2. Existing research on veteran educational attainment

While a substantial number of servicemembers remain in the military until retirement, the majority serve no more than 5 years (Wenger & Ward, Reference Wenger and Ward2022). Thus, a typical servicemember leaves the military at a relatively young age and usually without a 4-year college degree. Education benefits are a type of deferred benefit for servicemembers intended to facilitate the transition back to civilian life; these benefits also serve as a form of compensation to servicemembers as well as a substantial public investment.

Differences in educational access and costs faced by veteran students arise from time- and location-specific variation in post-service education benefits as well as geographic variation in tuition and housing costs. Based on these sources of variation, a sparse but growing literature estimates plausibly causal effects of military educational benefits on educational attainment. In the interest of brevity, we do not focus on this literature specifically in this article, but relevant works include Angrist (Reference Angrist1993), Bound and Turner (Reference Bound and Turner2002), Steele et al. (Reference Steele, Salcedo and Coley2010), Barr (Reference Barr2015), Barr (Reference Barr2019), and Kofoed (Reference Kofoed2020). Broadly, this research finds that the availability of education benefits for veterans increased educational attainment in a dose–response manner.Footnote ¹ A related literature focuses on barriers faced by Black veterans in using educational benefits and other specific factors relevant to Black veteran educational attainment including O'Neill (Reference O'Neill1977), Onkst (Reference Onkst1998), Turner and Bound (Reference Turner and Bound2003), and Katznelson (Reference Katznelson2005).

More closely related to this study, Loughran et al. (Reference Loughran, Martorell, Miller and Klerman2011) tracked the earnings and educational attainment of both servicemembers and military applicants who do not ultimately enlist. By following both groups from a 4-year application cohort (1991–1994) for nearly two decades, the authors were able to document the emergence and subsequent attenuation of a gap in educational attainment among servicemembers over a period of 18 years. Their results show an initial gap in the attainment of a 4-year degree of approximately zero between future service members and applicants at the time of application that grows to approximately 4% gap for veterans 7 years after application; 17 years after joining or having applied the gap had closed to around 2 percentage points. For attainment of a 2-year degree, they show that this gap begins to close roughly 6 years after the initial application and becomes statistically indistinguishable by year 13.

They also demonstrated that servicemembers were more likely than non-enlistees to attain a 2-year degree, but less likely to attain a 4-year degree. We note that the attainment gaps documented in this study are much smaller than those we document with the broader nonveteran population, consistent with their claim to estimate plausibly causal effects of military service using participation in the military application process to address selection effects.

To our knowledge, the nonparametric, cohort-based approach to using synthetic panel data employed in this study has not been used in the literature on educational attainment. We are aware of one recent study that has used this approach to estimate the relationship between educational attainment and health and mortality outcomes (Kaestner et al., Reference Kaestner, Schiman and Ward2020).

3. Data

For this study we use U.S. Census 5% extract data from 2000 and ACS data from 2005 to 2021. The 2000 Census data are a 1 in 20 random sample of the U.S. population. From 2005 onward, the ACS data are a 1% sample of the population. We use harmonized versions of these data from IPUMS (Ruggles et al., Reference Ruggles, Flood, Sobek, Brockman, Cooper, Richards and Schouweiler2023). ACS data from 2000 to 2004 exist, but do not provide geographic information on respondents below the state level. Because our empirical approach relies on using public use microdata area (PUMA) geographic identifiers as a key control, we exclude these years of ACS data from our analysis sample.Footnote ²

These data represent a rich set of demographic and socioeconomic measures at the individual level. However, there are important limitations of these data pertaining to measuring military service. Most importantly, they lack important details about veteran status and history of military service such as branch of service, officer/enlisted status, whether service was on active duty or in the reserves, years of service, or length of time since separation from the military. For this reason, our approach to identifying an appropriate age to begin comparing the educational attainment of veterans and nonveterans uses the modal path of military service, in terms of years of age. Specifically, we choose the initial age for our analysis based on the notion of a typical enlisted service member who transitions to veteran status at approximately age 23.Footnote ³ We document the empirical evidence supporting this approach in online Appendix A.1.

