Understanding the Benefits of Different Types and Timing of Education for Mental Health: A Sequence Analysis Approach

Abstract

Objectives

Individuals increasingly experience delays or interruptions in schooling; we evaluate the association between these nontraditional education trajectories and mental health.

Methods

Using year-by-year education data for 7,501 National Longitudinal Survey of Youth 1979 participants, ages 14–48 (262,535 person-years of education data), we applied sequence analysis and a clustering algorithm to identify educational trajectory groups, incorporating both type and timing to credential. Linear regression models, adjusted for early-life confounders, evaluated relationships between educational trajectories and mental health component summary (MCS) scores from the 12-item short form instrument at age 50. We evaluated effect modification by race, gender, and race by gender.

Results

We identified 24 distinct educational trajectories based on highest credential and educational timing. Compared to high school (HS) diplomas, <HS (β = −3.41, 95% CI: −4.74, −2.07) and general educational development credentials predicted poorer MCS (β = −2.07, 95% CI: −3.16, −0.98). The following educational trajectories predicted better MCS: some college immediately after HS (β = 1.52, 95% CI: 0.68, 2.37), Associate degrees after long interruptions (β = 1.73, 95% CI: 0.27, 3.19), and graduate school soon after Bachelor’s completion (β = 1.13, 95% CI: 0.21, 2.06). Compared to White men, Black women especially benefited from educational credentials higher than HS in predicting MCS.

Discussion

Both type and timing of educational credential predicted mental health. Black women’s mental health especially benefited from higher educational credentials.

Higher educational attainment predicts better mental health (Gilman et al., 2002; Lorant et al., 2003; Vable et al., 2016). However, most studies operationalize education using simplified measures based on either final credential (e.g., < high school [HS] vs ≥ HS) or total years of schooling, potentially masking important heterogeneities by type of educational credential and timing of educational attainment. The mental health effects of relatively nontraditional educational trajectories, such as some college (i.e., some postsecondary schooling, but no degree attainment), the general education development (GED) credential, and Associate’s degrees, are unclear (Montez et al., 2017; Zajacova et al., 2020). While the GED credential is broadly perceived as equivalent to an HS diploma, it is associated with poorer mental health outcomes (Zajacova & Montez, 2017).

Although nearly all prior work has evaluated total years of schooling or final credential attained, the timing of educational attainment could also influence mental health in several ways (Walsemann et al., 2012, 2018). Delayed (e.g., graduate HS, then pause some years before attending college) or interrupted (e.g., start college, then pause some years before completing degree) education may not confer mental health gains equivalent to completion of the same credential after continuous schooling. Specifically, delays or interruptions may reduce the financial benefits of credentials (Walsemann et al., 2012). Pursuing education in a nontraditional sequence may also limit opportunities for establishing professional networks and have larger opportunity costs due to foregone wages or costs associated with dependent care. Alternatively, completing or starting an educational credential that was a long-term goal, and/or the increased financial benefits degree completion affords, could benefit mental health.

Structurally minoritized groups, such as Black Americans, women, and those from lower socioeconomic backgrounds, are more likely to experience delays or interruptions in their educational trajectories than advantaged groups (Horn & Carroll, 1996; Milesi, 2010), likely due to disproportionate structural barriers imposed on minoritized groups. However, structurally minoritized groups also tend to benefit more from education than advantaged groups in predicting physical (Vable et al., 2018b, 2018c, 2020) and mental health markers (Ross & Mirowsky, 2006; Vable et al., 2016, 2018a). Resource substitution provides a possible theoretical explanation for the larger benefits of education for structurally minoritized groups (Ross & Mirowsky, 2006). Resource substitution posits that advantaged groups have access to multiple health-promoting resources besides education, such as social networks, economic resources, or political power. Structurally minoritized groups, however, are blocked from accessing health-promoting alternatives to education. Educational attainment is therefore especially relevant for structurally minoritized groups, resulting in larger effect sizes compared to structural advantaged groups, such as high socioeconomic status (SES), cisgender White men.

