Economic and Social Integration of Minorities: The Effect of WWII on Racial Segregation

Transcription

1 Economic and Social Integration of Minorities: The Effect of WWII on Racial Segregation Andreas Ferrara University of Warwick Preliminary draft March 13, 2017 Abstract This paper shows that the significant switch of black workers from low- to semiskilled occupations after WWII in the U.S. is causally related to casualties sustained by each state and county during the war. Difference-in-differences regressions estimate that a one standard deviation increase in the casualty rate among semi-skilled whites increases blacks probability of being employed in semi-skilled manufacturing jobs by 7 p.p., and raises the share of blacks in semi-skilled occupations by 1.6 p.p. Using data from the Negro Political Participation Study of 1961 and instrumenting the change in the share of blacks in semi-skilled jobs from 1940 to 1950 with the casualty rate, IV regressions find a significant and positive impact of the skill upgrade of blacks on their social standing and political participation in the South. Individuals living in counties with a stronger casualty induced skill upgrade of blacks have a higher probability for interracial friendships and interracial friendships formed at work, stronger preferences for integration over separation, as well as increased political participation by and reduced repercussions for political activity for blacks. JEL codes: J15, J24, N42 Keywords: Skill; minorities; social interaction; war. I wish to thank Sascha O. Becker, Clément de Chaisemartin, Price Fishback, Carola Frydman, Christoph König, Luigi Pascali, and seminar participants at the University of Arizona and Warwick for valuable comments and discussions. a.ferrara@warwick.ac.uk 1

2 1 Introduction Many countries, such as the United States, have a sizable share of minority groups. Despite being natives, these minorities are oftentimes neither economically nor socially well integrated with the rest of society. 1 Less fragmented societies, however, have been found to have better outcomes with respect to public goods provision (Alesina and La Ferrara, 2000), innovation in production (Goff et al., 2002), or urban poverty and inequality (Ananat, 2011). One potential policy to improve social integration is to help minority workers into employment where they are more frequently exposed to and cooperate with majority workers. A priori the effect of such a policy is ambiguous. Increased exposure of the two groups in the workplace may lead to more cooperation, reduced stereotyping and thus more social interaction, but can produce the exact opposite effect when the two groups perceive each other as rivals for employment, promotions and wages. This paper studies the question whether increased economic integration of minority workers leads to more social integration from a historic viewpoint using a natural experiment provided by World War II. During the war African Americans in the U.S. experienced a strong and persistent rise in their occupational standing (see figure 1). As blacks moved from low- to semi-skilled occupations, a predominantly white skill group at the time, this significantly increased the exposure of whites to blacks both at work and in the cities to which African Americans relocated from the rural areas during and after the war (Boustan, 2007; Boustan, 2010). While economists have paid much attention to the effects of the war on job market outcomes of women (see Goldin, 1991; Acemoglu et al., 2004; Goldin and Olivetti, 2013), its role in the skill upgrade of blacks and the implications for their social outcomes have not yet been studied. The purpose of this paper is therefore twofold. First, it offers a novel explanation for the unprecedented shift of black workers from low- to semi-skilled occupations at mid-century by arguing that this development was particularly pronounced in areas with higher WWII casualty rates among semi-skilled whites. Second, it exploits the quasiexperimental nature of the casualty induced skill upgrade of blacks to estimate the effect of economic integration on social outcomes of blacks and attitudes towards racial segregation in the South in (2015). 1 For examples in the European context see the reports by Froy and Pine (2011), or OECD/EU 2

3 Figure 1: Skill Composition of Blacks and Whites from 1900 to 2000 (a) Blacks (b) Whites % Laborers Farm laborers Service workers Operatives Craftsmen Sales workers Clerks Professionals Managers % Laborers Farm laborers Service workers Operatives Craftsmen Sales workers Clerks Professionals Managers Note: Graphs are based on the public use microdata files of the Decennial U.S. Censuses by Ruggles et al. (2015). The sample includes employed males aged 16 to 65. Black, dark-gray, and light-gray shades indicate low-, semi-, and high-skilled occupations, respectively. Skill groups are defined according to the 1950 Census Bureau occupational classification scheme. To test whether and how higher casualty rates among semi-skilled whites affected the occupational standing of African Americans, I match 8.3 million Army Enlistment Records with the WWII Honor List of Dead and Missing. The enlistment records contain information on a soldier s state and county of residence, ethnicity, and pre-war occupation. This allows to construct an accurate measure of how many white, semi-skilled workers in each state and county did not return from the war. Combining this casualty rate measure with individual level Census data from 1900 to 2000, difference-in-differences (DiD) estimates show that a one standard deviation increase in the the state-level casualty 1 rate among white semi-skilled workers raises the probability of being employed in a semiskilled manufacturing job for blacks by 7 p.p. The effect persists until the end of the sample period in Using county-level data for the U.S. South, a one standard deviation increase in the casualty rate is associated with a 1.6 p.p. increase in the share of blacks in semi-skilled jobs. This result is robust to different regression specifications, as well as observable and unobservable county characteristics. Given a 12.4% share of blacks in semi-skilled jobs in 1940, an increase of 1.6 p.p. corresponds to an employment increase of blacks by 12.9% in this skill group. 1 3

