The Impact of Work and Family Life Histories on Economic Well-Being at Older Ages

Abstract

Motivated by theoretical and empirical research in life course sociology, we examine relationships between trajectories of work and family roles across the life course and four measures of economic well-being in later adulthood. Using data from the Wisconsin Longitudinal Study (WLS), and multiple trajectory-generating methods, we first identify latent trajectories of work and family roles between late adolescence and age 65. We then model economic well-being (at age 65) as a function of these trajectories and contemporaneously measured indicators of older adults’ work, family, and health statuses. Our central finding is that trajectories of work and family experiences across the life course have direct effects on later-life economic well-being, as well as indirect effects that operate through more proximate measures of work, family, and other characteristics. We argue that these findings have important implications for how social scientists conceptualize and model the relationship between later-life economic outcomes and people’s work and family experiences across the life course.

Americans’ work and family lives have become increasingly complicated over the last half century. Declining employment security (Beck 2000; Farber 2010; Hacker 2008; Kalleberg 2000), growing exposure to “bad jobs” (Findlay, Kalleberg, and Warhurst 2013; Kalleberg 2009; Mishel, Bernstein, and Allegreto 2007; Raymo et al. 2011), and increases in divorce, cohabitation, and stepfamily formation (Bianchi and Milkie 2010; Cherlin 2009; Kennedy and Ruggles 2014) have all contributed to greater variability in work and family experiences across the life course. Together, these changes have produced a “new narrative” surrounding work and family life (Hollister 2011), in which long-term stability (e.g., continuous work for a single employer or a smooth, unidirectional progression from being single, to getting married, to having children, to being an “empty nester”) has been replaced by more frequent transitions, shorter-term arrangements, and considerably more intra-individual life-course complexity.

This shift presents a challenge for researchers who are interested in studying the relationship between individuals’ work and family experiences and their well-being during later adulthood. Due to the limited availability of detailed life history data, researchers are often forced to rely on static cross-sectional measures (e.g., employment status at a certain age or marital status at a particular point in time) or blunt summary indicators (e.g., “ever divorced” or “total number of jobs held”) to characterize what are often very complicated biographies, both at work and at home. Although these analyses are of considerable scientific value, they typically lack information concerning the timing, duration, and sequencing of adults’ work and family roles (Han and Moen 1999; Mayer 2010; Pearlin et al. 2005; Umberson, Pudrovska, and Reczek 2010). We believe that this limitation could have serious consequences, especially given increasing variation in the structure, stability, and patterning of people’s working careers and family lives.

In this article, we start from the premise that variation in later-life outcomes should be understood as a consequence of heterogeneity in earlier-life experiences and that these earlier-life experiences influence later-life outcomes through complex, temporally organized pathways that are typically not well understood. To understand well-being at a given age it is therefore not enough to know about people’s circumstances at that age; one must also consider longer-term trajectories in the labor market, at home, and in other social domains. To evaluate this argument, we examine the relationship between long-term trajectories of work and family experiences and individuals’ economic outcomes during later adulthood. Our primary objective is to determine how lifelong patterns of work and family experiences combine to influence people’s economic resources at older ages—and what kinds of data and tools are needed to model these processes properly. We believe that the answers to these questions have important implications for life course theory and research.

Background

Heterogeneity in economic well-being at older ages has received considerable scholarly and policy attention in recent years. Studies have shown that income inequality among Americans over the age of 65 is increasing (Burtless 2009; O’Rand 1996), and that many older adults are not financially prepared for life after retirement (Malone et al. 2010). Projections now indicate that as much as 35% of the early baby boom cohort (individuals born between 1946-1954) will not be able to maintain their pre-retirement standard of living, even if they work full time until age 65 (Munnell, Golub-Sass, and Webb 2007). Over the last decade, social scientists have linked these trends to the shift toward greater individual responsibility for financing the retirement years (Hacker 2008), the growing importance of financial literacy (Lusardi and Mitchell 2011), and the increasing need for long-term financial planning (Ekerdt 2004).

At the same time that Americans are required to assume more responsibility for their financial security, the labor market and family contexts in which individuals’ lives unfold have undergone profound changes. Since the mid-1970s, the labor market has been characterized by reductions in job security, wage security, and access to benefits (Hacker 2008), and increased participation on the part of women (Goldin 2006). Kalleberg (2009) argues that layoffs have become a basic component of employers’ restructuring strategies and that precarious employment has spread from unskilled, less-educated segments of the labor force to all sectors of the economy. Involuntary job loss is a widely-used indicator of employment insecurity (Brand, Levy, and Gallo 2008), and recent research shows that the proportion of employees who experience involuntary job loss in a given three-year period has fluctuated between 9% and 13% since the 1980s (Farber 2005). Estimates from the Bureau of Labor Statistics indicate that more than 30 million involuntary job losses occurred between the early 1980s and 2004 (Kalleberg 2009).

Family life also became more differentiated over this period, reflecting the reorganization of men’s and women’s work and family roles in response to shifting social expectations and economic pressures (Percheski 2008). Rates of marital dissolution increased dramatically, with the sharpest increase occurring among men and women born between 1936 and 1945 (Hughes and O’Rand 2004). Although divorce tends to reduce economic well-being in later life, especially for women, the impact of divorce on economic well-being depends on a number of factors, including age at divorce, gender, number of children, and work history (O’Rand and Henretta 1999). Variation in age at first childbirth has also increased in recent decades, with implications for the timing and duration of childrearing responsibilities (Smock and Greenland 2010). Taken as a whole, these dramatic changes in family behavior and family structure have resulted in increasing variation across the life course in marital status, the number of marriages, and the presence of children in the household (Cherlin 2009)—all of which are well-documented correlates of economic well-being (see, e.g., Smock, Manning, and Gupta 1999).

Theories of cumulative advantage and the life course

The combination of increasing individual responsibility for financial preparation for old age and growing diversity and instability in work and family lives has spurred efforts to understand the ways in which experiences across the life course contribute to within-cohort variation in economic resources at older ages. The theory of cumulative advantage posits that differences in individuals’ well-being are compounded over time to generate larger differentials at later points in the life course (DiPrete and Eirich 2006). Initial individual differences—by race, gender, or socioeconomic status—give rise to unique (and often unequal) opportunity structures and life pathways, which serve to further differentiate individuals as they age. The linkage of individual lives to historical events (e.g., wars, economic downturns) and institutional arrangements (e.g., employment, family) produces individual biographies that heighten intra-cohort heterogeneity in financial resources, health, and other important later-life outcomes.

This basic theoretical perspective is embedded within the more general life course framework. Life course research recognizes that individuals’ past experiences have enduring consequences for their later-life outcomes (Elder, Johnson, and Crosnoe 2004). Throughout their lives, individuals occupy particular roles or statuses—e.g., as workers, spouses, or parents. Trajectories refer to the sequences of these roles or statuses within specific social pathways, such as employment or family life (Elder 1985). Trajectories vary depending on (1) the particular statuses of which they are composed, (2) the duration of time spent in each status, and (3) the timing and frequency of transitions between statuses. According to cumulative advantage theory, individuals experience a unique temporal ordering of roles and statuses as they age, which produce trajectories that are increasingly dissimilar from other members of their cohort. Differentiation in these trajectories contributes to heterogeneity in well-being at older ages.

The application of these ideas to the study of later-life economic outcomes is straightforward. Based on DiPrete and Eirich’s (2006) “path-dependent” model of cumulative advantage, one would expect to see differences in financial well-being between two individuals whose work and family statuses at time t (e.g., married, employed, parent, holder of a well-paying job, etc.) are identical, but whose prior history of these characteristics (in all periods leading up to time t) differs in meaningful ways. Under this model, individuals’ past experiences (at work and at home) not only influence their future experiences (at work and at home), but also have direct effects on their later-life economic outcomes. The causal pathways implied by this model are depicted in Figure 1, with arrows linking individuals’ prior work and family circumstances to their current work and family circumstances (A → A’), their current work and family circumstances to their current financial well-being (A’ → Y), and their prior work and family circumstances to their current financial well-being (A → Y).

Figure 1.