Several pathways have been proposed to explain the association between education and mental health including health behaviors, access to health care, quality of health care, financial and occupational returns, and social networks (Cohen & Syme, 2013; Rehkopf et al., 2016). Given myriad social inequalities relevant to these paths, each of these mechanisms may have differential effects for men and women with different racial/ethnic identities. Critical race scholars have also argued it is important to shift discourse away from majority groups to minoritized groups, an approach called “centering in the margins,” to more fully understand the experience of those who are structurally minoritized (Ford & Airhihenbuwa, 2010). For these reasons, it is important to specifically evaluate if structurally minoritized groups differentially benefit from nontraditional educational trajectories.

Despite a growing empirical literature on the relationship between education and mental health, there are notable gaps that warrant further investigation, including mental health effects of alternative educational credentials, differential effects based on delays or interruptions in schooling, and differential effects for men and women from different racial/ethnic groups. As nontraditional (e.g., delayed or interrupted) educational trajectories become more common in the United States, especially among structurally marginalized people who also report poorer mental health outcomes, it is important to understand how nontraditional educational trajectories may influence mental health. We address these gaps in the literature by using sequence analysis to characterize educational trajectories from ages 14–48 using data from the National Longitudinal Survey of Youth (NLSY), 1979 cohort, and test for heterogeneous relationships by race, gender, and race by gender.

Method

Sample

Participants in the NLSY 1979 were aged 14–22 at baseline and interviewed annually through 1994 and biennially thereafter. We constructed individual-level educational trajectories for each respondent from ages 14 to 48; therefore, individuals who were followed until at least age 48 were eligible for analyses (N = 7,912). For respondents who were older than 14 at study entry in 1979, we imputed their educational trajectories from age 14 to study entry using data cleaning rules that have been detailed previously (Vable et al., 2020). We excluded 248 respondents with missing data on our primary outcome and 163 respondents missing information on one or more covariates, for a final analytical sample of 7,501 (94.8% of eligible). Those excluded from analysis due to ineligibility or missing data (N = 5,185 from the total NLSY sample of 12,686) completed fewer years of schooling, a lower proportion were female, a higher proportion were White, and a higher proportion were in poverty in 1979 (Supplementary Table 1).

Exposure

We used sequence analysis to collapse the thousands of individual-level educational trajectories NLSY respondents followed from age 14 to 48 into a meaningful, analytically tractable number of similar sequences. Implementation had three steps: (1) creation of education trajectories for each individual; (2) sequence analysis to calculate distance (i.e., dissimilarity) between education trajectories; (3) cluster analysis to group similar trajectories.

Creation of education trajectories for each individual

We used the year-by-year educational reports for each participant from age 14 to 48 to construct educational trajectories. We defined 10 mutually exclusive educational states and classified each year of each participant’s life into one of the following categories (Vable et al., 2020):

1. currently enrolled in HS;

2. not completed HS and not currently enrolled in HS (<HS);

3. HS graduate, not currently enrolled in higher education;

4. completed GED and not currently enrolled in higher education;

5. completed HS or GED and currently enrolled in higher education full-time;

6. completed HS or GED and currently enrolled in higher education part-time;

7. completed some college but not a college degree and not currently enrolled in formal schooling;

8. completed Associate’s degree (AA) and not currently enrolled in higher education;

9. completed Bachelor’s degree (BA) and not currently enrolled in higher education;

10. enrolled in or completed graduate school.

We collapsed educational trajectories after BA completion (started graduate school, master’s degree, professional degree) due to few observations. Data cleaning rules are reported elsewhere (Vable et al., 2020).

Sequence analysis to calculate distance (i.e., dissimilarity) between education trajectories

