Descriptive Estimates

In this section, I examine the impact of WWII veteran status on indicators for any college, having completed four or more years of college, labor force status, and employment status (employed/unemployed, conditional on being in the labor force). I also consider years of college completed and annual earnings from employment. As I mentioned earlier, because females had to be high school graduates to be eligible for WWII service, both my treatment (WWII veterans) and comparison (non-veterans) groups include only those with a high school diploma. Specifically, in the estimates in Panel A of Table 2, non-veterans are female high school graduates born between 1919 and 1925 who do not enlist in the armed forces during the WWII period. I include state fixed effects and controls for marital status, age (measured in quarters), and race in each specification.

Table 2

Estimates for WWII Veteran Status on Female Earnings and Education

Notes: Standard errors, corrected for heteroskedasticity, in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Dollar values are $1980. In Panel A, I restrict the sample to female, high school graduates, who turned 21 between 1940 and 1946. Non-veterans are, therefore, female high school graduates born between 1919 and 1925 who do not enlist in the armed forces during the WWII period. In Panel B, I restrict the sample to female, high school graduates, who turned 21 between 1940 and 1951, but then exclude non-veterans born between 1919 and 1925. In those estimates, therefore, non-veterans are female high school graduates born between 1926 and 1930 who do not enlist in the armed forces during the WWII period. All specifications include state fixed effects and control for marital status and race. In Panel A, I also control for age using quarter of birth indicators. In Panel B, I cannot control similarly for age because veteran status does not vary for individuals born prior to 1926 due to the sample restriction. To capture the effect of age, I instead add a linear birth year-quarter trend. Estimates without including this linear trend are very similar.

Source: 1980 PUMS 5 Percent Census Sample.

Using an OLS-based Linear Probability Model, I find that female veterans (in 1980) were 19 percentage points more likely to have at least some college, 7.8 percentage points more likely to have completed four years of college, and were 3.2 percentage points more likely to be in the labor force relative to non-veteran female high school graduates. Female veterans, however, were not statistically more or less likely to be employed (conditional on being in the labor force). Further, using a Poisson model, I find that female veterans completed 49.8 percent more years of college than non-veteran females, conditional on having a high school diploma. ^{Footnote 20}

The educational attainment of veterans suggests that veterans and non-veterans differed in ways that could affect earnings from employment later in life. Confirming this suspicion, OLS estimates in Column (6) suggest that WWII veterans earned $1,887 more per year compared to non-veterans in 1980. Given annual earnings of $9,518 for non-veterans, my findings imply that being a veteran is associated with a 19.8 percent boost in earnings relative to similar non-veterans. As an Online Appendix item, I show that accounting for selection into the labor force (using a Heckman selection approach with “other household income” as the selection variable) leads to very similar estimates of the greater earnings of female veterans.

Note that I intentionally do not control for education when examining earnings because education is highly correlated with veteran status (see Table 1) and would bias estimates of veterans’ additional earnings downward. Put differently, I am interested in the gross effect of veteran status, whereas controlling for education would provide a net effect. I also purposely do not control for occupation when looking at earnings because veteran status may predict selection into occupations or industries with higher wages. However, I present estimates where I control for education and occupation as an appendix item. In those estimates, while the additional earnings of veterans are still large and statistically significant, including education and occupation controls attenuates the effect of veteran status on earnings. Such a pattern suggests that the additional earnings of veterans come, at least in part, from the subsequent occupational choices of veterans enabled by greater educational attainment. Note that I also provide estimates where I exclude nurses and doctors from my sample as an appendix item. Those estimates are reassuringly similar to the estimates in Table 2 and highlight that my main findings are not driven solely by the enlistment of medical professionals.