4 In order to test the effect of economic integration of blacks via the skill upgrade on their social standing and attitudes towards racial segregation in the South, I utilize the Negro Political Participation Study by Matthews and Prothro (1975) from The randomly sampled 1,312 black and white Southerners were asked questions regarding individuals political and personal preferences with respect to racial topics and segregation. Instrumenting the change in the share of blacks in semi-skilled occupations from 1940 to 1950 with the casualty rate of semi-skilled whites, IV regressions show positive effects of the casualty induced skill upgrade on the social standing of blacks. Individuals living in treated counties have a higher probability for interracial friendships, and a higher propensity to preferring integration over segregation. Such friendships are especially formed in the workplace. In addition, affected counties see higher rates of political participation among African Americans and fewer instances in which they are punished by the majority group for their political involvement. The remainder of this paper is organized as follows. Section 2 provides a brief overview of previous work done on black economic history in the 20 th century and its relation to World War II and racial segregation in the South. Section 3 estimates the effect of the WWII casualty rate among semi-skilled whites on the probability of blacks being semi-skilled manufacturing employment and on the share of blacks in semi-skilled jobs in general. It first describes the construction of the casualty rate measure, and then estimates a long DiD regression from 1900 to 2000 using the Census micro data by Ruggles et al. (2015). The last part of the section repeats the DiD analysis at the county level to provide further evidence for the relationship between the skill upgrade of African Americans and the casualty rate. Section 4 uses data from the Negro Political Participation Study to estimate the effect of the casualty induced skill upgrade on outcomes regarding the social standing and political participation of blacks in the South. The final section concludes. 2 WWII, Black Economic Progress, and Interracial Relations Before WWII the economic situation of blacks can be summarized by the often cited assessment of Myrdal (1944): They own little property; even their household goods are mostly inadequate and dilapidated. Their incomes are not only low but irregular. 4

5 They thus live from day to day and have a scant security for the future. (p. 205). This is reflected in figure 1 and puts the influx of blacks into semi-skilled jobs into perspective. The war marks a turning point in black economic history (Collins, 2001), however, the majority of existing work has focused on its implications for the labor supply and occupational choice of women (see Goldin, 1991; Acemoglu et al., 2004; Goldin and Olivetti, 2013; Jaworski, 2014). As many young white men were drawn into the military, this opened up job opportunities for groups who had previously been excluded from certain jobs such as manufacturing occupations. This explanation also makes sense for African Americans but has received little attention. The employment gains of blacks during the war were considerable. One million black workers found civilian jobs during the war, mostly in manufacturing (Flynn, 1993), in addition to the 750,000 African-American soldiers who served during WWII (Wolfbein, 1947). Unlike women, most of whom were displaced from their jobs by returning war veterans, blacks managed to secure their wartime employment gains. Relating to previous work on the economic changes for blacks at mid-century, potential explanations for this non-displacement include the increase in schooling and school quality (Donohue et al., 2002; Aaronson and Mazumder, 2011), the temporary implementation of anti-discrimination laws (Collins, 2001), outmigration of blacks from the South (Boustan, 2010), or veteran benefits via the G.I. Bill and skills learned in the military (Collins, 2000; Turner and Bound, 2003). However, changes in schooling and school quality already occurred before the war and in a more continuous fashion (Smith, 1984), while gains from military service and fair employment laws were generally short-lived. Collins (2000) analyzes the Palmer Survey and rejects the idea that work experience in the armed forces significantly improved black veterans employment chances in the manufacturing sector. Instead he finds that the increased mobility during the war was a reason for the occupational upgrade of blacks. He also does not find evidence for black workers leaving agricultural jobs in order to move into semi-skilled manufacturing employment. However, these results are based on six cities only, none of which is located in the South where the situation might have been different. In a later paper Collins (2001) examines the impact of the first nationwide antidiscrimination law on black employment. The law was enacted under Roosevelt and implemented via the Fair Employment Practice Committee in His careful analysis 5