We evaluated similarity between educational trajectories using an algorithm that assigned an empirical penalty, or cost, to each change necessary to transform one education trajectory into another (Brzinsky-Fay et al., 2006). Costs were assigned as either substitutions (changing one state to another state, e.g., enrolled in HS to HS graduate) or insertions/deletions (inserting or deleting a state to make two sequences equivalent). Each insertion or deletion was assigned a cost of 0.5, and substitution costs were approximately 2 (i.e., approximately four times the cost of an insertion/deletion). The correct ratio of insertion/deletion costs to substitution costs is an area of ongoing scholarship and may vary by research question (Brzinsky-Fay et al., 2006). In sensitivity analyses (not shown), we set the substitution costs to be two times the insertion/deletion costs, and the results were substantively similar. Substitution costs were calculated using the “mean probability distance” so that the most common transitions (e.g., from enrolled in HS to HS graduate) had the lowest costs, and less common transitions had higher costs. We used Halpin’s optimal matching algorithm (Halpin, 2010), which applies a downward duration adjustment for insertion/deletion or substitution costs to reflect that the educational states respondents are in at, for example, ages 15 and 16 are not independent. All substitution costs ranged from 1.987499 to 2; see Supplementary Table 2 for the substitution cost matrix. Costs were summed for each pair of educational trajectories to calculate the “distance” between the trajectories; the result was a symmetric distance matrix for all unique educational trajectories represented in the data. Further explanation of the distance calculation is presented in Supplementary Table 3.

No prior research indicates the optimal calculation of costs, so we compared results using an alternative algorithm, Dynamic Hamming. This algorithm does not consider insertion–deletion changes and considers substitutions less costly when they occur at ages when transitions are common. We consider the Halpin results our primary findings because we are particularly interested in the less common transitions that occur in later adulthood.

Cluster analysis to group similar trajectories

We applied agglomerative clustering to combine sequences that were similar in the sense of having small distances to transform one sequence into another. Agglomerative clustering initially treats all unique sequences as distinct clusters, and then serially combines clusters with the lowest distances. We used a Wards linkage (which maximizes similarities within clusters) to calculate distance between clusters at each step and the Duda–Hart cluster stopping rules (which maximized distinct cluster structure; Stata.com, n.d.). Given the stopping rules indicated similar fit for alternative clustering solutions, we selected solutions that produced more clusters. We then used F tests to collapse trajectories culminating in the same terminal degree if they resulted in similar mental health component summary (MCS) scores (detailed below). This allowed us to avoid the pitfall of creating a large number of very small clusters, leaving no statistical power to detect differences between clusters.

Outcome

NLSY collected data on mental health using the MCS score from the 12-item short form instrument (SF-12) when participants were approximately age 50 (data collection: 2008–2016 when each participant was approximately 50). The SF-12 assesses if, due to emotional problems, the respondent accomplished less than they would like or did work less carefully than usual; assess how much of the time the respondent felt calm and peaceful, had a lot of energy, felt downhearted, and if physical health or emotional problems interfered with social activities. The measure has demonstrated good reliability (≥0.89) and validity (Ware et al., 1996) and detected active and recent depressive disorders (Vilagut et al., 2013). NLSY79 standardized MCS to mean 50 and SD 10, such that a score of 50 corresponds with the U.S. average, and a 1-point difference is 1/10 of an SD (NLSY79 Appendix 19: SF-12 Health Scale Scoring The, 1978).

Effect Modifiers

We evaluated effect modification by race/ethnicity (White, Black, Latino, Other race/missing), gender (male vs female), and by combinations of gender and race/ethnicity (e.g., Black women, Black men, White women, etc.). Due to small numbers and ambiguous interpretation, results for the Other race/missing category are neither presented nor discussed.

Confounders

We adjusted for the following potential confounders: birth year, birth in a Southern state (Glymour et al., 2007), rural residence at age 14, and numerous indicators of childhood SES: mother’s and father’s education (centered at 12 years, plus indicators for unknown or missing values of mother’s and father’s education, N = 436, 6.4% of mothers; N = 1,013, 14.9% of fathers); and the following variables for both parents when the respondent was age 14: worked for pay (yes/no), missing indicators for worked for pay, occupation skilled versus unskilled (dichotomized at 300 using 1970 three-digit census occupation codes), and missing indicator for skilled occupation.