We make two important sample restrictions to obtain our analysis sample. First, we exclude from the sample individuals who failed to complete high school or attain a Test of General Educational Development (GED) certification. This restriction increases comparability between veterans and nonveterans since such individuals are rare among enlistees and are typically admitted in relatively small numbers that are proportional to the annual difficulty of meeting end strength goals (Alvarez, Reference Alvarez2007). This sample restriction drops 11.7% of the nonveteran sample, but only 2.1% of the veteran sample (and nearly 40% of these excluded veterans are from the earliest birth cohort we consider, 1975–1977). Second, we restrict the sample to three large, self-identified racial/ethnic subgroups: non-Hispanic Whites, Blacks (regardless of Hispanic ethnicity), and non-Black Hispanics (which we refer to going forward simply as “Hispanic” for simplicity). These broad subgroup classifications mirror those commonly used in past research on gaps in educational attainment (see, for instance, Cameron & Heckman, Reference Cameron and Heckman2001; Desmond & López-Turley, Reference Desmond and López-Turley2009; McDaniel et al., Reference McDaniel, DiPrete, Buchmann and Shwed2011) and they include 91% of individuals in our analysis sample, yielding large enough subsample sizes for useful statistical inference even when split by gender.

In Table 1, we present the sample sizes of veterans in each of the five grouped birth cohorts we generate.Footnote ⁴ Each cell in this table represents the total number of veteran observations for each cohort by age grouping provided by the full, pooled data sample. We direct the reader's attention to two notable points. The first is that there is significant variation in the size of these cohorts; this variation is related to the survey sample sizes described above and how our veteran cohorts are spread across them. The second, related point is that these sample sizes are sufficiently large – the average cell size is around 9,000 veterans – to generate credible estimates relating to our research questions. On average, around 17–18% of each veteran cohort–age group is female. The smallest demographic group we consider in any of our analyses is Hispanic female veterans and average sample size for this subgroup is 335, with the smallest of these cohort-by-age group samples comprising 150 veterans.

Table 1.

Veteran birth cohort sample sizes at each age grouping

Source: Author calculations from ACS and Census data from IPUMS (Ruggles et al., Reference Ruggles, Flood, Sobek, Brockman, Cooper, Richards and Schouweiler2023) as described in text. Sample sizes for the age 23–25 cell for the 1978–1980 birth cohort and the age 26–28 cell for the 1975–1977 birth cohort are very small due to the primary years of data for these age and birth cohort combinations (the 2000–2004 ACS 1-year samples) lacking valid PUMA identifiers. We include these small cells in the main analysis of all veterans and note that the associated confidence intervals are a good deal larger for these two cells than for any others in the analysis, but we omit them in all subgroup analyses since in these subsamples the standard errors get very large and fail to provide useful inference.

Due to the absence of geographic identifiers in the 2001–2004 ACS data mentioned above, there are two age group-by-birth cohort cells that have particularly small sample sizes. These are individuals aged 26–28 in the 1975–1977 birth cohort and individuals aged 23–25 in the 1978–1980 birth cohort. These two cells are identified only by individuals on the youngest and oldest months of these 3-year age groups who were observed in the adjacent data years (the 2000 census and the 2005 ACS data). These sample sizes (of around 1,000 veterans) are sufficient to generate estimates among both the full sample and the White, non-Hispanic samples split by sex, but we omit them from the visual estimates for smaller subgroup analyses due to the overly large confidence intervals, while still reporting these intervals in the tables included in online Appendix B.Footnote ⁵

In Table 2 we present some basic characteristics of the veteran and nonveteran samples from our data. This table pools all veterans across the five birth cohorts, using observations between the ages of 23 and 34 (years of age for which all five cohorts are represented in the pooled data). Male veterans and nonveterans are broadly similar on dimensions including income, labor force participation, and employment. But other characteristics differ to varying degrees including age (veterans are slightly older, highlighting temporal shifts in the size of the armed forces over recent years related to the operational needs of the wars in Iraq and Afghanistan), presence and number of children, marital status, share non-Hispanic White, and share Hispanic.

Table 2.

Descriptive statistics for pooled observations of individuals between ages 23 and 34

Source: Author calculations from ACS and Census data (Ruggles et al., Reference Ruggles, Flood, Sobek, Brockman, Cooper, Richards and Schouweiler2023) as described in text.

Note: Sample restricted to respondents born between 1978 and 1990 who self-identify as non-Hispanic White, Black, or Hispanic who attained a high school degree (or equivalent credential) or higher, and who are observed in the data at ages 22–34. Estimates use person weights. Employment is conditional on labor force participation. High school degree includes those with a GED (high school diploma and GED are only disaggregated beginning in 2008). Race/ethnicity categories are mutually exclusive and as described in text. Shares may not sum to 100% due to rounding.