6 shows a significant effect on black employment in war related industries. Given that the committee was disbanded in 1946, it may not fully explain why and how black workers managed to secure the occupational gains achieved during wartime. Some states enacted their own anti-discrimination laws but it was not until 1964 that a comprehensive legal framework was established across the whole U.S. (see Collins, 2003). A marked improvement in schooling and school quality for blacks has been documented by Donohue et al. (2002). They show that the quality of school inputs converged to white levels between 1930 and 1960 for which philanthropic activism such as the Rosenwald Fund played a major role. Similar conclusions are reached by Aaronson and Mazumder (2011). The G.I. Bill, which subsidized college attendance for veterans, also had modest effects for blacks but only for those who were born outside the Southern states (Turner and Bound, 2003). Margo (1995) states the the occupational upgrade of blacks played an important role in the Great Compression for wages of African American workers but that the reasons for this upgrade are not well understood. Given the topical coverage of the previous literature, there are thus two questions which have received little attention. Firstly, how did the war affect the occupational standing of blacks? To fill this gap in the literature, the hypothesis here is that blacks transition more frequently from low- to semi-skilled employment in areas with higher WWII casualty rates among semi-skilled whites. Unlike the draft rate, this is a potential channel which is potentially unaffected by displacement effects after the war from returning soldiers. Secondly, how and through what channels, if any, did WWII affect the social standing of blacks in the post-war period? The move of blacks from rural to urban areas during and after the war not only meant higher exposure of whites to blacks in the cities (Boustan, 2010), but also at the workplace due to the shift of blacks from low- to semi-skilled work (see figure 1). Contemporary economists such as Wolfbein (1947) noticed already at the time that a, significant shift occurred from the farm to the factory as well as considerable upgrading of Negro workers, many of whom received their first opportunity to perform basic factory operations in a semiskilled or skilled capacity. (p. 663). The effect of increased exposure between the two groups is ambiguous. 6

7 One of the first theories on the improvement of racial relations between blacks and whites via economic cooperation was formulated by Allport (1954). His Intergroup Contact Theory poses that prejudice and conflict can be reduced by increasing contact between two groups if there is no unequal or hierarchical relationship between the two, and the environment of interaction is non-competitive but characterized by a common goal. The timing of the theory coincides with a significant improvement in the economic standing of black workers and the start of the Civil Rights movement. During the war the increased employment of blacks in factories was met with some resistance. For instance, the Philadelphia transit strike in 1944 was only broken up after the Army threatened to review white workers draft exemptions (Collins, 2001). In the Intergroup Contact Theory there may be an initial period of anxiety after first contact between the two groups that must be overcome. Hence the booming postwar economy provided a potentially more non-competitive, less anxious environment for interaction in the sense of Allport (1954). More recent studies on increased intergroup relations find mixed results. Alesina and La Ferrara (2000) model the conditions under which more ethnic diversity has positive effects on social participation. Their empirical findings show that more ethnic fragmentation in U.S. cities is associated with lower levels of social capital. More racially fragmented localities also display lower levels of public goods provision (Alesina et al., 1999). Ethnic diversity has also been found to have positive effects on innovation in production (Goff et al., 2002) and on urban poverty of blacks and inequality (Ananat, 2011). Even though Alesina and La Ferrara (2005) conclude that developed countries deal better with ethnic diversity in terms of cooperation, they also state that even within developed countries there are varying degrees of conflict and cooperation that are related to ethnic diversity. For instance, Colussi et al. (2016) argue that the construction of mosques and the increased salience of Muslims during Ramadan leads to more political extremism in German elections as well as increased numbers of politically motivated crimes against the Muslim minority. On the other hand, Grigorieff et al. (2016) show that expanding the majority group s information set about minority groups (immigrants in their case) positively affects attitudes towards these minorities. It is therefore not clear as to whether increased exposure between two groups at the workplace has positive or negative effects with respect to the social interaction of the two groups. 7