Variables corresponding with risk factors or experiences that occurred after age 14 were not considered potential confounders in our primary analyses because such factors may be influenced by education, and are therefore potential mediators of the relationship between education and MCS (Victora et al., 1997). In secondary analyses, we adjusted for a set of variables that might have been influenced by education but could also have influenced a respondent’s longer-term educational trajectory (i.e., potential time-varying confounders or mediators assessed after respondents were 14 years old, the age the educational trajectories start): (a) indicators for whether health limited the kind of work or amount of work the respondent could do, assessed in 1979, as proxy measures for childhood mental health preceding the educational trajectories; (b) poverty status assessed in 1979, and a missing indicator for poverty status; the armed forces qualification test (AFQT score), administered in 1981, and missing indicator for the AFQT. AFQT was conceptualized as a measure of aptitude or cognitive skills (NLSY, 2019), although critiques of the AFQT argue it actually measures characteristics such as socioeconomic background or English language proficiency (Cascio & Lewis, 2005; Dean et al., 1988; Kanarek, 2015). We additionally evaluate if results are robust to multiple imputation for the 411 individuals missing covariate or outcome data.

Analysis

Primary analyses used linear regression to evaluate the relationship between educational sequence clusters and MCS score at age 50, adjusted for confounders. For any two sequence clusters with the same terminal credential but different timing of degree completion (e.g., GED immediate, GED early, GED late; Supplementary Figure 1), we first tested if MCS differed between the two sequences using the F test to compare multiple point estimates; if we found no evidence that average MCS scores differed between two clusters (p > .2), we collapsed the clusters to improve statistical power to detect meaningful differences. In the interest of reproducible research, we report a large number of analytic findings in the Supplementary Appendix and retain only those essential to the primary results in the main manuscript.

All analyses were weighted to represent the NLSY 1979 sampling frame (U.S. residents aged 14–22 years in 1979), and standard errors were adjusted for the clustered sampling design using the svy command in Stata. Data cleaning and analyses were performed using Stata 15. A second programmer (not the primary analyst) reviewed all data cleaning and analysis code as was recently suggested as best practice (Vable et al., 2021).

Results

Educational Sequences

There were 3,185 unique educational trajectories among the 7,912 NLSY respondents (Figure 1A). The sequence analysis and initial clustering best supported a 24-cluster solution (Supplementary Figure 1). After collapsing clusters with the same terminal degree and similar average MCS score, we retained 10 sequences (Figure 1B): <HS, HS, GED, some college immediate, some college (most trajectories), AA (most trajectories), AA long interruption, BA, graduate school early, and graduate school long delay. The following groups were better represented in the education trajectories resulting in more advanced terminal degrees (Table 1): women, White Americans, those from higher childhood SES backgrounds, those born outside the South, those who lived in more urban areas at age 14, and those with higher AFQT scores.

Figure 1.

(A) Each individual is a row on the y-axis, while each year of life from 14 to age 48 is represented on the x-axis. There were 7,912 individuals followed until age 48 and included in the sequence analysis. We categorized each year of life from age 14 to 48 (35 years) into one of 10 mutually exclusive states. Of a total possible 10³⁵ educational trajectories, there were 3,185 unique educational trajectories. (B) From the 24-group solution of educational trajectories suggested by cluster stopping rules (Supplementary Figure 1), F tests indicated that the following 10 trajectories were most salient for predicting MCS: <HS, HS, GED, some college immediate, some college (most trajectories), AA (most trajectories), AA long interruption, BA, graduate school early, graduate school long delay. Notes: AA = Associate’s degree; BA = Bachelor’s degree; GED = general educational development; HS = high school; MCS = mental health component summary.

Table 1.

Distribution of Covariates Arranged by MCS Educational Trajectories, Weighted for Sample Representation