We include measures of personal and family income in these comparisons for the sake of demonstrating both similarities and differences among veteran and nonveteran households. Specifically, veterans and nonveterans have virtually identical personal income but, for both male and female veterans, total family income is 7–8% lower than nonveteran households. We conjecture that this is likely due to a persistent component of the well-documented earnings penalty that accrues to military spouses (Burke & Miller, Reference Burke and Miller2016; Meadows et al., Reference Meadows, Griffin and Pollak2015).

More directly relevant to the focus of this study, veterans in this age group are more likely to have completed some college or attained an associate's degree, and are much less likely to have obtained a bachelor's degree or greater.

Most demographic differences between women veterans and nonveterans are smaller in magnitude than for their male counterparts including marital status, number of children, and share non-Hispanic White. However, differences in educational attainment, our outcome of interest, largely mirror the relationships observed among men.

4. Empirical approach

The goal of our analysis is to characterize differences in the accumulation of education between veterans and nonveterans over the life cycle using available, large-scale, cross-sectional survey data. To accomplish this, we use the series of annual random samples described above to form synthetic longitudinal data at the birth cohort level. We then use a synthetic panel data approach to estimate educational attainment at a series of grouped years of age starting at 23–25.Footnote ⁶ We generate estimates to ages 32–34 for all cohorts, which is as far as we can estimate for the youngest cohort, born between 1987 and 1989, and as far as ages 44–46 for the oldest birth cohort, those born between 1975 and 1977.

Using the birth cohort born between 1975 and 1977 to provide an example of how this approach is implemented, we observe a large, randomly sampled group of individuals between 23 and 25 years old in survey years 2000–2002. We then observe a different randomly sampled group of individuals between 26 and 28 years old in survey years 2003–2005, and between 29 and 31 years of age in survey years 2006–2008, and so on. Because each of these samples is large and random (conditional on included sample weights) we can generate estimates of the average educational attainment for this birth cohort for each age group that we consider.

We estimate a series of semiparametric linear probability models separately for each 3-year birth cohort and, when indicated, for each relevant demographic subgroup (e.g., by sex and race/ethnicity). Otherwise, we control for these and other observable characteristics. These models are of the following basic form:

(1) $$\matrix{ {y_{iac} = \alpha + \mathop \sum \limits_{a = 26-28}^{38-40} \beta _{ac}^{nonvet} nonvet_{iac} + \mathop \sum \limits_{a = 23-25}^{38-40} \beta _{ac}^{vet} vet_{iac} + {\boldsymbol X}{\boldsymbol {^\prime}}_{iac}{\bf \Pi }_c + \delta _a + \varepsilon _{iac}.\;} \cr } $$

Equation (1) estimates, for binary educational outcome y for individual i in age group a (e.g., 26–28) born in birth cohort group c (e.g., 1978–1980), a set of nonparametric, age group-specific estimates of educational attainment levels for both veterans and nonveterans. This approach allows each of these estimated paths of the outcome variable for each birth cohort group to take on any shape across grouped years of age without being restricted by functional form assumptions.Footnote ⁷

For models that pool the entire sample we include indicator variables for race/ethnicity (as defined above) and female to control for the differential composition of veterans and nonveterans in each grouped birth cohort along these two dimensions.Footnote ⁸ We also control nonparametrically for marital status and linearly for total family size, number of children less than 5 years old, and number of children 5 years or older.Footnote ⁹ Finally, we interact the female dummy variable with the controls for marital status and the number of children to control for well-documented differences in gender roles with respect to child rearing and differing constraints potentially affecting school-going for men and women depending on marital and parental status (Bianchi, Reference Bianchi2011; Negraia et al., Reference Negraia, March Augustine and Chambers Prickett2018). This specification makes some potentially nontrivial assumptions, including uniformity of this gender/family size relationship across racial/ethnic groups (as a result of not interacting female with these variables) and no important differences in these relationships across cohorts or ages. While conceptually restrictive, testing the inclusion of these dimensions suggested they failed to qualitatively influence the estimates, so we retained this simpler specification.