To the extent that military service delayed entry into civilian life (including career choices, marriage timing, fertility, and so on) for veteran females, females from the same age cohort (those born 1919 to 1925) are perhaps not the right comparison group. Also, and as I describe earlier, educational attainment among non-veteran was somewhat negatively affected by WWII. In Panel B of Table 2, therefore, I designate non-veterans to be female highschool graduates born between 1926 and 1930. These females were just too young to serve in WWII. The estimates in Panel B confirm that WWII veterans experience large and statistically significant increases in educational attainment and annual earnings, regardless of the comparison group.

$$Edu{c_i} = \alpha + \tau Ag{e_{i,1944}} + {X_i}\Pi + {\varepsilon _i}$$

At issue is whether the additional earnings of veterans are causally related to the education benefits provided by the G.I. Bill. The problem is that being a veteran could have affected earnings in at least two other ways. First, female veterans selected into service and may have had greater (unobserved) productivity than non-veterans, whether due to innate differences or veterans becoming more productive because of their service (by acquiring skills and/or experience). Second, veterans, as a group, may have greater educational attainment than otherwise comparable females even before they enlist. The effect of WWII veteran status on annual earnings in the final column of Panel A of Table 2 simply estimates the size of the veterans’ earning premium and does not attempt to disentangle such competing explanations.

For my estimates to be causal, the additional educational attainment of veterans must be explained by the benefits provided by the G.I. Bill. Moreover, it must also be the case that such educational attainment causes at least some of the veterans’ earnings premium. In the next two subsections, to help establish a causal relationship, I show that the additional educational attainment of veterans is almost entirely explained by veterans attending college after their service that otherwise would not have attended college. I also show that there is no evidence that attending college affected veterans’ earnings differently relative to non-veterans. Then, I attempt to separate the effect of veterans’ experience and innate characteristics from the effect of education by using age at the time of the G.I. Bill’s announcement to instrument for veterans’ later educational attainment.

Selection on Pre-Service Educational Attainment

One threat to a causal interpretation for my findings is that female WWII enlistees might have been more likely to have attended college before their service relative to high school graduates who did not enlist. To see that this is not likely, first note that the Department of Veteran’s Affairs reports that 19.5 percent of eligible females used their benefits to attend college, and the estimates in Column (1) of Table 2 highlight that WWII veterans are 19 percentage points more likely to report having attended at least some college when responding to the 1980 long-form Census. ^{Footnote 21} The alignment between these measures suggests veterans were not especially more likely to have attended college prior to their service relative to other female high school graduates. ^{Footnote 22} If my 1980 census sample is representative, it, therefore, appears that differences in educational attainment between female veterans and non-veterans are explained by those veterans who attend college after serving in WWII.

To further support such a claim, I turn to the publicly available Women’s Army Corps (WAC) Enlistment Records. ^{Footnote 23} The enlistment records refer to females in just one particular branch of the military but provide educational attainment and other background information including age, race, marital status, birthplace, 1940 residence information, and army enlistment date for almost 37,000 female veterans aged between 21 and 26 at the time of enlistment. ^{Footnote 24} Using the WAC enlistment records, I compare educational attainment (by age) at the time of enlistment to educational attainment for female high school graduates (at the same ages) in the 1940 complete count census (Ruggles et al. Reference Ruggles, Flood, Foster, Goeken, Pacas, Schouweiler and Sobek2020). Table 3 presents a summary of educational attainment for female WAC enlistees age 21 to 26 at the time of their enlistment compared to females age 21 to 26 in the 1940 census. The summary statistics in Table 3 illustrate that WAC enlistees were a little more likely to have some college but somewhat less likely to have completed college. This may represent enlistees interrupting their education to join the military, especially when we consider that enlistees were much more likely to be 21 or 22, relative to the age distribution in the 1940 Census. The key takeaway, however, is that enlistees, when I combine those who have some and those who have completed college—were not especially more likely to have a college education prior to service relative to high school graduates in the population.

Table 3

Educational Attainment in 1940 Census and Enlistment Records

Sources: The data in Panel A refer to the 1940 Complete Census restricted to female high school graduates age 21 to 26 in 1940. Panel B contains data from Women’s Army Corps Enlistment Records.