A simple two-period, path-dependent model of cumulative advantage. According to this model, financial well-being in period 2 (Y) is dependent on work and family circumstances in period 2 (A’) and work and family circumstances in period 1 (A). The arrows emanating from A indicate that prior work and family experiences have both direct (A Y) and indirect (A → A’ → Y) effects on individuals’ subsequent outcomes.

To make this model more concrete, consider the relationship between employment instability and later-life net worth. In most cases, unstable work patterns during mid-life—characterized by experiences of dismissal, part-time work, working in “bad” jobs, and/or frequent employer changes—will limit the ability to accumulate assets (Gangl 2006). These disruptions could, in turn, place constraints on a person’s future financial resources—a direct effect stemming from a gap in employment and an associated loss in earnings and investment (A → Y). If they are sufficiently severe, these disruptions could also have collateral consequences for important downstream variables (A → A’). Marriage (Amato and Beattie 2011), health (Young 2012), and future work behaviors (Gangl 2005) have all been linked to employment instability—and could all have an influence on later-life outcomes (A’ → Y). The end result is a combination of direct and indirect effects, with future work, family, and health-related variables serving as the primary mediators.

Although this conceptual model may seem intuitive, it has rarely been subjected to direct empirical testing. Instead, scholars have typically focused on estimating the relationship between A’ and Y, with little attention to direct (A → Y) and indirect (A → A’ → Y) effects emanating from individuals’ prior work and family roles (A). Examples include Farkas and O’Rand’s (1998) study of employment status and pension receipt among older women; Yabiku’s (2000) study of marriage, parenthood, and job benefits; Lum and Lightfoot’s (2003) study of health and later-life retirement savings; and Gustman and Juster’s (1996) study of marital status and income. In all of these analyses, the implicit (but generally untested) assumption is that personal biographies (with respect to employment histories, workplace opportunities and exposures, marital histories, parenthood, and so on) are irrelevant once more proximate indicators of people’s work and family circumstances (i.e., A’) are taken into account.

As a strategy for identifying factors that contribute to variation in financial well-being at older ages, we worry that this approach may obscure important variation in individuals’ prior work and family experiences. Even among older adults who are all married, who all have children, and who all work full time, there will be differences with respect to when they married; whether and when they divorced or were widowed and then remarried; when they had children; the age at which they first worked full time; how many jobs they ever lost involuntarily; how many jobs they held across their careers; how the characteristics and/or quality of their jobs varied over time; and so on. According to the life course perspective and to theories of cumulative advantage (see, e.g., Dannefer 2003; DiPrete and Eirich 2006; O’Rand and Henretta 1999), this heterogeneity in lived experiences should matter for people’s later-life economic outcomes regardless of their work and family circumstances at the time that their economic outcomes are assessed.

To thoroughly examine this hypothesis, one must have access to finely-grained data that include detailed information about people’s past work and family experiences. As we noted earlier, this requirement presents a steep hurdle. Because the collection of detailed life histories is time-consuming and expensive, many sources of data on older adults’ financial well-being offer an incomplete record of their work and family biographies. Helpful summary measures (e.g., number of years in the labor force or number of jobs held) can be constructed using nationally-representative surveys like the Health and Retirement Study (HRS), but these data generally do not offer the level of detail that would be required to characterize the sequencing, timing, structure, and context of older adults’ work and family roles. Despite well-developed theory about the importance of life-long trajectories at work and at home, actual measures that capture the richness and complexity of people’s lived experiences remain very difficult to produce.

In this article, we address this limitation using nearly 50 years’ worth of detailed life history data that were collected as a part of the Wisconsin Longitudinal Study (WLS). Drawing on the life-course perspective and the conceptual framework outlined by DiPrete and Eirich (2006), we specify a model of later-life economic outcomes that allows for complex direct and indirect effects stemming from individuals’ previous work and family circumstances (i.e., A → Y, and A → A’ → Y). Our approach—which we describe in substantially more detail below—allows us to (1) characterize the timing, duration, and sequencing of individuals’ experiences in the labor market and at home; and (2) evaluate life course ideas about the ways in which heterogeneity in later-life outcomes are related to people’s long-term trajectories of work and family experiences. In carrying out these analyses, we seek to provide new and important information about the lifelong processes that contribute to economic inequality among older adults.

Data and methods

In this section, we provide a detailed description of our data and measures. We then go on to describe the methodological approach that we use to (1) categorize people’s long-term histories of work and family experiences into a limited number of qualitatively distinct trajectory groups; and (2) relate these trajectories to respondents’ later-life economic outcomes.

Data

The Wisconsin Longitudinal Study (WLS) is a long-term study of a random sample of 10,317 men and women who graduated from Wisconsin high schools in 1957 (Sewell et al. 2004). WLS “graduates” were interviewed in 1957, 1975, 1992, 2004, and 2011; and spouses of the original respondent were interviewed in 2004. The WLS graduate sample is broadly representative of white, non-Hispanic Americans who have completed at least a high school education—a group that includes approximately two-thirds of all Americans belonging to this particular cohort (Hauser 2005). Response rates to WLS telephone and mail surveys have been consistently high. Responses were obtained from 88% of surviving graduates’ parents in 1964 and from 90% of surviving graduates in 1975. In 1993, 87% of surviving graduates responded to the telephone survey and 71% responded to the mail survey. In 2004, 78% of surviving graduates responded to the telephone survey and 76% responded to the mail survey.

The 1975 through 2004 WLS telephone surveys collected detailed information on marriage and most of the jobs that respondents ever held, allowing us to construct work and family histories in greater detail than in previous studies. In 1993 and 2004, the WLS telephone surveys obtained essentially complete employment histories for graduates covering the period 1975 through 2004 (or ages 36 through 65). The 1975, 1993, and 2004 telephone surveys obtained complete marital histories through 2004 (or age 65). Rich information on health, work, and family circumstances collected in the 2004 survey allows us to construct a comprehensive set of established temporally proximate correlates of economic well-being (A’ in Figure 1), as observed at or around age 65. Finally, the WLS collected multiple measures of economic well-being in 2004 (Y in Figure 1), including personal and household income, net worth, and home equity (as we describe in more detail below). All measures are available for both male and female respondents.

Our analyses are initially restricted to the 3,249 male and 3,785 female graduates who responded to the 1993 and 2004 telephone surveys; without this restriction we cannot observe complete trajectories of work and family roles. We then removed respondents whose birth year was not ascertained (and who may not have been a member of the modal birth cohort) and whose life history data were completely missing (n = 18). Our final analytic sample of 7,016 individuals includes about 78% of the 9,030 graduates who survived to 2004, nearly a half century after they were first enrolled in the WLS. To limit the impact of item-level missing data, we multiply imputed missing values using chained equations in Stata, where the number of imputed data sets equaled 10. Following von Hippel’s (2007) recommendations, we included observations with missing values on the dependent variables during imputation but deleted them prior to estimating our substantive models. This leads to some minor variation across models with respect to sample size (as shown below).

Measuring economic well-being at older ages (Y)

Economic well-being (or Y) is a multi-dimensional construct that can be measured in a variety of ways (see, e.g., Angel, Jiménez, and Angel 2007; Haveman et al. 2003; Kahn and Pearlin 2006; Sorokina, Webb, and Muldoon 2008; van der Klaauw and Wolpin 2008). In an effort to capture as many different aspects of economic well-being as possible (including important conceptual distinctions between “stock” and “flow” components of people’s financial portfolios), and in order to allow for the possibility that different economic outcomes exhibit different relationships to respondents’ work and family trajectories, we examined four commonly-used indicators of financial well-being—all of which were collected in the 2004 phone survey when most respondents were age 64. Independent reports obtained through the 2004 survey of WLS spouses suggest that the reliability of our measures is relatively high (e.g., the correlation between graduates’ and spouses’ answers to identically worded questions about home equity was r = 0.74). Below, we describe the four measures that we use in our analysis, and in Table 1 we provide basic descriptive statistics for each variable.

Table 1.