8 3 WWII Casualties and the Semi-Skilled Employment of Blacks The discussion of how WWII casualties among white, semi-skilled workers affected the probability for employment in semi-skilled manufacturing jobs for blacks and the share of blacks in semi-skilled jobs more generally is presented in three parts. First, I discuss the data sources and construction of the casualty rate among semi-skilled whites. Second, using individual level data from the U.S. decennial Census, I show the evolution of the probability of obtaining semi-skilled manufacturing employment for blacks between 1900 and 2000, and estimate the long-run casualty effect in a difference-indifferences regression while controlling for a wide set of controls. Lastly, to move to a more granular spatial level, I provide county level evidence on how the casualty rate affected the share of blacks in semi-skilled jobs using county level aggregates from the 1940 and 1950 Census. This is limited to states in the South as these are the only states for which employment statistics distinguish occupational groups between black and white workers at the county level. The purpose of using the county level data is to show that the rather coarse state level casualty measure is not accidentally picking up some other unobserved factor which increased the share of blacks in manufacturing in the aftermath of the war. Counties in the Southern states are also interesting for the granular analysis given that the majority of African American population is clustered in this geographic region. 3.1 Constructing the Casualty Rate for Semi-Skilled Whites In order to construct a casualty rate for semi-skilled whites, I match 8.3 million soldiers from the WWII Army Enlistment Records with 309,456 entries from the WWII Honor List of Dead and Missing. The enlistment records were digitized by the National Archive and Records Administration (NARA) while the casualty records were digitized specifically for this paper. The enlistment records include a soldier s name, Army serial number, state and county of residence, place and terms of enlistment, the data of enlistment, grade and branch, nativity, year of birth, race, education, and pre-war occupation. For a small subsample the variable holding a soldier s height was filled in with data on their Army General Classification Test (AGCT) score, the predecessor of the modern AFQT (see Aaronson and Mazumder, 2011). 8

9 Figure 2: Number of Drafted and Fallen Soldiers by Month and Year (a) Draft Numbers (b) Casualty Numbers Inductions per Month Casualties Battle of the Bulge D-Day Battle of Anzio Battle of Sicily Operation Torch Guadalcanal Okinawa 1940m1 1941m1 1942m1 1943m1 1944m1 1945m1 1942m1 1943m1 1944m1 1945m1 1946m1 Note: Draft numbers (inductions) include those who enlisted voluntarily prior to when voluntary enlistment was forbidden in Both draft and casualty figures are for the Army and Army Air Force only. Panel (b) shows major battles and operations involving U.S. Army and Army Air Force personnel. The enlistment data cover soldiers drafted the Army and Army Air Force, however, the 8.3 million individuals in the data cover the majority of the 10 million drafted men in WWII. The records include enlisted men as well. Also the casualty records are for the Army and Army Air Force only. Due to the high manpower demands by the armed forces there was little scope to choose a service branch for drafted soldiers (Flynn, 1993). Volunteering was forbidden early on in the war in 1942 which is before the largest battles and casualties sustained, so it would have been difficult to form a prior as to which branch was the least dangerous. This is shown in figure 2. Deferments were1 obtained by fathers with dependents, workers 1 in war-related industries and farmers, or conscientious objectors. Out of 40 million men who had been assessed by their local draft boards only 11,896 men registered as conscientious objectors based on religious reasons (Flynn, 1993). Out of 16 million WWII soldiers some 50,000 deserted as compared to the 200,000 out of 2.5 million Civil War soldiers (Glass, 2013). Overall, there is little evidence that draft evasion and avoidance were a major issue during WWII, especially after Pearl Harbor. Casualty and draft records are matched based on soldiers unique Army serial number which results in a match rate of 74.4%. Matches are imperfect because both data sources include errors given that they have been digitized via Optical Character Recognition (OCR). This is particularly true for the casualty records for which the original scans are 9

10 not of very high quality. Another 3.4% of unique casualties were identified by matching the enlistment records with the American Battle Monuments Commission (ABMC) burials and memorials data. These include the names and serial numbers of buried soldiers serving in WWII. Again, this is not a complete list of casualties as not all soldiers could be buried given that some were never retrieved after the war. The remaining 22.3% of casualties were matched via the probabilistic string matching algorithms provided by Wasi and Flaaen (2015). A one-to-one match was used to link each casualty with a potential enlistment record based on name and serial number stratified by state of residence. Names are matched via a tokenization and serial numbers via a bigram algorithm, and the match with the highest matching score was kept. This results in a final match rate of 94%. From a random sample of 1,000 matches the error rate was 0.6%. The OCR quality of the remaining 6% of casualty observations was too poor in order to clearly identify whether a given match was correct. These cases were dropped as were a small number of soldiers from Alaska and Hawaii which were not independent states at the time. Table 1 provides summary statistics for the enlistment records and some of the key variables, including the indicator for whether a soldier did not return from the war. Blacks faced both the lowest draft and casualty rates. This was due to segregation in the armed forces and the insufficient amount of available facilities to maintain this segregation. The other reason was that service required a minimum standard of education and illiteracy was a main reason for rejection. In 1940, 11.5% of blacks were illiterate as compared to 2.9% of whites (Smith, 1984). Given that the AGCT test was taken after formal schooling was completed, the score is essentially also a function of the amount and quality of education received and not necessarily a signal for lower ability. Enlisted men, most of whom joined between late 1941 and the end of 1942, tended to be younger, single, more able and educated, 2 but faced the same unconditional death probability as drafted men as shown in panel B. Also the comparison of Southerners in panel C shows differences in terms of education, marital status, and age but not in the share of fallen soldiers. 2 An AGCT score of 130 and higher was required to be classified in the top grade on the 5-scale assessment scale of the Army. 10