	<HS	HS	GED	Some college immediate	Some college (all other trajectories)	AA (all other trajectories)	AA long interruption	BA	Graduate school early	Graduate school long delay
Unweighted sample size	N = 750	N = 1,679	N = 828	N = 454	N = 1,046	N = 961	N = 127	N = 951	N = 519	N = 186
Birth year (mean ± linearized SE)	1,960.4 ± 0.1	1,960.3 ± 0.1	1,960.8 ± 0.1	1,960.8 ± 0.1	1,960.2 ± 0.1	1,960.3 ± 0.1	1,960.9 ± 0.2	1,960.6 ± 0.1	1,960.6 ± 0.1	1,960.6 ± 0.2
Female	0.44	0.49	0.47	0.55	0.52	0.59	0.62	0.52	0.51	0.58
White (non-Hispanic)	0.59	0.77	0.61	0.67	0.72	0.72	0.62	0.78	0.84	0.84
Black (non-Hispanic)	0.21	0.14	0.24	0.21	0.16	0.16	0.20	0.10	0.06	0.08
Latino	0.13	0.04	0.09	0.06	0.06	0.06	0.05	0.03	0.03	0.02
Other race/missing	0.06	0.05	0.06	0.05	0.04	0.05	0.09	0.08	0.06	0.05
Southern birth	0.45	0.32	0.40	0.33	0.32	0.29	0.21	0.28	0.27	0.27
Immigrant	0.07	0.02	0.07	0.04	0.05	0.04	0.08	0.03	0.04	0.03
Rural residence at age 14	0.24	0.28	0.21	0.24	0.19	0.24	0.19	0.19	0.16	0.17
Childhood SES measures
Mom’s education (mean ± linearized SE)	8.2 ± 0.2	10.3 ± 0.1	9.2 ± 0.2	11.3 ± 0.2	10.8 ± 0.2	11.3 ± 0.2	11.1 ± 0.4	12.6 ± 0.1	13.5 ± 0.2	12.7 ± 0.3
Mom’s education missing	0.12	0.06	0.09	0.04	0.06	0.04	0.06	0.02	0.02	0.04
Dad’s education (mean ± linearized SE)	6.9 ± 0.3	9.7 ± 0.2	7.9 ± 0.2	11.0 ± 0.3	10.5 ± 0.2	10.9 ± 0.2	10.9 ± 0.5	13.1 ± 0.2	13.9 ± 0.3	13.0 ± 0.4
Dad’s education missing	0.23	0.09	0.19	0.08	0.09	0.10	0.10	0.04	0.04	0.06
Mom worked for pay at age 14	0.47	0.50	0.51	0.55	0.51	0.54	0.60	0.54	0.58	0.51
Mother’s work for pay at age 14 missing	0.05	0.02	0.05	0.01	0.01	0.02	0.01	0.01	0.02	0.02
Mom’s job skilled	0.57	0.57	0.55	0.55	0.59	0.58	0.53	0.65	0.69	0.66
Mom’s job unskilled	0.39	0.40	0.41	0.42	0.38	0.39	0.44	0.33	0.29	0.33
Mom’s job skill missing	0.03	0.03	0.04	0.03	0.03	0.02	0.03	0.02	0.02	0.01
Dad work for pay at age 14	0.67	0.85	0.73	0.81	0.82	0.82	0.77	0.88	0.89	0.81
Dad work for pay at age 14 missing	0.22	0.10	0.21	0.15	0.14	0.12	0.20	0.10	0.08	0.15
Dad’s job unskilled	0.55	0.64	0.58	0.51	0.56	0.47	0.42	0.34	0.26	0.34
Potential mediators (included in secondary analyses)
AFQT (mean ± linearized SE)	16.69	36.54	25.22	48.60	45.19	52.70	43.31	69.14	78.43	69.82
AFQT missing	0.06	0.05	0.07	0.02	0.05	0.04	0.01	0.03	0.02	0.02
Poverty	0.22	0.09	0.18	0.11	0.11	0.09	0.06	0.06	0.04	0.02
Poverty missing	0.23	0.21	0.21	0.16	0.18	0.18	0.16	0.19	0.20	0.24
Outcome
MCS at age 50 (mean ± linearized SE)	50.1 ± 0.6	53.4 ± 0.3	51.4 ± 0.5	54.8 ± 0.4	52.4 ± 0.4	52.6 ± 0.4	54.8 ± 0.7	53.4 ± 0.3	54.0 ± 0.3	52.4 ± 0.7

Notes: Women, White Americans, those from higher childhood SES backgrounds, those born outside the South, those who lived in more urban areas at age 14, and those with higher AFQT scores were better represented in the higher educational trajectories. Distributions weighted to represent the NLSY 1979 sampling frame. AA = Associate’s degree; AFQT = armed forces qualification test; BA = Bachelor’s degree; GED = general educational development; HS = high school; MCS = mental health component summary; NLSY = National Longitudinal Survey of Youth; SES = socioeconomic status.