Financial Outcomes in 2004 Used as Dependent Variables

	Men						Women
	N	Avg	(sd)	Min	Max	Dist	N	Avg	(sd)	Min	Max	Dist
Financial Circumstances in 2004 ($ln)
Household Net Worth	3,191	12.55	(2.72)	0.00	16.30		3,725	11.76	(3.27)	0.00	16.30
Household Income	3,207	10.05	(2.71)	0.00	13.47		3,742	9.33	(3.18)	0.00	13.47
Personal Wage & Salary Income	3,206	9.98	(2.84)	0.00	13.22		3,741	8.73	(3.26)	0.00	13.22
Net Value of Home	3,186	10.58	(3.94)	0.00	14.22		3,721	9.64	(4.72)	0.00	14.22

Note: WLS sample restricted to graduates who responded to the 1975, 1993, and 2004 telephone surveys. All estimates are weighted to account for non-random sample attrition. See text for variable descriptions.

Our first measure of economic well-being is respondents’ net worth; this variable reflects the total market value of respondents’ home(s), other real estate, farms, vehicle(s), savings, and investments, minus any debt owed on those assets. We express net worth in logged dollars; for the small number of respondents with negative or zero net worth, we assigned a value of $1 before taking the log. Our second measure, home equity, is based on a series of items that assess whether respondents own their own homes, how much the home is worth, and how much they owe on their homes; respondents who did not own their homes at the time of the interview are assumed to have $0 in home equity. As with net worth, we express home equity in log dollars (after adding a small constant for respondents with $0 in home equity). Both of these measures can be considered “stock” variables as they represent more permanent assets that accumulate over time.

Our third and fourth measures of economic well-being—which can be thought of as “flow” variables—provide information about respondents’ personal and household income, respectively. Household income includes wages, farm income, interest income, social security benefits, pensions, public assistance, other government programs, child support, alimony, and other sources of income for all members of respondents’ households (to adjust for differences in household size and account for economies of scale, we divide household income by the square root of household size and reexpress in logged dollars). Personal income is the total amount that respondents received from the various sources listed above, and is also expressed in log dollars. The descriptive statistics shown in Table 1 suggest that, on average, women earn less than men and have lower net worth; that men tend to live in homes with a higher net value; and that women’s later-life financial outcomes are somewhat more variable (as indicated by larger standard deviations). These descriptive patterns are generally consistent with prior research (see, e.g., Bernasek and Shwiff 2001).

Proximate predictors of economic well-being (A’)

Our models predicting economic well-being include several measures of respondents’ labor force and family circumstances as observed at age 64—these indicators can be thought of as A’ in Figure 1. Measures of labor force conditions include employment status (employed or not employed); retirement status (partially retired, completely retired, or not retired); whether respondents’ current or most recent employer offered them pension benefits and/or health insurance; and the prestige of respondents’ current or last occupation (measured in terms of occupational earnings). Measures of family and other circumstances at age 64 include number of children; an indicator of whether respondents have any long-term activity-limiting health condition; a measure of respondents’ self-assessed overall health (excellent/very good or good/fair/poor); an indicator of marital status (married or not married); an indicator of spouses’ employment status (employed, not employed, or not married); and an indicator of spouses’ overall health (excellent/good, fair/poor/very poor, or not married). Descriptive statistics for each of these variables can be found in Table 2.

Table 2.

Current Labor Force, Family, and Other Circumstances in 2004

	Men (N = 3,238)					Women (N = 3,778)
	Mean	(SD)	Min	Max	Dist.	Mean	(SD)	Min	Max	Dist.
Labor Force Circumstances in 2004
Currently Employed (1=Yes)	0.51		0.00	1.00		0.41		0.00	1.00
Retired: Completely (1=Yes)	0.46		0.00	1.00		0.52		0.00	1.00
Retired: Partly (1=Yes)	0.27		0.00	1.00		0.19		0.00	1.00
Retired: Not at All (1=Yes)	0.27		0.00	1.00		0.28		0.00	1.00
Pension Plan on Current/Last Job (1=Yes)	0.50		0.00	1.00		0.43		0.00	1.00
Health Insurance at Current/Last Job (1=Yes)	0.53		0.00	1.00		0.42		0.00	1.00
Occupation Earnings of Current/Last Job	38.84	(20.21)	4.49	87.60		24.56	(18.12)	3.70	82.90
Family Circumstances in 2004
Current Spouse: None (1=Yes)	0.14		0.00	1.00		0.27		0.00	1.00
Current Spouse: Not Employed (1=Yes)	0.47		0.00	1.00		0.45		0.00	1.00
Current Spouse: Employed (1=Yes)	0.39		0.00	1.00		0.28		0.00	1.00
Current Spouse: Excellent/Good Health (1=Yes)	0.74		0.00	1.00		0.59		0.00	1.00
Current Spouse: Fair/Poor/Very Poor Health (1=Yes)	0.12		0.00	1.00		0.13		0.00	1.00
Number of Children	2.92	(1.64)	0.00	10.00		3.13	(1.77)	0.00	10.00
Health in 2004
Self-Assessed Health (1=Good, Fair, or Poor)	0.36		0.00	1.00		0.35		0.00	1.00
Activity Limiting Condition (1=Yes)	0.25		0.00	1.00		0.27		0.00	1.00

Note: WLS sample restricted to graduates who responded to the 1975, 1993, and 2004 telephone surveys. Distributions of continuous variables are shown in the “Dist.” column. Missing values imputed using chained equations in Stata. All estimates are weighted to account for non-random sample attrition. See text for variable descriptions.

Identifying trajectories of work and family experiences (A)

We characterize respondents’ work and family experiences (or A) in two ways: first, we model work- and marital-history data separately using three different statistical approaches that (1) estimate the number of latent trajectories in the population; (2) assign individuals to one of these trajectories using a probabilistic approach; and (3) quantify uncertainty in both (1) and (2). Second, we use work- and marital-history data to construct comparatively simple summary measures. Both types of measures are based on data collected in the 1975 through 2004 telephone surveys. The resulting work history data reflect employment circumstances at six month intervals from 1975 (age 36) through 2004; the resulting marital history data characterize marital status at six month intervals from 1955 (age 16) through 2004. Using each method, we model trajectories of employment status (i.e., employed or not); occupational standing (i.e., whether respondents’ occupations were in the bottom quartile of the occupational earnings distribution); pension and health insurance availability (i.e., whether respondents’ employers offered these benefits); and marital status (i.e., married or not).

Our statistical approaches to operationalizing trajectories of work and marital roles include repeated measures latent class analysis (Collins and Lanza 2010), latent class growth models (Nagin 2005), and growth mixture models (Muthén 2004). Using each method, and separately for each variable, we begin by estimating models that specify 1, 2, …, k trajectory groups. Based on a combination of formal (BIC, AIC, and the Lo-Mendell-Rubin (LMR)-LRT) and informal (size of the smallest group, and the distinctiveness of the trajectory shapes) criteria, we identify the best-fitting model and infer the appropriate number of trajectory groups; the results of each model also identify the latent trajectory that best describes each individual’s observed experiences. Our decision to use multiple methodologies was informed by Warren et al.’s (2013) work, which suggests that different estimators can sometimes produce different results with respect to the number, shape, and composition of latent trajectory classes. Information on model fit—presented separately for men and women and for each of the three methods that we employ—can be found in Appendix Tables A1a-c (online).

For reasons that we describe below, we also use work- and family-history data to construct a series of simpler summary measures. Our work measures—which describe the period 1975 (usually age 36) through 2004—include number of years employed; whether respondents ever worked at a job that offered pension benefits; whether respondents ever worked at a job that offered health insurance benefits; and whether respondents ever lost a job involuntarily. Our family measures—which describe the period between 1955 (usually age 16) and 2004—include number of times married; age at first marriage; and whether respondents were ever divorced or widowed. The measures of age at first marriage and age at parenthood are categorical, and express whether respondents made those transitions at the modal sex-specific ages (as opposed to making those transitions earlier or later than the average WLS respondent). These measures—which are summarized in Table 3—are meant to approximate the kinds of variables that can be constructed using simple retrospective reports.

Table 3.