11 Table 1: Summary Statistics - WWII Enlistment Records Panel A: Black (n = 807,116) White (n = 7,228,570) mean st. dev. min. max. mean st. dev. min. max. Age Education AGCT Married Height (in.) Weight (lbs.) Died Panel B: Enlisted (n = 1,670,352) Drafted (n = 6,622,454) mean st. dev. min. max. mean st. dev. min. max. Age Education AGCT Married Height (in.) Weight (lbs.) Died Panel C: South (n = 2,249,203) Non-South (n = 6,043,984) mean st. dev. min. max. mean st. dev. min. max. Age Education AGCT Married Height (in.) Weight (lbs.) Died Note: Summary statistics for data from drafted soldiers in the Army or Army Air Force between 1940 and AGCT is the Army General Classification Test, an ability test administered during the draft examinations. This measure is only available for a subset of men drafted in The similarities in the minimum values for the AGCT, education levels, and height across groups are due to the minimum requirements imposed by the Army on the draft. The indicator for a soldier s death equals one for those who were killed in combat or who died due to all other reasons such as battle and non-battle injuries, accidents, self-inflicted wounds or diseases. Occupations in the enlistment records were classified by the Army in three-digit groups using the Dictionary of Occupational Titles of This classification scheme includes a division into high-, semi-, and low-skilled occupations according to which semi-skilled workers here are defined (see table A2). Using the Enlistment Records merged with the casualty information, the casualty rate for white semi-skilled workers in each region r {state, county} is then constructed by summing over the I casualties and J draftees in each region and computing the rate as, casualty rate r = I i=1 1(white semi-skilled casualties) ir J j=1 1(white semi-skilled drafted) jr 100 (1) 11

12 Figure 3: Spatial Distribution of the Casualty Rate among Semi-Skilled Whites (a) State Level (b) County Level Note: The semi-skilled casualty rate is defined as the number of fallen soldiers with a semi-skilled pre-war occupation divided by the number of drafted semi-skilled workers times one hundred. State and county boundaries are as of 1940 excluding Alaska and Hawaii. Polygon shades refer to quintiles of the casualty rate distribution with the percentage ranges being reported in the legend. Darker shades show higher quintiles of the casualty rate among semi-skilled whites. Rather than computing the casualty rate over all semi-skilled workers, this measure takes only those into account who were actually missing during the war and thus exposed 1 to the risk of dying. The second motivation for using the number of drafted semi-skilled workers as the denominator rather than the semi-skilled workers of draft eligible age is that i) the draft age was changed at multiple points and ii) workers in war related industries had a higher chance of receiving deferments. Without further knowledge on what exact sub-industries and occupations in the Census received such deferments, this will lead to an inaccurate measure of the number of workers who were actually at risk of being 12

13 killed. Using all semi-skilled workers as denominator also makes the assumption that a 50 year old is a substitute for a 20 year old draftee. The denominator of (1) represents those whites who were missing during the war and therefore needed replacement, whilst the denominator represents those who were replaced but did not return, i.e. those who did not potentially displace their wartime replacement. The spatial distribution of the semi-skilled casualty rate by state is shown in figure 3 at the state- and county-level. More variation is available when considering the casualty rate measure at the county level. The county level variation cannot be exploited in the public use micro samples of the Census given that county of residence is not provided after 1940 due to the 72-year rule. The spatial pattern of the casualty rate distribution is not random and nowhere in the following statistical analysis will such an assumption be made. The implication though is that plain OLS is unlikely to be sufficient for estimating the casualty effect on black semi-skilled employment. 3.2 Difference-in-Differences Estimates Using State-Level Casualty Rates The DiD strategy employed in this section uses the state-level casualty rate together with the microdata files of the decennial U.S. Censuses between 1900 and 2000 which are made available by Ruggles et al. (2015). The estimation sample includes African American male employed manufacturing workers aged 16 to 65 of the non-institutionalized population who are not currently attending school. The focus on manufacturing is due to the large proportion of semi-skilled work available in this sector. According to the 1950 Census, 46.5% of all semi-skilled employment and 30.7% of the total private sector employment was in manufacturing. Individuals with missing occupation or industry information were dropped. Classification of workers and industries follows the 1950 Census Bureau industrial and occupational classification schemes. As in Acemoglu et al. (2004), some of the Census samples such as the % file require the use of sampling weights which are employed here to account for the underrepresentation of individuals living in large households. The concerned samples are also listed in appendix A. The Census data include several variables of interest to be utilized as control variables in the statistical analysis. These are the completed years of education attained, age, skill 13