Overall Results

Generally, trajectories that resulted in more advanced terminal degrees regardless of timing predicted better mental health; the contrast was most marked for individuals with <HS or GED (Figure 2; Supplementary Table 4). Compared to those with an HS diploma only, those with <HS (β = −3.41, 95% CI: −4.74, −2.07) or GEDs (β = −2.70, 95% CI: −3.16, −0.98) reported poorer mental health. Those with GED reported equivalent mental health to those with <HS (p = .10). Compared to those with an HS diploma only, individuals who continued on to some college immediately (β = 1.52, 95% CI: 0.68, 2.37), received an AA after a long interruption (β = 1.73, 95% CI: 0.27, 3.19), or went to graduate school early (β = 1.13, 95% CI: 0.21, 2.06) reported better mental health. The positive outcomes for individuals who completed some college immediately after HS were in contrast to individuals who completed some college after a delay or interruption (some college [most trajectories]), whose average MCS was nonsignificantly worse than HS graduates (β = −0.83, 95% CI: −1.74, 0.87), and worse than those with some college immediate (p < .0005). Timing also mattered for graduate school trajectories, such that those who started graduate school early had better mental health than those who started graduate school after a long delay (p = .05). Those who started an AA but did not complete it for several years (AA long interruption) had better mental health than all other AA trajectories (p = .004; see Supplementary Table 5 for results for all 24 educational trajectories). For context, a 2-unit difference in the MCS score, similar to the difference between the GED and HS credentials, approximates the difference in mental health between Americans aged 40 and 60 years; mental health tends to get poorer with age, although relationships are nonlinear (Utah Health Department, 2001).

Figure 2.

Results adjusted for birth year, birth in a Southern state, rural residence at age 14, childhood socioeconomic status, and weighted to represent the National Longitudinal Survey of Youth 1979 sampling frame. Notes: AA = Associate’s degree; BA = Bachelor’s degree; GED = general educational development; HS = high school.

Effect Modification Results

After evaluating models stratified by both race/ethnicity (Supplementary Table 6) and gender (Supplementary Table 7), we concluded that racial/ethnic differences were modified by gender, so we show only race by gender results (Figure 3; Supplementary Table 8; Supplementary Figure 2). Black women seemed to benefit more from education than White men; interaction point estimates for Black women (i.e., the difference in benefit of the degree for Black women vs White men) were all in the same, positive direction, with the greatest benefits for Black women in the highest attainment trajectories: BA (N = 124; interaction β = 3.58, 95% CI: 1.47, 5.70), graduate school early (N = 49; interaction β = 4.84, 95% CI: 2.25, 7.43), graduate school long delay (N = 25; interaction β = 4.14, 95% CI: −0.09, 8.37; 95% CI contains the null). Similarly, among Black men who started graduate school, the graduate school advantage was larger than among White men: graduate school early (N = 37; interaction β = 2.85, 95% CI: 0.57, 5.14), and graduate school after long delay (N = 12; interaction β = 4.78, 95% CI: 0.97, 8.59); with similar findings for Latino men who started graduate school early (N = 22; interaction β = 2.83, 95% CI: 0.20, 5.45), and Latino women who started graduate school after a long delay (N = 11; interaction β = 3.72, 95% CI: −0.44, 7.88; 95% CI contains the null). See Supplementary Table 9 for the sample sizes for each educational trajectory by race/ethnicity and gender.

Figure 3.

Results adjusted for birth year, birth in a Southern state, rural residence at age 14, childhood socioeconomic status, and weighted to represent the National Longitudinal Survey of Youth 1979 sampling frame. See Supplementary Table 8 for interaction results, and Supplementary Figure 2 for plots of all trajectories stratified by race × sex. Notes: AA = Associate’s degree; BA = Bachelor’s degree; HS = high school.

Robustness Checks

Educational trajectory clusters indicated by the Dynamic Hamming cost algorithm returned similar trajectories (Supplementary Figure 3), point estimates were in the same direction, and our main conclusions were unchanged (Supplementary Table 10). Results also changed little when additionally adjusting for a childhood mental health proxies (Supplementary Table 11), for poverty and AFQT score (Supplementary Table 12), or when performing imputations for individuals missing covariate or outcome data (Supplementary Table 13).