Summary Measures of Labor Force and Family Experiences through 2004

	Men (N = 3,248)					Women (N = 3,785)
	Avg	(sd)	Min	Max	Dist.	Avg	(sd)	Min	Max	Dist.
Summary Measures of Labor Force Experiences through 2004
Years Employed, 1975 to 2004	26.48	(4.48)	0.00	31.00		21.52	(8.47)	0.00	31.00
Number of Employer Spells, 1975 to 2004	3.71	(2.36)	0.00	26.00		3.60	(2.41)	0.00	23.00
Ever Offered Health Insurance on Job (1=Yes)	0.92		0.00	1.00		0.72		0.00	1.00
Ever Offered Pension on Job (1=Yes)	0.88		0.00	1.00		0.69		0.00	1.00
Ever Lost a Job Involuntarily (1=Yes)	0.15		0.00	1.00		0.15		0.00	1.00
Summary Measures of Family Experiences through 2004
Number of Times Married through 2004	1.22	(0.59)	0.00	4.00		1.17	(0.54)	0.00	4.00
Age at First Marriage: Never Married (1=Yes)	0.04		0.00	1.00		0.04		0.00	1.00
Age at First Marriage: Younger than Mode (1=Yes)	0.16		0.00	1.00		0.13		0.00	1.00
Age at First Marriage: Modal Ages (1=Yes)	0.40		0.00	1.00		0.48		0.00	1.00
Age at First Marriage: Older than Mode (1=Yes)	0.41		0.00	1.00		0.35		0.00	1.00
Ever Divorced (1=Yes)	0.25		0.00	1.00		0.24		0.00	1.00
Ever Widowed (1=Yes)	0.05		0.00	1.00		0.14		0.00	1.00
Age at First Child: No Children (1=Yes)	0.07		0.00	1.00		0.07		0.00	1.00
Age at First Child: Younger than Mode (1=Yes)	0.15		0.00	1.00		0.12		0.00	1.00
Age at First Child: Modal Ages (1=Yes)	0.31		0.00	1.00		0.37		0.00	1.00
Age at First Child: Older than Mode (1=Yes)	0.47		0.00	1.00		0.44		0.00	1.00

Note: WLS sample restricted to graduates who responded to the 1975, 1993, and 2004 telephone surveys. Missing values imputed using chained equations in Stata. Distributions of continuous variables are shown in the “Dist.” column. All estimates are weighted to account for non-random sample attrition. See text for variable descriptions.

Testing for direct (A → Y) and indirect effects (A → A’ → Y)

Life course theory and theories of cumulative advantage both suggest that individuals’ prior work and family experiences should have an influence on their later-life well-being—even after controlling for more proximate, point-in-time predictors. To examine this hypothesis, we fit a series of multiple mediator models (Breen, Karlson, and Holm 2013), where the key independent variables (A) are our trajectory-based measures; where the mediators (A’) are our contemporaneously-measured work, family, and health status indicators; and where the outcomes (Y) are the four economic variables described earlier. All models were fit separately by gender to allow for the possibility that the processes in question operate differently for men and women. The specific parameters that we estimate are (1) the total effect of A on Y, (2) the direct effect of A on Y, and (3) the indirect effect of A on Y through A’. If theories of cumulative advantage hold, we would expect to see statistically significant estimates for each of these three quantities. That is, people’s past work and family experiences should have both direct and indirect effects on their later-life economic outcomes.

We fit our multiple mediator models using the user-written khb routine in Stata. The routine—which can flexibly estimate models with linear or non-linear outcomes, and which can accommodate continuous and/or categorical mediators—allows researchers to decompose the total effects associated with a given variable (or set of variables) into direct and indirect effects. The decomposition is accomplished by comparing the coefficients obtained from a “reduced model” (without mediators) to a “full model” (with mediators). The coefficients obtained from the reduced model provide an estimate of the total effects associated with the variables of interest (i.e., respondents’ trajectories of work and family experiences); the coefficients obtained from the full model provide an estimate of the direct effects; and differences between these two sets of coefficients provide an estimate of the indirect effects (i.e., the effects that are transmitted via more proximate indicators of respondents’ work and family circumstances). See Breen, Karlson, and Holm (2013) for more details.

Comparing sophisticated trajectory-based measures to simpler summary indicators

The theoretical framework outlined earlier also predicts that the temporal properties of people’s experiences (e.g., the amount of time they spent in a “good” job or the age at which they entered the labor force) should bear directly on their financial situation later in life. One corollary of this hypothesis is that detailed trajectory-based indicators—which summarize the general temporal patterning of people’s experiences—should provide additional information above and beyond what is contained in simpler summary measures. This corollary can be assessed by comparing the results obtained from models that include trajectory-based indicators to the results of models that include less-detailed characterizations of respondents’ work and family experiences. If the trajectory-based indicators provide additional explanatory power, we can infer that detailed life history data (and methods that are capable of summarizing such complex data) are needed to effectively model heterogeneity in later-life economic well-being. This finding would provide further support for the life course perspective, and would underscore the value of prospective studies that collect detailed information on work and family experiences across respondents’ life course.

Results

Before addressing our two main research questions, we first summarize the results we obtained from our three methods of generating work and family trajectories. Table 4 presents the distribution of the sample with respect to trajectory group membership, separately by sex and method, with trajectory groups ordered from highest to lowest prevalence. Each trajectory has been given a qualitative label that describes its basic shape (e.g., “Consistently employed”); the relative size of each trajectory is provided in the columns adjacent to the labels. Reassuringly, there appears to be little variation across estimators with respect to the number of estimated trajectory groups. For most of our work and family variables we identified four distinct trajectories—with women’s marital status being the only exception (the preferred LCGA model for this variable included three trajectory groups).

Table 4.

Distribution of Respondents Across Trajectory Groups, by Estimation Strategy and Gender

	Men			Women
Type of trajectory	Latent Class Analysis	Latent Class Growth Models	Growth Mixture Models	Latent Class Analysis	Latent Class Growth Models	Growth Mixture Models
Employment status
Consistently employed	51.9%	51.7%	49.2%	43.5%	42.4%	30.2%
Retired in early 60s	28.2%	27.6%	24.6%	25.2%	25.3%	23.7%
Retired in mid 50s	14.4%	14.2%	17.0%	–	–	–
Intermittent/no employment	5.5%	6.4%	9.1%	15.1%	14.7%	29.2%
Late entry	–	–	–	16.2%	17.6%	16.8%
Job quality
Never held a bad job	60.5%	60.3%	52.7%	55.3%	55.1%	54.3%
Always held a bad job	17.4%	17.6%	20.5%	16.0%	16.7%	19.7%
Transitioned into a bad job	13.2%	13.7%	13.8%	14.3%	13.6%	13.2%
Transitioned out of a bad job	9.0%	8.4%	13.0%	14.4%	14.6%	12.7%
Health insurance benefits
Received health insurance until 60s	32.8%	53.7%	42.1%	33.1%	34.5%	23.3%
Received health insurance until 50s	28.6%	25.8%	27.6%	19.2%	18.3%	25.0%
Intermittent	24.4%	10.3%	16.7%	–	–	–
Never received insurance	14.1%	10.2%	13.6%	30.5%	29.8%	34.3%
Transitioned into a job w/ coverage	–	–	–	17.1%	17.4%	17.4%
Pension benefits
Received pension benefits until 60s	42.8%	43.8%	50.4%	27.5%	28.6%	23.8%
Received pension benefits until 50s	26.9%	26.0%	23.9%	19.0%	18.3%	24.7%
Never received benefits	17.1%	16.9%	14.7%	35.1%	34.3%	35.2%
Intermittent	13.2%	13.3%	11.0%	–	–	–
Transitioned into a job w/ benefits	–	–	–	18.4%	18.8%	16.2%
Marital status
Marriage in early to mid-20s	54.2%	48.9%	47.0%	59.4%	62.7%	63.5%
Marriage after age 30	27.4%	25.2%	25.7%	17.8%	22.5%	11.1%
Multiple marriages	12.7%	16.6%	17.4%	14.6%	–	12.3%
Divorce with no remarriage	5.7%	9.2%	10.0%	8.2%	14.8%	13.1%

Note: WLS sample restricted to graduates who responded to the 1975, 1993, and 2004 telephone surveys. The qualitative labels provided in the leftmost column describe the basic shape of the trajectory. Job quality was measured using quartiles of the occupational earnings distribution; jobs in the bottom quartile were considered to be “bad” (see, e.g., Raymo et al. 2011). See text for a description of the methods we used to infer the most appropriate number of latent trajectories and trajectory group membership.

