlomo and Kuh; Kuh et al. 2003). Future research relying on more thorough and

sensitive information gathering techniques like the calendar method (Belli, Stafford, and Alwin 2009) would provide more validity and reliability to early life adversity measures. Lastly, though conceptually sound, the GGMM-LTC is still novel with regards to its application to cognitive aging and other outcomes related to older age (Thomas, 2012). Further refinement of the method is necessary, particularly with regards to handling attrition and addressing mortality selection. Recall that the GGMM-LTC allows for respondents with but one data point to be retained in the sample. Such a technique does little to address differential rates of attrition due to mortality and/or survey refusals. Considering that proportion of waves present is included in the models as a statistical control, I am able to adjust for differential levels of attrition. However, results should be interpreted with the understanding that mortality selection may be

• perating.

33 References Angel, J. and Hogan, D. 2004. “Population Aging and Diversity in a New Era.” Pp 1-12 in Closing the Gap: Improving the Health of Minority Elders in the New Millennium, edited by K.E. Whitfield. Washington, DC: Gerontological Society of America. Batty, G. David, Debbie A. Lawlor, Sally Macintyre, Heather Clark, and David A. Leon. 2005. “Accuracy

• f Adults’ Recall of Childhood Social Class: Findings from the Aberdeen Children of the 1950s

Study.” Journal of Epidemiology and Community Health 59:898-903. Belli, Robert F., Frank P. Stafford, and Duane F. Alwin. 2009. Calendar and Time Diary Methods in Life Course Research Thousand Oaks, CA: SAGE. Ben-Shlomo, Yoav., and Diana Kuh. 2002. “A Life Course Approach to Chronic Disease Epidemiology: Conceptual Models, Empirical Challenges, and Interdisciplinary Perspectives.” International Journal of Epidemiology 31:285-293. Brown, Tyson H., Liana J. Richardson, Taylow W. Hargrove, and Courtney S. Thomas. 2015. “Using Multiple-hierarchy Stratification and Life Course Approaches to Understand Health Inequalities: The Intersecting Consequences of Race, Gender, SES, and Age.” Journal of Health and Social Behavior 57(2):200-22. Coughlin, Steven S. 1990. “Recall Bias in Epidemiologic Studies.” Journal of Clinical Epidemiology 43(1):87-91. Freedman, Vicki A., Hakan Aykan, and Linda G. Martin. 2002. “Another Look at Aggregate Changes in Severe Cognitive Impairment: Further Investigation Into the Cumulative Effects of Three Survey Design Issues.” Journal of Gerontology: Social Sciences 57B(2):S126-S131. Glymour, M. Maria and Jennifer J. Manly. 2008. “Lifecourse Social Conditions and Racial and Ethnic Patterns of Cognitive Aging.” Neuropsychology Review 18:223-254. Herzog, A. Regula, and Robert B. Wallace. 1997. “Measures of Cognitive Functioning in the AHEAD Study.” The Journals of Gerontology Series B: Psychological and Sociail Sciences 52B(Special Issue):37-48. Hill, Terrence D., Amy M. Burdette, John Taylor, and Jacqueline L. Angel. 2016. “Religious Attendance and the Mobility Trajectories of Older Mexican Americans: An Application of the Growth Mixture Model.” Journal of Health and Social Behavior 57(1):118-134. Johnson, David Richard, and James C. Creech. 1986. “Ordinal Measures in Multiple Indicator Models: A Simulation Study of Categorization Error.” American Sociological Review 48(3):398-407. Jones, Bobby L., and Daniel S. Nagin. 2007. “Advances in Group-Based Group Trajectory Modeling and a SAS Procedure for Estimating Them.” Sociological Methods & Research 35:547-571. Jones, Bobby L., and Daniel S. Nagin. 2013 “A Note on a Stata Plugin for Estimating Group-Based Trajectory Models.” Sociological Methods & Research 42(4):608-613. Kessler, R.C. and J.D. McLeod. 1984. “Sex Differences in Vulnerability to Life Events.” American Sociological Events 49:620-631. Kuh, Diana, Yoav Ben-Scholomo, J. Lynch, J. Hllqvist, and C. Power. 2003. “Life Course Epidemiology: Glossary.” Journal of Epidemiology and Community Health 57(10):778-783. Lee, Katherine J., and John B. Carlin. “Multiple Imputation for Missing Data: Fully Conditional Specification Versus Multivariate Normal Imputation.” American Journal of Epidemiology 171(5):624-632. Lei, Pui-Wa. 2009. “Evaluating Estimation Methods for Ordinal Data in Structural Equation Modeling.” Quality & Quantity 43:495-507. Luo, Ye, and Linda J. Waite. 2005. “The Impact of Childhood and Adult SES on Physical, Mental, and Cognitive Well-Being in Later Life.” Journals of Gerontology: Series B Social Siences 60(2): 93-101. Masel, Meredith C. and M. Kristen Peek. 2009. “Ethnic Differences in Cognitive Function Over Time.” Annals of Epidemiology 19(11):778-83.