Selection on Post-Service Educational Attainment

While it appears that females were not significantly more likely to have attended college before enlisting, it is possible that veterans would attend college after their service absent any G.I. Bill benefits. For example, those who enlist might also be those who were planning to attend college in the future. For this to be a significant source of bias, there would have to be a large number of female enlistees who were high school graduates and were planning to go to college, and yet (because females had to be 21 to enlist) did not attend college between the ages of 18 and 21. It is perhaps more likely that selection would work in the opposing direction. That is, the overall probability of attending college for females (or, indeed, males) age 21 or older, conditional on not attending college between age 18 and 21, is likely small. Furthermore, among those choosing to enlist (because few could have anticipated any educational benefits at the time they enlisted), the ex-ante probability of subsequent college attendance, absent G.I. benefits, could be lower relative to those who choose not to enlist.

The mid-1944 announcement of the G.I. Bill additionally limits the value of selecting into service primarily to obtain educational benefits. Given WWII G.I. benefits were awarded to those who served between September 1940 and July 1947 (see Stanley Reference Stanley2003, p. 674), only those who were already enlisted at the time of the announcement would have enough time to obtain the quantity of G.I. Bill benefits sufficient to complete a college degree (each month of active duty provided an additional month of G.I. Bill benefits). ^{Footnote 25} It is possible that some delayed their exit from service to maximize their G.I. Bill benefits. Such behavior clouds identification only if there is a correlation between that delay (thereby increasing the quantity of G.I. benefits) and the individual’s pre-enlistment likelihood of attending college after their period of service. In such a case, the causation would be reversed, college attendance would, at least to some degree, “predict” G.I. Bill benefits. Given that the available evidence suggests few female veterans were likely to attend college without G.I. benefits and the fact that enlistment was concentrated prior to the announcement of G.I. benefits, it seems unlikely that such behavior could be a significant source of bias.

At the same time, because I do not observe the amount of benefits available nor individual army enlistment dates in my 1980 Census data, I cannot completely rule it out. Instead, to try to directly address this potential source of bias, I present estimates where I interact WWII veteran status with an indicator for those who turned 21 after June 1944 in Table 4. These later-born individuals could have been aware of the available G.I. benefits prior to being old enough to enlist and could be driving the increased educational attainment of veterans purely via selection effects once the benefits became known. On the other hand, summary statistics in Table 1 illustrate that those who were old enough to enlist in 1943 or earlier (rather than after 1943) were significantly more likely to attend college. The estimates in Table 4 reinforce that pattern. Looking at the interaction term only, female veterans who served in WWII but who were only eligible to serve after 1943 were somewhat less likely to attend college, complete their degree, and had fewer years of college compared to those who could enlist before the G.I. Bill was announced. They also had lower earnings, but the estimate is not statistically significant. The lack of precision perhaps arises because (1) the estimates are limited to only those who are employed, and (2) only a relatively small proportion of the estimation sample consists of female veterans born after 1923. In any case, these estimates suggest that females selecting into service upon learning of the G.I. Bill’s potential benefits are not driving my estimates.

Table 4

Selection after G.I. Bill Announced

Note: In all specifications, non-veterans are female high school graduates born between 1919 and 1925 who do not enlist in the armed forces during the WWII period.

Source: 1980 PUMS 5 Percent Census Sample.

Note that the estimates in Table 4 also suggest that the experience of serving in the military is not driving my findings. That is, while the mechanism would be far from clear, the experience of being in the army could have caused veterans to attend college even absent any financial assistance to do so. If that were true, however, we would not expect differences in subsequent college attendance among those who were eligible to enlist earlier relative to those who could enlist later. The estimates in Table 4 also ease concerns that those who were planning to attend college in the future (for whatever reason) were significantly more likely to enlist across the WWII period. For example, if veterans were generally more likely to attend college after service, economic theory would suggest that the G.I. Bill’s announcement should further increase that proportion, at least at the margin. In turn, then those who turn 21 after 1943 should be especially likely to attend college relative to enlistees in prior years, but they are not.