A closer inspection of the percentages presented in Table 4 reveals some interesting patterns. According to the results, the modal experience for both women and men involved marrying in their mid-20s, working full-time throughout mid-adulthood, having access to employer-provided health and pension benefits, and spending very little time in what we classify as “bad jobs”. These can all be considered fairly stable trajectories, given how few changes people made across ages with respect to work and family roles. In contrast, we also found evidence of more complicated trajectory groups, especially among women. These included a late marriage and multiple marriage group; an early retirement group; a group that moved out of bad jobs; a group that moved into them; a group that lost benefits; a group that gained them; and, among women only, a group that delayed entry into the labor market until middle adulthood. In the remainder of our analysis, we turn our attention to the relationship between these diverse—and often complicated—life pathways and respondents’ economic well-being at older ages.

The relationship between long-term trajectories and later-life outcomes

Our first research question is whether detailed trajectories of work and family experiences are related to economic outcomes in later adulthood even after we control for work, family and other circumstances as measured at the time that respondents’ financial outcomes are observed (i.e., at age 65). As mentioned above, we sought to answer this question by estimating the total, direct, and indirect effects associated with older adults’ long-term work and family experiences (using a series of multiple mediator models). Results from these analyses are presented in Table 5, separately by gender, outcome, and trajectory-generating method. To facilitate interpretation, each cell provides a p-value corresponding to an F-test of the hypothesis that the trajectory-based indicators have jointly significant (direct, total, or indirect) effects on the economic outcome in question. Bolded p-values are less than or equal to 0.05.

Table 5.

P -values from Hypothesis Tests of the Total, Direct, and Indirect Effects of Respondents’ Past Work and Family Experiences

	Men			Women
Outcome in 2004	Latent Class Analysis	Latent Class Growth Models	Growth Mixture Models	Latent Class Analysis	Latent Class Growth Models	Growth Mixture Models
Household Net Worth ($ln)
Total effect of trajectories	< 0.01	< 0.01	< 0.01	< 0.01	< 0.01	< 0.01
Direct effect of trajectories	0.01	0.02	0.02	< 0.01	< 0.01	< 0.01
Indirect effect of trajectories (via proximate measures)	0.00	< 0.01	< 0.01	0.01	< 0.01	< 0.01
Household Income ($ln)
Total effect of trajectories	< 0.01	< 0.01	< 0.01	< 0.01	< 0.01	< 0.01
Direct effect of trajectories	0.02	0.13	0.03	0.02	0.02	0.05
Indirect effect of trajectories (via proximate measures)	0.01	< 0.01	< 0.01	0.25	0.16	0.01
Personal Wage & Salary Income ($ln)
Total effect of trajectories	< 0.01	< 0.01	< 0.01	< 0.01	< 0.01	< 0.01
Direct effect of trajectories	< 0.01	0.01	0.01	< 0.01	< 0.01	< 0.01
Indirect effect of trajectories (via proximate measures)	0.02	0.03	< 0.01	< 0.01	< 0.01	< 0.01
Net Value of Home ($ln)
Total effect of trajectories	< 0.01	< 0.01	< 0.01	< 0.01	< 0.01	< 0.01
Direct effect of trajectories	0.01	0.01	0.01	< 0.01	< 0.01	< 0.01
Indirect effect of trajectories (via proximate measures)	< 0.01	< 0.01	< 0.01	0.04	0.01	0.09

Note: WLS sample restricted to graduates who responded to the 1975, 1993, and 2004 telephone surveys. Missing values on exogenous variables were imputed using chained equations in Stata. P -values for the indirect effects were calculated using the procedures described by Breen et al. (2013). Proximate measures include indicators of respondents’ work, family, and other circumstances as measured contemporaneously with the dependent variable in 2004. Bolded p -values are less than or equal to 0.05. All estimates are weighted to account for non-random sample attrition. See text for more details.

Overall, the results appear to be consistent for men and women, similar for stock and flow measures of financial well-being, and (mostly) robust to our choice of trajectory estimation strategy. In nearly every case, we found that respondents’ trajectories of work and family experiences have significant total, direct, and indirect effects on later-life financial well-being. This implies that our trajectory-based measures are predictive of older adults’ economic outcomes (i.e., past experiences matter); that the effects persist after controlling for more proximate predictors of those same outcomes (i.e., past experiences have lagged effects on adults’ later life well-being); and that the total effects associated with past work and family experiences are partially, but not completely, transmitted through more proximate channels (i.e., past experiences have effects that ripple through the life course, influencing subsequent statuses, and, in turn, later life well-being). We take this as important evidence of the long-term influence of life trajectories that has been posited, but not empirically validated, in previous research on processes of cumulative advantage.

To give a better sense for how trajectories of work and family experience are related to later-life economic outcomes, Table 6 provides results from models predicting (logged) net worth when trajectories are identified using latent class growth models; detailed results for other outcomes and/or trajectory-based estimators are available upon request. The coefficients pertaining to the trajectory-based indicators—which we entered into the model simultaneously along with our proximate predictors (see Appendix 1 for the full set of results, including results for the proximate measures)—are fairly intuitive: we observed significant and negative effects of having jobs in the bottom quartile of the occupational earnings distribution (for the entire time or at older ages), not receiving pension benefits (or only receiving such benefits later in life), and marrying multiple times or remaining unmarried throughout the majority of adulthood.

Table 6.

Detailed Estimates from Models Predicting Respondents’ Net Worth at Age 65

	Men		Women
Type of work/family trajectory	b	(s.e.)	b	(s.e.)
Employment status trajectories
Consistently employed	[ Reference Group ]		[ Reference Group ]
Retired in early 60s	−0.22	(0.12)	−0.08	(0.13)
Retired in mid 50s	−0.00	(0.18)	–	–
Intermittent/no employment	−0.53	(0.26)	−0.10	(0.25)
Late entry	–	–	−0.08	(0.16)
“Bad job” trajectories
Never held a bad job	[ Reference Group ]		[ Reference Group ]
Always held a bad job	−0.84	(0.15)	−0.96	(0.17)
Transitioned into a bad job	−0.76	(0.16)	−0.75	(0.16)
Transitioned out of a bad job	−0.32	(0.15)	−0.32	(0.15)
Health insurance trajectories
Received healthinsurance until 60s	[ Reference Group ]		[ Reference Group ]
Received health insurance until 50s	0.26	(0.20)	0.04	(0.22)
Intermittent	0.06	(0.21)	–	–
Never received health insurance	0.19	(0.33)	−0.17	(0.23)
Transitioned into a job with insurance	–	–	0.29	(0.19)
Pension benefits trajectories
Received pension benefits until 60s	[ Reference Group ]		[ Reference Group ]
Received pension benefits until 50s	−0.34	(0.18)	−0.03	(0.22)
Never received benefits	−0.68	(0.25)	−0.42	(0.22)
Intermittent	−0.32	(0.20)	–	–
Transitioned into a job with pension benefits	–	–	−0.43	(0.20)
Marital status trajectories
Marriage in early to mid-20s	[ Reference Group ]		[ Reference Group ]
Marriage after age 30	0.04	(0.12)	−0.20	(0.12)
Multiple marriages	−0.33	(0.15)	–	–
Divorce with no remarriage	−1.03	(0.21)	−1.38	(0.17)
Constant	13.19	(0.11)	12.63	(0.11)

Note: Indicators of trajectory group membership were obtained from LCGA models; results for other models are available upon request. Bolded coefficients have p -values that are less than or equal to 0.05. All estimates are weighted to account for non-random sample attrition. All of our work and family trajectory-based measures were entered into the model simultaneously. See text for further details.

We should note that in these analyses we also observed significant and negative effects of having a weak or intermittent attachment to the labor force among male respondents (i.e., men with especially volatile work histories), but not among women. Although our data do not allow us to speak to this question directly, it is possible that this finding reflects differences in the meaning of “intermittent” work between men and women of this particular cohort.