In all models we include fixed effects for each single birth year (δ _a) to allow for potential differences in age-specific factors affecting attainment within our grouped birth cohorts.Footnote ¹⁰ We weight all regressions using the provided person weights for each of the pooled datasets. As mentioned earlier, we also include small geographical area fixed effects (Census Bureau public use microdata areas or PUMAs) in our analysis, an approach that helps to control for differences in several otherwise difficult to observe or truly unobservable factors including proximity to post-secondary institutions, variation in average community educational attainment, local labor market opportunities and employment conditions, and local cultural and demographic factors that affect propensity to serve in the military (Goldberg et al., Reference Goldberg, Cheng, Huff, Kimko and Saizan2018).

The coefficients $\beta _{ac}^{nonvet} {\rm \;and\;}\beta _{ac}^{vet}$ from these models, which are the estimated educational attainment levels for nonveterans and veterans net of our included controls, are then combined to construct graphical presentations of the results in either levels or gaps (as discussed further below) with appropriate confidence intervals.Footnote ¹¹ The outcomes we consider are attainment of a bachelor's degree or higher, attainment of an associate's degree or higher, and enrollment in school.

While we have taken care to develop an empirical approach that accounts for observable and certain unobservable differences that could render the comparisons in this study less informative, we reiterate that the empirical results in this study are descriptive in nature and we do not propose that these estimates represent a causal effect of military service on educational attainment. Our focus is on presenting what we believe is new and novel evidence on the noteworthy evolution of secular trends in the educational attainment of veterans relative to nonveterans over time.

6. Mechanisms driving increased educational attainment

There are multiple mechanisms that may be driving veterans to increase their educational attainment relative to nonveterans, both over time within-cohort, and across cohorts. First, the share of veterans attending school may be increasing across cohorts, leading to greater attainment through veterans crossing this extensive margin to pursue post-secondary education. Second, veterans may be increasing their persistence of school attendance over time, leading to increased attainment conditional on beginning studies. Finally, veterans may be enrolling in school more intensively, shortening their time to attainment.Footnote ¹⁷

We have one measure in the Census/ACS data, school enrollment at the time respondents are surveyed, that can provide evidence on the relative plausibility of the first two of these three mechanisms in influencing the trends we present. Comparing enrollment levels indicates that this channel appears to be an important factor in explaining increased attainment. For both male and female nonveterans, enrollment at ages 23–25 ranged from around 22% (men) to 26% (women) for the 1975–1977 birth cohort. Over subsequent cohorts, this level increased to between 27% (men) and 31% (women). Enrollment declines by 8–10 percentage points by ages 26–28 and continues to decline through the rest of the ages we assess. Veteran enrollment for the 1975–1977 cohort was similar in level to nonveteran enrollment at ages 23–25 but for all subsequent (younger) cohorts in our analysis, enrollment levels increased such that our youngest cohort (1987–1989) had enrollment levels 10 percentage points (men) to 18 percentage points (women) higher at ages 23–25. In some cases these levels increased slightly by ages 26–28 and in some they were flat or declined slightly and continued to decline and subsequent ages, but for male veterans, the average (positive) gap in enrollment with similar nonveterans between ages 26–28 and 32–34 was 10–15 percentage points across the five birth cohorts and for women the gap across the same years of age was between 13 and 25 percentage points across the five birth cohorts. (We include these figures in our online Supplementary Appendix material.)

In Figure 5, we show these differences in enrollment as gaps for male and female veterans. In both panels of Figure 5 two patterns are readily apparent. The first is that the relative share of veterans enrolled at ages 23–25 (compared to nonveterans) has been growing in more recent cohorts, indicating support for the first mechanism. The second is that the gradient of this enrollment gap has increased in more recent cohorts as well, suggesting that the second mechanism, increased persistence, is also likely playing a role in closing attainment gaps.Footnote ¹⁸

Figure 5.

Veteran enrollment gap by sex. Figures depict outcomes derived from equation (1) as described in text for each indicated subsample. Heteroskedasticity-robust standard errors clustered at the PUMA level are used to calculate 95% confidence intervals, which are shown as capped whiskers for each estimate. Panel a: men and panel b: women.

Source: American Community Survey and Census data from IPUMS (Ruggles et al., Reference Ruggles, Flood, Sobek, Brockman, Cooper, Richards and Schouweiler2023) as described in text.