34 Matud, M.P. 2004. “Gender Differences in Stress and Coping Styles.” Personality and Individual Differences 37:1401-1415. Meyer, I.H. S. Schwartz, and D.M. Frost. 2008. “Social Patterning of Stress and Coping: Does Disadvantaged Social Status Confer More Stress and Fewer Coping Resources?” Social Science & Medicine 67:368-379. Montez, Jennifer Karas and Mark D. Hayward. 2014. “Cumulative Childhood Adversity, Educational Attainment, and Active Life Expectancy among U.S. Adults.” Demography 51:413-435. Muthen, Bengt. 2004. “Latent Variable Analysis: Growth Mixture Modeling and Related Techniques for Longitudinal Data. Pp. 345-368 in Handbook of Quantitative Methodology for the Social Sciences, edited by D. Kaplan. Newbury Park, CA: Sage. Muthen, Bengt, and Linda K. Muthen. 2000. “Integrating Person-Centered and Variable- Centered Analyses: Growth Mixture Modeling with Latent Trajectory Classes.” Alcoholism: Clinical and Experimental Research 24(6):882-891. Olson, Julie Skalamera, and Robert Crosnoe. 2017. “Are You Still Bringing Me Down? Romantic Involvement and Depressive Symptoms from Adolescence to Young Adulthood.” Journal of Health and Social Behavior 58(1):102-115. Pudrovska, Tetyana, Eric N. Reither, Ellis S. Logan, and Kyler J. Sherman-Wilkins. 2014. “Gender and Reinforcing Associations between Socioeconomic Disadvantage and Body Mass over the Life Course.” Journal of Health and Social Behavior 55(3):283-301. Ram, Nilam and K.J. Grimm. 2009. “Growth Mixture Modeling: A Method for Identifying Differences in Longitudinal Change among Unobserved Groups.” International Journal

• f Behavioral Development 33:565-76.

Reuser, Mike, Frans J. Willekens, and Luc Bonneux. 2011. “Higher Education Delays and Shortens Cognitive Impairment: A Multistate Life Table Analysis of the US Health and Retirement Study.” European Journal of Epidemiology 26:395-405. Rhemtulla, Mikje, Patricia E. Brousseau-Liard, and Victoria Savlei. 2012. Psychological Methods 17(3):354-373. Rosenfield, Sarah and Dawne Mouzon. 2013. “Gender and Mental Health.” Pp. 277-296 in Handbook of the Sociology of Mental Health: Second Edition, edited by C.S Aneshensel, J.C. Phelan and A. Bierman. New York: Springer. Royston, Patrick, and Ian R. White. “Multiple Imputation by Chained Equations (MICE): Implementation in Stata.” Journal of Statistical Software 45(4):1-20. Thomas, Patricia A. 2011. “Trajectories of Social Engagement and Limitations in Late Life.” Journal of Health and Social Behavior 52(4):430-443. Thomas, Patricia A. 2012. “Trajectories of Social Engagement and Mortality in Late Life.” Journal of Aging and Health 24(4):547-68. Turner, Ralph J., and D.A. Lloyd. 2004. “Stress Burden and the Lifetime Incidence of Psychiatric Disorder in Young Adults: Racial and Ethnic Contrasts.” Archives of General Psychiatry 61:481-488. Umberson, D., M.D. Chen, J.S. House, K. Hopkins, and E. Slaten. 1996. “The Effects of Social Relationships on Psychological Well-Being: Are Men and Women Really So Different?” American Sociological Review 61:837-857. Warner, David F. and Tyson H. Brown. 2011. “Understanding How Race/Ethnicity and Gender Define Age-Trajectories of Disability: An Intersectionality Approach.” Social Science & Medicine 72:1236-1248. Warren, John Robert, Liying Luo, Andrew Halpern-Manners, James M. Raymo, and Alberto Palloni.

• 2015. “Do Different Methods for Modeling Age-Graded Trajectories Yield Consistent and Valid

Results?” American Journal of Sociology 120(6):1809-1856. White, Ian R., Patrick Royston, and Angela M. Wood. “Multiple Imputation Using Chained Equations: Issues and Guidance for Practice.” Statistics in Medicine 30(4):377-399.

35 Zhang, Zhenmei, Danan Gu, and Mark D. Hayward. 2008. “Early Life Influences on Cognitive Impairment among Oldest Old Chinese.” Journal of Gerontology: Social Sciences 63B(1):S25-S33. Zhang, Zhenmei, Mark D. Hayward, and Yan-Liang Yu. 2016. “Life Course Pathways to Racial Disparities in Cognitive Impairment among Older Americans.” Journal of Health and Social Behavior 57(2):184-99.