Trajectory-based measures versus simpler summary indicators

Our second empirical question is whether statistically sophisticated measures of work and family trajectories offer additional explanatory power over and above simpler summary indicators. To answer this question, we compare the model fit obtained from models that include summary indicators and proximate variables to the model fit obtained when we include summary indicators, proximate variables, and trajectory-based measures. Because the first set of models is nested within the second, we can evaluate their relative performance using a Wald test. The null hypothesis for the test is that the inclusion of the trajectory-based measures provides no additional explanatory power (i.e., the joint effect of the trajectory measures is not significant net of the summary indicators and proximate variables). Rejecting this hypothesis would imply that the timing, duration, and/or sequencing of respondents’ work and family experiences—which are not captured by summary measures and proximate variables, but which are captured by our trajectory-based indicators—have important implications for our understanding of later-life economic well-being.

Results from these model comparisons are provided in Table 7. As in the previous tables, we provide p-values indicating whether our trajectory-based measures are jointly significant after adjusting for the other variables in the model. Taken together, the results suggest that finely-grained life-history data—and trajectory-based methods that are able to exploit those data—provide useful additional information when modeling the relationship between work and family experiences and economic outcomes at older ages. Regardless of the method we used, or the financial outcomes we examined, the inclusion of trajectory-based measures improved our ability to explain respondents’ later-life well-being (14 out of the 24 hypothesis tests that we conducted produced significant results at the p < 0.05 level; and 18 out of 24 were significant at the p < 0.10 level). These findings are consistent with empirical predictions derived from the life course perspective: knowing something about the temporal characteristics of people’s work and family experiences allows for a more refined model of important later-life variables.

Table 7.

The Joint Significance of Trajectory-Based Measures in Models that Include Proximate Variables and Summary Indicators

	Men			Women
Outcome in 2004 ($ln)	Latent Class Analysis	Latent Class Growth Models	Growth Mixture Models	Latent Class Analysis	Latent Class Growth Models	Growth Mixture Models
Household net worth	0.02	0.01	< 0.01	< 0.01	< 0.01	< 0.01
Household income	0.15	0.32	0.49	0.02	0.06	0.27
Personal wage/salary income	0.02	0.09	0.29	0.01	0.03	0.16
Net value of home	< 0.01	< 0.01	< 0.01	0.06	0.10	0.01

Note : WLS sample restricted to graduates who responded to the 1975, 1993, and 2004 telephone surveys. Missing values on exogenous variables imputed using chained equations in Stata. All models include summary measures of respondents’ work and family histories as well as proximate measures of the work, family, and other circumstances (as observed in 2004). Bolded p -values are less than or equal to 0.05. All estimates are weighted to account for non-random sample attrition. See text for further details.

Discussion

According to theories of cumulative advantage (see, e.g., Crystal and Waehrer 1996; Dannefer 2003; DiPrete and Eirich 2006; O’Rand 1996; Willson, Shuey, and Elder 2007), variation in economic well-being, health, and other later-life outcomes is (at least in part) the product of people’s long-term exposure to various work and family roles. How much time people spend in different work and family roles, their longitudinal sequencing and timing, how and when they transition between them, their qualitative attributes, and whether and when they experience unanticipated “turning points” all combine to shape later life well-being. A central empirical premise of this perspective is that life course trajectories of work and family roles should matter for later-life well-being, and that the effects associated with these trajectories should operate both directly and indirectly via more proximate predictors. Unfortunately, a lack of suitable data has constrained researchers’ ability to test this perspective in a way that does justice to its conceptual and theoretical promise.

To address these issues, we examined two related research questions using rich longitudinal data from a long-term study of men and women who are now entering their retirement years. In the first part of our analysis, we asked whether adults’ lifelong trajectories of work and family experiences are independently related to their later-life financial outcomes. The answer to this question is a decided “yes.” Consistent with theoretical expectations, we found a significant and robust relationship between our trajectory-based measures of adults’ work and family experiences and several dimensions of later-life economic well-being. Our results suggest that these effects operate both directly and indirectly, with proximate indicators of work, family, and other circumstances serving as mediating variables. These findings were insensitive to the trajectory-based estimation strategies that we used (latent class analysis, latent class growth models, and growth mixture models), and held for each of the financial outcomes that we examined.

In the second part of our analysis, we asked whether statistically sophisticated measures of work and family trajectories—typically based on expensive, prospectively-collected longitudinal data—offer additional information, net of simpler summary measures that can be constructed from more conventional (and easier to obtain) retrospective reports. The answer to this question is also “yes.” For most of the models that we estimated, we found that the inclusion of trajectory-based measures improved our ability to predict economic outcomes—even after controlling for point-in-time variables and broad summary measures. Our interpretation of this finding is that simpler summary measures—such as number of years worked, whether the respondent ever received pension or health insurance benefits, whether the respondent ever lost a job involuntarily, and whether the respondent was ever divorced or widowed—do not capture the full temporal dimensions of life course pathways that matter for financial well-being later in life.

We believe that these findings have important theoretical implications. Our results suggest that people’s work and family experiences have consequences that reverberate across the life course, impacting their future experiences, and ultimately shaping their circumstances as older adults. One can think of this chain of events as a non-Markovian process. Work and family experiences early in life set the stage for work and family experiences later in life, but they also have lagged effects on later-life well-being. This result is consistent with theories of cumulative advantage (DiPrete and Eirich 2006), which suggest that earlier advantages and disadvantages accumulate with time, interacting with unexpected life events (e.g., divorce, job loss) to produce increasing within-cohort variation with respect to financial well-being, health, and other adult outcomes (see, e.g., Dannefer 2003; Ferraro and Shippee 2009; Garbarski 2014; Giudici and Pallas 2014; O’Rand and Hamil-Luker 2005; Petersen et al. 2011; Willson et al. 2007).

Our results also have important practical implications. Researchers and funding agencies are currently making major investments in large-scale longitudinal studies. At the same time, methods for developing complex measures of trajectories are proliferating (Warren et al. 2013). This is despite a lack of sound evidence about whether trajectories matter (for financial outcomes) net of proximal measures; a lack of consensus about how to model trajectory data; and a lack of information about whether simpler summary measures might serve as well as more advanced trajectory-based indicators. Our findings help to address these concerns. In this article, we showed that trajectories do matter, that simpler summary measures are not a perfect substitute for trajectory-based indicators, and that neither of these findings depend on which methods researchers use to model people’s long-term work and family experiences. Social scientists who collect, analyze, and/or disseminate life history data should find these results encouraging.

This is not to say that more research on this topic is not needed. In our analyses we considered an important but limited set of work and family characteristics, and economic outcomes at a single point in time, using a specific cohort of respondents. Future work could build on our results by (1) examining a wider array of variables (including people’s experiences with disability, illness, childrearing, and death in the family); (2) incorporating time-varying measures of economic well-being; and (3) carrying out additional sub-group analyses. The ultimate objective would be to provide a dynamic portrait of people’s work and family lives that traces interdependencies across social spheres and allows scholars to “see” cumulative advantage processes play out in real time. We are optimistic that the analytic framework that we employed here-modeling trajectories of work and family characteristics that summarize the temporal properties of people’s experiences in these two domains—will prove useful as social scientists move forward with this important line of research.

Conclusion

Economic well-being during the retirement years is increasingly dependent upon people’s own planning and resources. This shift toward increased responsibility has taken place within the context of growing heterogeneity in people’s working careers and family lives. Within the past 50 years, Americans have experienced fundamental changes in the structure of the labor market, in employment relationships, in women’s labor force participation rates, and in rates of marital dissolution. These changes have important consequences for how researchers think about and model the relationship between later-life well-being and people’s longer-term work and family experiences. Our research shows that economic well-being among older Americans is best understood as the product of a lifetime of accumulated experiences in the two most important social roles that most people occupy: their families and their jobs. We take this as good evidence in support of the life course perspective.

Biographies

About the Authors

Andrew Halpern-Manners is an assistant professor in the Department of Sociology at Indiana University. His research interests include social demography, the sociology of education, stratification, and various issues related to longitudinal data analysis. He has recently published articles in these areas in Sociological Methods & Research, Public Opinion Quarterly, Social Forces, and Demography.