Discussion

Using a novel application of sequence analysis, we evaluated how educational trajectories across the life course, incorporating the type of education credential and the timing of educational attainment, predicted mental health at age 50, and if there were differential relationships by race/ethnicity and gender. Our findings indicate that higher educational attainment predicts better mental health; however, we also observed meaningful differences by both type of education credential and timing of educational attainment. Specifically, we found (a) terminal GED was associated with poorer mental health than HS diploma; (b) the relationship between educational timing and mental health differed by educational credential, with earlier attainment generally predicting better mental health; and (c) Black women benefited more from education than White men in predicting mental health, with most consistent differential benefits for Black women with higher educational trajectories.

Our results add to a growing literature suggesting health researchers should refrain from conflating GEDs and HS diplomas (Liu et al., 2013; Zajacova, 2012; Zajacova & Everett, 2014; Zajacova & Montez, 2017). Rather, we found that participants with terminal GEDs reported similar mental health to those who have less than an HS education. The nonequivalence between GEDs and HS diplomas in predicting mental health could arise via several mechanisms: GEDs may be less valued in the labor market than HS diplomas (Heckman & Rubinstein, 2001), which, in turn, has implications for economic returns to education, health insurance status, social networks, and experiences of stigma (Paul & Moser, 2009). Our results suggest collapsing GEDs and HS diplomas masks meaningful heterogeneities; future research should disaggregate GED and HS diploma credentials when collecting, analyzing, and reporting health by education unless formal statistical tests indicate it is appropriate to collapse. Our findings also support policy and practice efforts to retain students in HS until degree completion.

We also found that education timing mattered for mental health; however, whether earlier versus later matriculation was more beneficial differed by credential. For some college and graduate trajectories, immediate or early matriculation predicted better mental health outcomes, compared to later matriculation. Conversely, attaining an AA after a long interruption benefitted mental health compared to HS only. Prior work has suggested that attaining higher education at later ages benefits mental health (Walsemann et al., 2012), but we know of no work finding that education attained at later ages may be detrimental for health. These differences could reflect variation in workforce opportunities at different times (e.g., early exit from college to pursue promising employment opportunities vs later reentry due to loss of employment or economic recessions), field of study pursued or skills gained at different times (e.g., later reentry for AA credential to advance in existing workplace opportunities vs earlier AA in more general field of study), as well as how these skills are valued by society and in the labor market (Paul & Moser, 2009). Future research, possibly using qualitative approaches, is warranted to explore potential mechanisms underlying these observed associations and how they vary across credentials.

When examining differential effects by race and gender, a particularly salient finding was that Black women consistently benefitted more from education than White men, especially in the higher educational trajectories. That is, Black women attaining a BA or greater benefited more from their educational attainment in predicting mental health than White men in the same trajectory. Our findings are consistent with a majority of studies that note protective effects of education for Black women in particular, using varying operationalizations of education and mental health (Assari, 2017, 2018; Roxburgh, 2009; Vable et al., 2018a), and in contrast to a smaller literature which finds no association between education and mental health, particularly major depressive disorder, among Black women (Hudson et al., 2012; Williams et al., 2007).

Literature on structural racism as a determinant of inequitable access to health-promoting resources suggests structurally advantaged groups are systematically supported in accessing health-promoting alternatives to education (e.g., economic and political power) from which marginalized groups are excluded (Jones, 2000; Williams & Mohammed, 2013). Therefore, accessing higher education could be particularly important for mental health among structurally marginalized people, such as Black women, compared to White men (Ross & Mirowsky, 2006; Vable et al., 2018a). Alternatively, Black women may be less likely to self-report adverse mental health, even when it is present (Williams et al., 2007; Woods-Giscombé, 2010). While our findings for Black women’s mental health were consistent with the broader literature, other work suggests adverse effects of higher education on Black women’s physical health (Geronimus, 1992). These seemingly conflicting results could be explained by informal social support systems Black women establish while navigating higher education (Baker, 2015) in order to buffer some of the adverse mental health effects of training in environments in which racism is operating. While these networks may confer mental health benefits for Black women, the persistence of structural and institutional racism has adverse implications for physical health (Baker, 2015; Leath & Chavous, 2018).