14 and industry group, as well as veteran, marital, urban residency, and inter-state migration status. The latter is an indicator which equals one if an individual s current residence state is not the same as their birth state. Given that several factors changed for African Americans at mid-century such as the increase in education (Smith, 1984; Donohue et al., 2002; Aaronson and Mazumder, 2011), stronger inter-state mobility, and service in the armed forces that potentially provided training and practical skills (Wolfbein, 1947; Collins, 2000), or benefits from the G.I. Bill for Veterans (Bound and Turner, 2003), this is an attempt to take these changes into account. In order to address further aggregate factors that had an impact on the employment of black workers in skilled and semi-skilled occupations, I supplement the Census data with several other data sources. Fishback and Cullen (2013) document slight employment effects in manufacturing related to WWII spending across U.S. counties. War related spending comes from the 1947 edition of the County and City Data Book published by the U.S. Census Bureau. Information on state specific anti-discrimination laws was taken from Collins (2003). An indicator variable for these regulations equals one for implementing states in the nearest Census year after the law was enacted and is zero otherwise. Another concern is technological change in the agricultural sector. Olmstead and Rhode (2001) document a doubling in the number of tractors between 1940 and 1950, moving from 1.5 to 3 million tractors. With blacks being the main farm labor group affected, the improvement in farm technologies might be a push factor which led to a move from farms to factories. This is particularly important given that Olmstead and Rhode (2001) notice a comparatively stronger spread of the tractor in the Southern states during and after WWII. To address this concern, I use data from the 1940 and 1950 agricultural Census to control for the average number of tractors per farm, the total share of acres in farming, the share of cash- and share-tenants to account for differential skill in agriculture, and the average value of farm machinery per farm. The difference-in-differences regression to be estimated is as follows: Pr(semi-skilled = 1) ist = α s +λ t + β t casualty rate s + X istγ + α s t + η ist (2) t 1940 where the outcome is an indicator for whether a black individual i, in state s, in decade 14

15 t is employed in a semi-skilled manufacturing job. This type of jobs refers to the Census classifications of operatives and craftsmen. In this specification the WWII casualty rate among white semi-skilled workers is interacted with decade fixed effects. To uniquely identify the coefficients of interest, a baseline period must be omitted which here was chosen to be 1940, i.e. the last pre-treatment period. For ease of interpretation the casualty rate is standardized to have unit variance. The treatment coefficients β t thus measure the average percentage point increase in the outcome in each period (relative to 1940) for a one standard deviation increase in the casualty rate. Time-invariant factors leading to cross-state variation in the employment probabilities are captured by the state fixed effects α s while time-varying shocks common to all states are accounted for by the decade fixed effects λ t. In addition, state-specific linear time trends α s t control for systematic long run trends in the employment probability for blacks in semi-skilled manufacturing jobs across states. The vector X ist consists of individual and state-level covariates. These include age, an education score variable, 3 a quartic in potential job market experience, and indicators for marital, veteran, inter-state migration, and urban residence status. The indicator for inter-state migration equals one if an individual does not live in their state of birth. Even though the treatment here need not be random, it still has to be exogenous from the viewpoint of the individual worker. Otherwise the casualty rate coefficient potentially capture additional migration effects. The inter-state migration dummy attempts to account for such possible treatment status manipulation. The state level controls include the interactions of decade fixed effects with the draft rate, and with the log of WWII related spending. The draft rate is the number of soldiers inducted for service over the number of men in the draft-eligible age range of 18 to 49. Also included are state-specific fair employment laws before Time-variant covariates are the share of workers employed in the manufacturing sector, the share of acres used in agricultural production, the average value of machinery per farm, the share of sharetenants and cash-tenants, and the average number of tractors per farm. Differentiating agricultural employment into share- and cash-tenants takes care for differential skilllevels in the agricultural sectors while the average number of tractors per farm accounts for potential push factors via technology driven labor release. 3 A consistent measure of years of schooling is only available from 1940 onwards. In prior Censuses only literacy is reported. 15