John Robert Warren is Professor of Sociology at the University of Minnesota and Training Director of the Minnesota Population Center. He has been involved with the WLS since 1994; is co-leading an effort to field follow-up surveys of the 1980 High School & Beyond cohorts; and is co-PI of a project to harmonize, fully link, document, and disseminate data and metadata from the Current Population Surveys. He is Editor of Sociology of Education through 2016.

James Raymo is Professor of Sociology at the University of Wisconsin-Madison where he is also the current director of the Center for Demography and Ecology. His research focuses on evaluating patterns and potential consequences of demographic changes associated with rapid population aging in Japan. He has published widely on key features of recent family change in Japan, including delayed marriage, extended coresidence with parents, and increases in premarital cohabitation, shotgun marriages, and divorce.

D. Adam Nicholson is a doctoral student in the Department of Sociology at Indiana University. His areas of specialization include political sociology, social stratification, and quantitative methods. His current research examines the effects of voter identification laws on voter turnout by minority groups, as well as trends in political trust and efficacy.

Technical appendix

In recent years, many researchers have turned to finite mixture models as a way to model change within people over time. Repeated measure latent class analysis (RMLCA), latent class growth analysis (LCGA), and growth mixture models (GMMs) are three different types of finite mixture model (Muthén 2004). In all three cases, individuals are assigned to latent trajectory groups on the basis of their observed experiences or behaviors. Each trajectory group represents a qualitatively distinct latent subpopulation within the data, composed of individuals with relatively similar measurements on some age-sequenced variable (e.g., employment status or marital status or job quality). GMMs allow for residual variation within these trajectory groups (i.e., within-class heterogeneity or random effects), whereas RMLCA and LCGA assume homogeneity conditional on trajectory group membership.

In these models, the number of underlying trajectories, their shape, their prevalence within the population, and the assignment of individuals to them are all inferred from the data; see Nagin (2005), Muthén (2004), and Bollen and Curran (2006) for technical details and relevant formulas. Models with increasing numbers of trajectory classes are typically fit in an iterative fashion. The “preferred model” is then determined using a combination of formal and informal criteria. In many applications, researchers base decisions about the number of trajectories on measures of model fit (e.g., BIC, AIC, and/or the Lo-Mendell-Rubin Likelihood Ratio Test), the share of cases assigned to the smallest trajectory group, interpretability, the distinctiveness of the trajectories, and model convergence (see, e.g., Marcell et al. 2011). We followed a similar approach in our analysis, relying on a blend of formal and informal criteria to identify our preferred specifications. Fit statistics for each of the models that we estimated can be found in Appendix Tables A1a-A1c (online).

A key concern when estimating finite mixture models is the presence of local maxima. Likelihood functions for mixture models are exceedingly complex (e.g., not always concave), which can sometimes lead to false solutions and/or non-convergence problems. In an effort to guard against these possibilities, we randomly generated at least 400 sets of starting values for every LCGA, RMLCA, and GMM that we estimated, using the automated STARTS routine available in recent versions of Mplus (Muthén and Muthén 2010). We then optimized the 100 best sets, as identified by a comparison of the log-likelihoods. Our final estimates were obtained from model estimates in which the highest log-likelihood was replicated at least once, suggesting that the global maximum likelihood solution was successfully reached. If the best log-likelihood was not replicated, we increased the number of start values to 1,000 and the number of second-stage optimizations to 250, and then reestimated the model. Similar approaches have been used in other substantive applications (see, e.g., McLeod and Fettes 2007).

Because our primary substantive interest was in quantifying the direct and indirect effects associated with people’s work and family trajectories, we employed a “two-step” approach where we first identified trajectories (in Mplus) and then, in a separate set of models, related people’s trajectories to their later-life outcomes (in Stata using the khb routine). Although similar models can be estimated in one-step using Mplus, valid estimates of the indirect effects associated with respondents’ trajectory groups cannot be obtained. To determine whether our choice of estimation strategy influenced our results, we refit our trajectory models after including our list of proximate predictors as covariates and our measures of financial well-being as distal outcomes. Results from these analyses were similar with respect to the number of trajectories, their shape and prevalence within the sample, and their relationship to people’s later-life well-being (i.e., the general patterns that we observed in Table 6).

Table A1a.

Fit Statistics by Model Specification, Gender, and Type of Work/Family Trajectory (LCA) [To be made available online]

	Men						Women
	BIC	AIC	LMR	Entropy	Distinct	Size	BIC	AIC	LMR	Entropy	Distinct	Size
Employment status
k = 1	108004.3	108004.3	–	–			246182.9	245809.1	–	–
k = 2	80968.9	80968.9	0.00	0.99			180998.0	180244.1	0.00	0.98
k = 3	70367.1	70367.1	0.00	0.99			157213.0	156079.0	0.00	0.98
k = 4	62267.5	62267.5	0.00	0.99			139044.7	137530.6	0.00	0.98
k = 5	59029.9	59029.9	0.00	0.99			130359.2	128465.1	0.00	0.98
Job quality
k = 1	179098.7	178734.3	–	–			165596.0	165226.1	–	–
k = 2	86170.3	85435.4	0.00	0.99			93419.4	92673.4	0.00	0.98
k = 3	68541.3	67435.9	0.00	0.99			77056.6	75934.6	0.00	0.96
k = 4	56999.5	55523.6	0.00	0.99			64298.3	62800.2	0.00	0.96
k = 5	52182.6	50336.2	0.06	0.98			59449.7	57575.5	0.09	0.95
Health insurance
k = 1	190714.8	190350.1	–	–			280784.9	280411.1	–	–
k = 2	133390.5	132655.0	0.00	0.99			180356.3	179602.5	0.00	0.99
k = 3	104616.7	103510.4	0.00	0.99			150087.5	148953.6	0.00	0.99
k = 4	93240.0	91762.9	0.00	0.99			128626.9	127112.9	0.00	0.98
k = 5	84131.0	82283.1	0.02	0.99			119423.1	117529.1	0.00	0.98
Pension benefits
k = 1	220187.8	219823.1	–	–			276866.8	276493.0	–	–
k = 2	144156.0	143420.4	0.00	0.99			175945.7	175191.9	0.00	0.99
k = 3	116867.2	115760.9	0.00	0.99			142566.8	141432.8	0.00	0.99
k = 4	102139.7	100662.6	0.00	0.99			122146.6	120632.6	0.00	0.98
k = 5	92646.7	90798.7	0.42	0.98			113949.8	112055.8	0.56	0.98
Marital status
k = 1	242874.2	242267.3	–	–			320220.2	319597.9	–	–
k = 2	159746.1	158526.1	0.00	1.00			198124.8	196873.9	0.00	1.00
k = 3	137632.7	135799.8	0.00	0.99			173563.7	171684.4	0.00	0.99
k = 4	123035.5	120589.6	0.00	0.99			154398.8	151890.9	0.00	0.99
k = 5	113527.4	110468.4	0.00	1.00			141495.1	138358.6	0.78	0.99

Note: Models with k = 1, …, 5 trajectories were fit for each work/family charcteristic, separately by gender. The “LMR” column provides p -values from a Lo-Mendell-Rubin likelihood ratio test, comparing a model with k trajectories to a model with k – 1. The null hypothesis for the test is that the two models are equivalent. Filled boxes in the “Distinct” column indicate that trajectories identified by the model were substantively distinct; empty boxes indicate otherwise. We made these determinations by visually inspecting each of the trajectories. Filled boxes in the “Size” column indicate that the smallest of the resulting trajectory groups contained at least 5% of the sample, using posterior probabilities of group membership to make trajectory group assignments.

Table A1b.