We do not have an explicit measure to use in exploring how the third factor may be affecting the growth in relative attainment by veterans (increased enrollment intensity among those who do enroll in school leading to faster completion) but using the conceptual implications of recent changes in the generosity of veteran educational benefits, we can provide some evidence on the role of this third mechanism.

The increased generosity of the post-9/11 GI Bill (PGIB) has two primary components: increased tuition benefits and a set of significant nontuition benefits. Overall, nearly half of the total increase in the overall generosity of the PGIB is due to a housing allowance that provides money for room and board for 36 months for students enrolled in school for at least half of full-time enrollment level (Bass, Reference Bass2019). Additionally, the PGIB provides for up to $1,000 per year for books and supplies. This increase in housing and school supplies benefits under the PGIB likely reduces the need to be employed while enrolled in school. As a result, veteran students are less likely to mix schooling and employment. Since the PGIB covers 36 months of schooling (equivalent to 9 months for 4 years), these time limits may further incentivize veteran students to complete coursework and, hence, educational attainment more quickly.

One sufficient measure of the likelihood that this mechanism plays a meaningful role would be a decline in labor force participation among more recent cohorts of veterans who were increasingly likely to be fully eligible for the PGIB. Online Appendix Table A.1.2 shows that there is an increasing probability for our successively younger cohorts to be eligible for the PGIB. A number of recent studies on the effect of the PGIB have used this same exposure-based approach to identify causal effects of the PGIB on enrollment (e.g., Barr, Reference Barr2015). We incorporate this notion in the following model that estimates differences in the labor force participation for young adult veterans and nonveteran students in birth cohorts that are increasingly likely to be eligible for PGIB benefits:

(2) $$\matrix{ {LFP_{iac} = \pi + \mathop \sum \limits_{c = 2}^5 \gamma _c^{nonvet} nonvet_{iac} + \mathop \sum \limits_{c = 1}^5 \gamma _c^{vet} vet_{iac} + {\boldsymbol X}{\boldsymbol {^\prime}}_{iac}{\bf \Pi }_c + \delta _a + \varepsilon _{iac}.} \cr } $$

This model regresses the labor force participation of individuals in our analysis data who are enrolled in school when they are between 23 and 28 years of age on a set of mutually exclusive birth cohort group indicator variables for both veterans and nonveterans (indexed by c). The model controls for the same measures in our main analyses and includes year-fixed effects for common factors that may have affected labor force participation (e.g., the effects of the Great Recession of 2008).

As in our main analyses above, we use the difference in coefficients to construct a measure of veteran/nonveteran labor force participation gaps for each grouped birth cohort. But to provide context on the levels behind these gaps, for the 1975–1977 cohort, veterans enrolled in school (after controlling only for year, year of age, and PUMA fixed effects) had a labor force participation rate of 60% while fewer than 53% of enrolled nonveterans were in the labor force. These differences declined over time due to lower levels of veteran labor force participation at these younger ages such that the gap in labor force participation between these groups was approaching zero for the 1987–1989 cohort. These labor force participation gaps by cohort are presented in Figure 6. For the 1975–1977 cohort, the figure shows a positive labor force participation gap for enrolled veterans relative to nonveterans of around 7.5 percentage points for the three oldest birth cohorts. However, the magnitude of this positive gap declines among more recent birth cohorts such that, for the most recent cohort, the difference is no longer statistically distinguishable from zero at the 95% confidence level. This first big decline, from around 7.5 percentage points to around 5 percentage points, occurred for the 1981–1983 cohort, who would have been among the first to serve as young adults in the period making them prospectively eligible for the PGIB. This evidence is consistent with the third mechanism, increased course taking conditional on enrollment, playing a role in closing attainment gaps that appears likely related to the increasing nontuition benefits of the PGIB.Footnote ¹⁹

Figure 6.

Differences in labor force participation among veteran and nonveterans ages 23–28 over five birth cohorts. Positive estimates indicate that veteran labor force participation is, on average, higher at these ages than nonveterans. Estimates for each grouped birth cohort shown are relative to the level of labor force participation of the 1975–1977 nonveteran birth cohort. Heteroskedasticity-robust standard errors clustered at the PUMA level are used to calculate 95% confidence intervals, which are shown as capped whiskers for each estimate.

Source: American Community Survey and Census data from IPUMS (Ruggles et al., Reference Ruggles, Flood, Sobek, Brockman, Cooper, Richards and Schouweiler2023) as described in text.