16 Figure 4: Share of Blacks in Manufacturing in High- and Low-Casualty Rate States (a) Split by Median of the Casualty Rate (b) Split by Quartiles of the Casualty Rate Share of blacks in manufacturing Share of blacks in manufacturing Below Median Casualties Above Median Casualties 1 st Quartile 2 nd Quartile 3 rd Quartile 4 th Quartile Note: Figures based on the Census files by Ruggles et al. (2015) for workers employed in the manufacturing sector. Sample splits are based on above- and below median casualty rate states in panel (a), and on quartiles of the casualty rate in panel (b). The casualty rate refers to the share of fallen white semi-skilled workers relative to the total number of drafted white semi-skilled workers. The last pre-war sample is The main identifying assumption is that the probability of employment in semi-skilled manufacturing jobs for African Americans evolved in a parallel way over time across states with higher and lower casualty rates during the war. Additionally, no other omitted but significant and time-varying determinant of the outcome must be correlated with the casualty rate and change as a result of the war. Panel (a) of figure 4 shows the average share of blacks in manufacturing for above- and below-median casualty rate states. These move in a parallel fashion in the pre-war period and diverge after the war which lends 1 graphical evidence for the parallel trends assumption. Despite the ease of interpretation and visibility, the median split does not describe the full picture. Panel (b) therefore splits states according to the casualty rate quartiles. No effect is visible for states with the lowest casualty rates, the increase in the post-war share of blacks is monotonically increasing in the treatment intensity. The specification in (2) is interesting for two reasons. First, it presents a formal test for the identifying assumption which is given by the pre-1940 β t coefficients. The casualty rate measure should not significantly affect the outcome before there are any casualties. This would cast doubt on the identification strategy and hint towards pre-existing trends in the share of blacks across high- and low-casualty rate states. Compared to the visual inspection, the regression based test for pre-trends has the advantage of conditioning on covariates as well rather than observing the unconditional evolution of the outcome. Secondly, the coefficients with the post-1940 decade interactions measure i) the casualty 1 16

17 Figure 5: Long-Run Difference-in-Differences Results Casualty Rate Effect βt Note: Coefficient plot for regressions of the semi-skilled manufacturing employment indicator on the WWII casualty rate of semi-skilled white workers interacted with decade fixed effects. Coefficients are in terms of a one s.d. increase in the casualty rate. The estimation sample uses 142,850 individual African American workers employed in manufacturing from the decennial U.S. Census files by Ruggles et al. (2015). The regression includes decade and state fixed effects, and state-specific linear time trends. Individual level controls include age, education, a quartic in potential job market experience, and indicators for skill group, marital, veteran, inter-state migration, black, and urban residence status. State level controls include interactions of decade fixed effects with the draft rate and WWII related military spending per capita, as well as information on state-specific anti-discrimination laws, the total employment share in manufacturing, and the number of tractors per acre. Error bars show 95% confidence intervals. Standard errors are clustered at the state level. rate effect on the outcome and ii) the fading-out time (or persistence, if any) of the treatment effect. Figure 5 shows coefficient plots from regression (2). Coefficients are reported in terms of a one standard deviation increase in the casualty rate. Standard errors are clustered at the state level to account for heteroscedasticity and serial correlation of the residuals. The vertical line marks the last pre-treatment period in Before the war none of the coefficients are statistically significant meaning that there are no systematic differences in the employment probabilities for blacks in states with different casualty rates. After the war occurs, a one standard deviation increase in the casualty rate among semi-skilled whites leads to an average increase of 7 p.p. in the employment probability for blacks 1 in semi-skilled jobs. All coefficients are significant at the 5% level until the end of the sample period. While the effect seems to be continually increasing over time, none of the later coefficients are significantly different from the 1950 effect. 17

18 3.3 Difference-in-Differences Estimates Using County Data for the Southern States Despite the rich set of controls, one concern with the previous estimation strategy is that the rather coarse state level casualty rate might be picking up time-varying factors which are correlated with both the casualty rate and the outcome. The Census files for the Southern states report county level aggregates for occupational groups such as operatives and craftsmen for blacks and whites separately. This makes it possible to take the analysis to a more granular level. The second interesting feature of this data is that it is located in the part of the U.S. where the majority of African Americans are located. Data is available for 17 states and Washington D.C. with a total of 1,387 counties using the state and county boundaries of Figure 6 shows the spatial distribution of the casualty rate among semi-skilled whites in the Southern counties in panel (a). Compared to the state level measure, the county level casualty rate provides much more variation with both a larger maximum and standard deviation, but with a similar mean. Table A3 displays the summary statistics of the casualty rate for each level of aggregation. Comparing Southern and non-southern counties, the distribution functions are similar at the center but differ in the tails. Southern counties have more extreme values while non-southern counties have a higher number of zeroes. The estimated density functions of the casualty rate in the Southern counties compared to all non-southern counties are displayed in panel (b). The DiD specification using the county aggregates is share of blacks ct = α c + d βcasualty rate c d X ctγ + ɛ ct (3) where the outcome is the share of blacks in semi-skilled occupations, i.e. craftsmen and operatives, α c are county fixed effects, d 1950 is an indicator for 1950 which is also interacted with the county casualty rate. The vector of controls includes the interactions between the 1950 indicator and: the draft rate, the average casualty rate in the counties adjacent to county c to control for spatial spillover effects, the log of WWII spending per capita, the share of acres flooded by the Mississippi in 1927 to control for a major 4 States included are Alabama, Arkansas, Delaware, Florida, Georgia, Kentucky, Louisiana, Maryland, Mississippi, North and South Carolina, Oklahoma, Tennessee, Texas, Virginia, and West Virginia. 18