Fit Statistics by Model Specification, Gender, and Type of Work/Family Trajectory (LCGA) [To be made available online]

	Men						Women
	BIC	AIC	LMR	Entropy	Distinct	Size	BIC	AIC	LMR	Entropy	Distinct	Size
Employment status
k = 1	107756.3	107738.0	–	–			246107.9	246089.2	–	–
k = 2	81604.4	81561.8	0.00	0.96			180937.6	180893.9	0.00	0.98
k = 3	70409.8	70343.0	0.00	0.97			158064.6	157996.1	0.00	0.97
k = 4	62086.3	61995.1	0.00	0.98			140174.5	140081.0	0.00	0.97
k = 5	58822.8	58707.3	0.01	0.98			131808.9	131690.5	0.00	0.97
Job quality
k = 1	178653.3	178635.0	–	–			165206.1	165187.6	–	–
k = 2	85794.2	85751.7	0.00	0.99			92735.6	92692.5	0.00	0.98
k = 3	68184.0	68117.2	0.00	0.99			77599.3	77531.5	0.00	0.96
k = 4	56228.5	56137.4	0.00	0.99			63986.1	63893.6	0.00	0.96
k = 5	51518.6	51403.2	0.32	0.98			59105.3	58988.1	0.21	0.95
Health insurance
k = 1	191019.0	191000.7	–	–			281426.0	281407.3	–	–
k = 2	134368.1	134325.6	0.00	0.99			181782.6	181739.0	0.00	0.99
k = 3	106318.6	106251.7	0.00	0.99			152654.7	152586.1	0.00	0.99
k = 4	95968.7	95877.5	0.00	0.99			130451.0	130357.5	0.00	0.98
k = 5	86923.4	86807.9	0.59	0.98			121366.0	121247.6	0.11	0.97
Pension benefits
k = 1	220608.6	220590.3	–	–			220608.6	220590.3	–	–
k = 2	146256.3	146213.8	0.00	0.99			146256.3	146213.8	0.00	0.99
k = 3	119539.0	119472.2	0.00	0.99			119539.0	119472.2	0.00	0.99
k = 4	104179.5	104088.4	0.00	0.99			104179.5	104088.4	0.00	0.99
k = 5	95843.9	95728.4	0.00	0.98			95843.9	95728.4	0.08	0.98
Marital status
k = 1	258773.8	258755.6	–	–			347675.0	347656.3	–	–
k = 2	165121.4	165078.9	0.00	0.99			209433.3	209389.8	0.00	1.00
k = 3	141890.9	141824.1	0.00	0.98			177877.5	177809.1	0.00	0.99
k = 4	128224.4	128133.3	0.02	0.98			162922.9	162829.5	0.38	0.99
k = 5	118075.6	117960.2	0.12	0.98			147968.2	147849.9	0.98	0.99

Note: Models with k = 1, …, 5 trajectories were fit for each work/family charcteristic, separately by gender. The “LMR” column provides p -values from a Lo-Mendell-Rubin likelihood ratio test, comparing a model with k trajectories to a model with k – 1. The null hypothesis for the test is that the two models are equivalent. Filled boxes in the “Distinct” column indicate that trajectories identified by the model were substantively distinct; empty boxes indicate otherwise. We made these determinations by visually inspecting each of the trajectories. Filled boxes in the “Size” column indicate that the smallest of the resulting trajectory groups contained at least 5% of the sample, using posterior probabilities of group membership to make trajectory group assignments.

Table A1c.

Fit Statistics by Model Specification, Gender, and Type of Work/Family Trajectory (GMMs) [To be made available online]

	Men						Women
	BIC	AIC	LMR	Entropy	Distinct	Size	BIC	AIC	LMR	Entropy	Distinct	Size
Employment status
k = 1	71432.4	71408.1	–	–			157517.3	157492.4	–	–
k = 2	54310.2	54261.6	0.00	0.93			123314.5	123264.7	0.00	0.87
k = 3	51631.2	51558.2	0.00	0.93			113104.2	113029.4	0.00	0.82
k = 4	52106.7	52009.4	0.00	0.93			105520.2	105420.5	0.00	0.81
k = 5	47372.7	47251.2	0.02	0.84			100454.1	100329.5	0.00	0.87
Job quality
k = 1	66166.5	66142.2	–	–			73930.2	73905.5	–	–
k = 2	47660.4	47611.8	0.00	0.95			54795.6	54746.3	0.00	0.79
k = 3	41176.9	41104.0	0.00	0.93			48665.7	48591.7	0.00	0.80
k = 4	39352.1	39254.9	0.02	0.91			43760.4	43661.7	0.12	0.84
k = 5	36063.5	35942.1	0.31	0.85			40710.2	40586.9	0.02	0.73
Health insurance
k = 1	111383.2	111358.9	–	–			156353.9	156329.0	–	–
k = 2	78153.4	78104.8	0.00	0.90			116037.2	115987.3	0.00	0.92
k = 3	96222.6	96149.7	0.01	0.98			101109.3	101034.5	0.00	0.87
k = 4	67152.1	67054.8	0.04	0.89			92837.2	92737.6	0.04	0.91
k = 5	64058.9	63937.4	0.08	0.88			88922.1	88797.5	0.10	0.89
Pension benefits
k = 1	130045.5	130021.2	–	–			153865.0	153840.1	–	–
k = 2	86527.6	86478.9	0.00	0.93			111806.3	111756.4	0.03	0.94
k = 3	77035.1	76962.1	0.00	0.91			96832.5	96757.8	0.00	0.76
k = 4	74270.4	74173.2	0.00	0.90			88441.3	88341.6	0.00	0.93
k = 5	67695.3	67573.7	0.00	0.92			87778.9	87654.3	0.64	0.92
Marital status
k = 1	154791.6	154767.3	–	–			217097.3	217072.4	–	–
k = 2	112341.9	112293.3	0.00	0.99			153784.1	153734.3	0.00	0.99
k = 3	99862.0	99789.1	0.00	0.98			130719.2	130644.5	0.00	0.98
k = 4	95692.2	95595.1	0.00	0.93			115571.1	115471.6	0.00	0.99
k = 5	90845.0	90723.6	0.09	0.93			110470.9	110346.4	0.52	0.93

Note: Models with k = 1, …, 5 trajectories were fit for each work/family charcteristic, separately by gender. The “LMR” column provides p -values from a Lo-Mendell-Rubin likelihood ratio test, comparing a model with k trajectories to a model with k – 1. The null hypothesis for the test is that the two models are equivalent. Filled boxes in the “Distinct” column indicate that trajectories identified by the model were substantively distinct; empty boxes indicate otherwise. We made these determinations by visually inspecting each of the trajectories. Filled boxes in the “Size” column indicate that the smallest of the resulting trajectory groups contained at least 5% of the sample, using posterior probabilities of group membership to make trajectory group assignments.

Table A2.

Estimates from Models Predicting Respondents’ Net Worth at Age 65 (Proximate Predictors) [To be made available online]

	Men		Women
Proiximate variable (observed at age 65)	b	(s.e.)	b	(s.e.)
Labor Force Circumstances in 2004
Not employed [Reference Group]	–	–	–	–
Currently Employed	0.39	(0.19)	0.27	(0.21)
Retired: Not at All [Reference Group]	–	–	–	–
Retired: Completely	0.01	(0.16)	−0.45	(0.20)
Retired: Partly	−0.18	(0.19)	−0.32	(0.22)
No pension [Reference Group]	–	–	–	–
Pension Plan on Current/Last Job	0.02	(0.16)	0.20	(0.18)
No Health Insurance [Reference Group]	–	–	–	–
Health Insurance at Current/Last Job	−0.01	(0.15)	−0.25	(0.16)
Occupation Earnings of Current/Last Job	0.02	(0.00)	0.02	(0.00)
Family Circumstances in 2004
Current Spouse: Not Married [Reference Group]	–	–	–	–
Current Spouse: Not Employed	−1.90	(1.19)	−4.77	(2.25)
Current Spouse: Employed	−1.90	(1.19)	−4.79	(2.25)
Current Spouse: Excellent/Good Health	2.60	(1.18)	5.10	2.24
Current Spouse: Fair/Poor/Very Poor Health	1.93	(1.18)	4.84	2.25
Number of Children	−0.04	(0.03)	−0.05	(0.03)
Health in 2004
Self-Assessed Health: Excellent/Very Good [Reference Group]	–	–	–	–
Self-Assessed Health: Good/Fair/Poor	−0.35	(0.10)	−0.81	(0.12)
No Activity Limiting Condition [Reference Group]	–	–	–	–
Activity Limiting Condition	−0.15	0.11	−0.21	0.13

Note: Models also include trajectory-based indicators obtained from LCGA; estimates for these measures are given in Table 6. Bolded coefficients have p -values that are less than or equal to 0.05. All estimates are weighted to account for non-random sample attrition. See text for further details.