To further ease concerns that those who took up the offer of G.I. Bill benefits were positively selected, I present estimates that interact veteran status with indicators for educational attainment in Table 5. There I find suggestive evidence of mild negative selection, in the sense that veterans experienced a smaller boost in earnings from education relative to non-veterans. It is possible, therefore, that instrumental variable (IV) estimates will be larger than OLS estimates of the impact of education on earnings when focusing on veterans. Essentially, a valid IV may estimate a larger local average treatment effect because the instrument changes the behavior of veterans only for whom the positive effects of more education might be larger than average. Such a pattern would be consistent with the higher average earnings of veterans relative to non-veterans among those who did not attend any college (see Table 1).

Table 5

Education Interaction Estimates

Notes: Standard errors, corrected for heteroskedasticity, in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Dollar values are $1980. All estimates include controls for age, race, marital status, and state.

Source: 1980 PUMS 5 Percent Census Sample.

IV Estimates

Before turning to an IV approach, I briefly summarize my findings. In Table 2, I show that female WWII veterans’ educational attainment and annual earnings from employment were significantly greater than comparable non-veterans. Combining administrative data along with enlistment patterns and requirements, I then show that female veterans’ additional educational attainment must be due to veterans obtaining an education after their service, that veterans were not especially likely to attend college even if they did not obtain G.I. benefits (Table 4 and related discussion), and that veterans were not likely to benefit disproportionately from education (Table 5 and related discussion).

My final empirical exercise demonstrates that those induced to attend college by the G.I. Bill’s generous benefits are driving most of the overall earnings premium experienced by female veterans, further limiting concerns that service itself or innate differences in productivity are driving the earnings gap among veterans and non-veterans. Specifically, to try to isolate how increased education, caused by the G.I. Bill, affected veterans’ earnings, I use age at the time of the G.I. Bill’s announcement to instrument for increased educational attainment among veterans.

The idea with this approach is that army enlistment patterns, the 1944 announcement of the G.I. Bill, and the fact that longer periods of service granted more G.I. benefits, mean those female veterans born in 1919, and therefore old enough to enlist several years prior to the G.I. Bill’s announcement, could take greater advantage of the G.I. Bill’s unexpected benefits. My data are at least consistent with such a claim; in Table 1, I show that female veterans born earlier were much more likely to attend college compared to later-born veterans. Further validating my instrument, enlistment records (see Table 3) do not suggest that older enlistees were more likely to have a college education at the time of enlistment (relative to younger enlistees and relative to the general population). Using predicted education values, I then estimate how additional education affects veterans’ earnings.

$$Annual\,Earning{s_i} = \theta + \phi \widehat {Edu{c_i}} + {X_i}\Pi + {\mu _i}$$

In practice, I use an indicator for being born after the second quarter of 1923 (identifying those turning 21 after the G.I. Bill was announced in June 1944) to instrument for three different measures of educational attainment for veterans: any college, four or more years of college (equivalent to a Bachelor’s degree or more), and total years of college. ^{Footnote 26} I present two-stage least squares estimates, alongside OLS estimates, in Table 6. All specifications include state fixed effects while controlling for marital status and race. I do not control for year-quarter of birth in these estimates because I use age as an instrument. The OLS estimates show that attending any college increases veteran’s earnings by $2,424, completing at least four years of college increases earnings by $3,597 and that each additional year of college corresponds to a $925 increase in earnings relative to veterans with only a high school education. ^{Footnote 27} Each estimate is significant at the 1 percent level. Because veterans obtained more education, and because more education appears to increase earnings, my OLS estimates suggest it is possible that the G.I. Bill (via increased education) explains most or all of the $1,887 earnings premium for female veterans in 1980.