Our findings have implications for policy, practice, and future research. First, given that structurally marginalized people face greater barriers to advanced degrees, programs and policies to increase access to higher education could be powerful mechanisms to reduce racial inequities in mental health (e.g., Affirmative Action [Thibaud 2019; Venkataramani et al., 2019], Pell Grants [Denning et al., 2018], the GI Bill [Vable et al., 2016, 2018b, 2018c]). However, we need to ensure that such programs are actually successful in increasing access to higher education for structurally marginalized people (Herbold, 1994; Turner & Bound, 2013). Further, while removing structural barriers to accessing advanced degree trajectories is necessary, it is also insufficient for effectively addressing the mental health implications of our study. Rather, such interventions would need to be coupled with those that seek to disrupt how racism operates (Bailey et al., 2017; Leath & Chavous, 2018). Our findings indicate it is important to consider how the institution of education may differentially affect mental health across axes of structural marginalization. Evaluating only the overall relationship, or the relationships by race or by gender, without exploring the intersection of race and gender, may result in erroneous conclusions and miss important opportunities for intervention (Ford & Airhihenbuwa, 2010). Thus, this work adds support to the argument that we cannot simply sum the regression coefficients, for example, of Black participants and women to understand the experiences of Black women (Bauer, 2014; Vable et al., 2018a).

There are limitations to these analyses. Self-reports on education may have some misclassification, although participants were interviewed biennially. Our outcome measure does not comprehensively capture all aspects of mental health or clinical diagnoses. Unobserved confounding is a concern in this observational analysis, although we controlled for a comprehensive list of potential confounders available in NLSY79 and performed sensitivity analyses in which we additionally controlled for variables that may be either confounders or mediators. Our findings may be susceptible to reverse causation such that mental health challenges could affect educational trajectories. Relatedly, we are not able to model selection processes that occur during our 35-year exposure window; most notably for these analyses, poor mental health during HS or college could affect educational trajectories. Inability to model life-course selection processes is a limitation of sequence analysis. Our main findings were robust to several additional controls; however, any causal interpretation rests on the strong, untestable assumption of no unmeasured confounding.

Our findings may not generalize to other birth cohorts given secular trends in education systems and policy; we encourage researchers to replicate our analyses in other data sets and other birth cohorts to determine if results are robust to variations in time, place, and population. Most importantly, in some categories the sample is fairly small and replication of our results is important because policy solutions are most wisely undertaken in the context of a cumulative body of findings rather than in response to the results found in any single study.

Our approach also has several strengths. We used prospective data from a national cohort that followed respondents who were ages 14–22 at baseline, thereby avoiding potential recall bias in timing of educational attainment. An important challenge in evaluating the heterogeneous educational trajectories represented in the real world has been choosing a principled strategy to reduce thousands of possible educational trajectories that individuals follow into an analytically tractable number of exposure categories. Leveraging a novel methodologic approach—sequence analysis—facilitated a nuanced assessment of the association between education and mental health, and allowed us to empirically identify meaningful differences. Finally, by exploring the intersection of race and gender we were able to contribute to a growing literature on the ways in which Black women may be uniquely affected.

Our main findings are: (a) there were fewer mental health benefits of terminal GEDs compared to HS diplomas; (b) the importance of timing of degree completion; and (c) the differential benefits of higher education for Black women’s mental health. Further research is warranted to understand the mechanisms that underlie who benefits from education, why timing of degree completion matters, and to determine if these relationships are robust. As the landscape of educational options expands, it is critical for researchers to provide rigorous and nuanced evidence on the long-term implications of diverse educational trajectories to inform intervention and policy.

Funding

This work was supported by the National Institute on Aging (grant number AG056360, PI: I. H. Yen).

Conflict of Interest

None declared.

Author Contributions

A. M. Vable, M. M. Glymour, and I. H. Yen conceptualized the study. A. M. Vable led the analyses. A. M. Vable and C. dP. Duarte prepared the first draft of the manuscript. C. dP. Duarte, S. R. Wannier, and A. M. Chan-Golston provided analytic support including code review, the creation of graphics, and sensitivity analyses. A. K. Cohen and R. K. Ream reviewed manuscript drafts and helped contextualize our findings in the wider literature.