Table 6

Two-Stage Least Squares Estimates

Notes: Standard errors, corrected for heteroskedasticity, in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Dollar values are $1980. All specifications include state fixed effects and control for marital status and race. I do not control for year-quarter of birth in the OLS or IV estimates because I use age as an instrument. “Age at time of G.I. Bill Announcement” refers to an indicator for being born before versus after the second quarter of 1923, making a female 21, and thus eligible to serve, before versus after the announcement of the G.I. Bill.

Source: 1980 PUMS 5 Percent Census Sample restricted to female veterans born 1919 to 1925.

On the other hand, the OLS estimates combine veterans who already attended college (or would have attended absent any G.I. benefits) with those who attended only because of the G.I. Bill. When using an IV approach, I find evidence that OLS likely underestimates the effect of G.I. Bill-induced education on earnings for female veterans. Specifically, the IV estimates suggest that female veterans induced to attend at least some college by the G.I. Bill have $6,495 greater annual earnings, those who complete four or more years of college experience $8,214 greater annual earnings, and that, each additional year of college corresponds to a $1,350 increase in annual earnings (an 11.6 percent annual return). The effects on earnings are significant at the 5 percent level for two of the three outcomes (falling just short of the 5 percent level when looking at “Any College”), and the first stage F-statistics indicate that the instrument is not weak.

Note again that using age in 1944 as an instrument is conceptually valid because age in 1944 is clearly correlated with educational attainment (see Table 1) and is unlikely to be directly correlated with differences in earnings in 1980, other than through its effect on education. Furthermore, the difference between the OLS and IV estimates is aligned with the literature on the returns to education described by Card (Reference Card2001). ^{Footnote 28} While an 11.6 percent return to an additional year of education seems large, Card (Reference Card, Ashenfelter and Card1999) finds that IV estimates range between 6 percent (Angrist and Krueger Reference Angrist and Krueger1991) and 15.3 percent (Harmon and Walker Reference Harmon and Ian1995) in studies that use data from the 1970s, 1980s, and 1990s. Moreover, Dougherty (Reference Dougherty2005) highlights that females benefit more from a college education. ^{Footnote 29} My two-stage least squares estimates suggest that 72.5 percent of the overall earnings boost for female veterans can be explained by the additional earnings of those who used their G.I. benefits to attend college. This back-of-the-envelope calculation relies on my earlier OLS estimates of the increase in any college (19 percentage points) and college completion (41 percent of those who complete any college) along with my IV estimates of the returns to education for those induced to attend ($6,495 per year) or complete college ($8,214 per year). Using those values, $6,495 × .19 × .59 + $8,214 × .19 × .41 = $1,368, which is 72.5 percent of the $1,887 overall difference between veteran and non-veteran females in 1980.

My IV estimates further limit concerns regarding positive selection into service by females with greater future earnings potential. Given the pattern of estimates, we would not expect veterans to be significantly more productive absent their additional education, as might be the case if IV estimates were smaller than OLS estimates. Moreover, because the estimates in Table 6 are restricted to veterans, they ease concerns that military service itself, by providing experience or on-the-job training, explain all of the veterans’ earnings premium.

Cohort Level Analyses

As I mentioned earlier, if we are concerned that comparisons between veterans and non-veterans are invalid, one alternative approach would be to perform a cohort-level analysis, following Bound and Turner (2002). Specifically, Bound and Turner use birth year-quarter cohort aggregates to develop estimates of the relationship between the share of veterans in a cohort and overall educational outcomes. Their OLS estimates suggest that the share of male veterans in a cohort is positively related to greater male educational attainment. They also present complementary IV estimates where they use an indicator variable for being born prior to the fourth quarter of 1927 (thus turning 18 prior to Victory over Japan Day (V-J Day), 15 August 1945) as an IV to predict the share of a cohort that are veterans. V-J Day is a valid IV because Japan’s surrender marked the end of the war. Bound and Turner show that the share of WWII veterans by cohort declines swiftly for cohorts born after early 1927. Highlighting that V-J Day was important, their IV estimates closely match their OLS estimates.

For Bound and Turner, the primary motivation for their cohort level analyses is that many veterans would have attended college (instead of serving), and non-veteran males in the same cohorts are those who failed physical and/or mental fitness, limiting their value as a comparison group. Put differently, when looking at male veterans, the main empirical issue is that the within-cohort non-veteran comparison group is small and severely negatively selected, requiring an alternative approach. As I explained when discussing Bound and Turner’s findings earlier, such concerns are less of an issue when focusing on females. When studying female veterans, the small share of female veterans in any given birth cohort (no more than 4.54 percent in my data) means that the within-cohort non-veteran comparison group is both large and unlikely to be severely negatively selected. For that reason, the primary concern when looking at the effect of the WWII G.I. Bill on female veterans is positive selection into service, either from having greater education prior to service or from innate ability/productivity differences that lead to greater future earnings even absent further educational attainment. To the extent that female WWII enlistees (1) did not have greater educational attainment than comparable non-veterans prior to enlisting and (2) were not likely to attend college after serving, as I demonstrated earlier, then comparisons between veteran and non-veteran females will provide reasonable estimates of the effects of interest.

That being said, I present estimates in Online Appendix C that mirror Bound and Turner’s cohort level analyses, using both OLS and the same “eligible for service prior to V-J Day” instrument. ^{Footnote 30} However, regardless of my approach to estimation (OLS or IV), I find mostly null effects on cohort level labor force participation, employment, and earnings. Counter-intuitively, the share of a female birth year-quarter cohort that is a WWII veteran is associated with reduced educational attainment at the cohort level, even when focusing only on WWII-eligible cohorts. Specifically, the coefficients in Table C1 refer to the effect on the outcome of interest for a 1 percentage point increase in the female veteran share in the cohort. Looking at females born between 1919 to 1925, my between-cohort OLS estimates show that if around 3 percent of the sample are WWII veterans, then WWII was associated with a 1.5 percentage point reduction in college completion and about a 0.05 reduction in the number of years of college completed at the cohort level.

As I mentioned earlier, however, negative cohort-level effects on education would be consistent with Acemoglu, Autor, and Lyle (Reference Acemoglu, Autor and Lyle2004) and Jaworski (Reference Jaworski2014), who find that the war effort caused many non-veteran females to join the labor force rather than continuing their education. As a result, despite the boost provided by some veterans in eligible cohorts gaining additional education via the G.I. Bill, those cohorts obtained less education overall. In turn, cohort level IV estimates have the “wrong” sign because the correlation between the instrument (being born early enough to turn 21 before V-J Day, regardless of veteran status) and educational outcomes in relevant cohorts is negative.

Any war-related decline in educational attainment among non-veteran females means that comparisons between veterans and non-veterans in the same birth cohorts will tend to overestimate the effect of the G.I. Bill on female veterans’ outcomes. Helpfully, the effects on educational attainment are relatively small at the cohort level. Moreover, Table 1 shows that the effect of the war effort on college attendance for females seems to have mostly dissipated when we look at those born between 1926 and 1930. The estimates in Panel B of Table 2, because they compare outcomes for female WWII veterans to those of females born from 1926 to 1930, therefore help to bound the size of any overestimate. As one example, the difference between the estimates in Panel A and Panel B of Table 2, in terms of years of education, amounts to 0.169 years in 1980. Looking at years of education completed in 1960, Jaworski (Reference Jaworski2014) finds a war-related 0.163 decline in years of education for females at the median level of mobilization. If only non-veteran females experienced a decline in years of education because of WWII, then any comparison of veterans to non-veterans would likely overestimate the effect of veteran status on years of education by 0.163, right in line with the difference in estimates I report in Table 2. However, to the extent that the war effort also disrupted education among some females who served, it is perhaps best to view the estimates in Panel B of Table 2 as representing a lower bound on the true effect, with the estimates in Panel A representing an upper bound. Either way, it remains clear that female veterans experienced significant G.I. Bill-related increases in educational attainment and earnings.