19 Figure 6: County Level Casualty Rates (a) Spatial Distribution (Southern counties) (b) County Casualty Rate Densities Distribution of Casualty Rates (County Level) Southern States Non-Southern States Note: Panel (a) displays the casualty rates among semi-skilled whites in Southern counties. Polygon colors indicate quintiles of the casualty rate distribution with the percentage ranges of each quintile being reported in the legend. Panel (b) shows the estimated casualty rate densities for counties in Southern and non-southern states. migration event for black workers (Hornbeck and Naidu, 2014), as well as the number of slaves in 1860 and the number of lynchings of blacks between 1900 and 1930 per thousand blacks to account for anti-black sentiment. Further time-variant controls include the number of Rosenwald schools per one thousand blacks to account for the impact of improved education (Aaronson and Mazumder, 2011), the share of manufacturing and agricultural employment, the share of black males, the share of women in manufacturing, the value of agricultural output and the number of tractors per acre to control for technological change in agriculture, the share of acres 1 in cotton production, the median values of education and family income, as well as the Republican s vote shares in the 1940 and 1950 Congress elections. 1 Summary statistics for 1940 and 1950 are reported in table A4. Table 2 reports the results from estimation of (3) for various specifications. The addition of controls does not significantly affect the estimated increase in the share of blacks in semi-skilled occupations of p.p. for a one s.d. increase in the casualty rate. Even though the interaction of the covariates with the 1950 indicator reduces the estimated treatment effect, the selection procedure of Belloni et al. (2014) in column 4 discards most of these interactions. Their doubly-robust selection method first regresses the outcome on all covariates, their squares, and cross-term interactions, then regresses the treatment on the same 19

20 Table 2: County Level Difference-in-Differences Results, Outcome: share of blacks in semi-skilled jobs (1) (2) (3) (4) Casualty Rate (0.367) (0.451) (0.527) (0.384) Controls No Yes Yes Yes Controls Time No No Yes Yes LASSO-Selection No No No Yes Observations 2,771 2,544 2,544 2,691 Within R Oster s δ Note: Difference-in-differences regressions of the share of blacks in semi-skilled occupations in each county on the county casualty rate interacted with the 1950 dummy. The estimation sample contains 1,387 counties in Southern states in 1940 and Coefficients are expressed in terms of a one standard deviation increase in the casualty rate. All regressions include county fixed effects and an indicator for Controls include: draft rate d 1950, the average casualty rate of counties adjacent to a given county, the share of manufacturing and agricultural employment, the share of black males, the share of women in manufacturing, the value of agricultural output per acre, number of tractors per thousand acres, median education, median family income, log WWII spending per capita, the share of agriculture acres in cotton production, share of acres flooded by the Mississippi in 1927 d 1950, the number of lynchings of blacks between per thousand blacks, the number of Rosenwald schools per thousand blacks, the vote share of the Republican Party in the 1940 and 1950 Congress elections, and the number of slaves in 1860 d Monetary values are deflated to 2010 dollars. The LASSO-selection model uses the variable selection procedure for testing model stability and improving inference by Belloni et al. (2014). Standard errors are clustered at the county level. Significance levels are denoted by * p < 0.10, ** p < 0.05, *** p < set of variables. Those covariates which are significant predictors in either of the two regressions as judged by LASSO or other high quality selection methods are kept. Lastly, the outcome is regressed on the treatment variable and the selected variables from the previous step. The casualty rate effect from the doubly-robust DiD regression produces a coefficient which is statistically indistinguishable from the original specification. A sensitivity analysis for the coefficient stability of the casualty rate coefficient with respect to the observable is shown in table 3 and a robustness check with respect to the unobservables is shown in table 2 by reporting Oster s (2016) δ statistic. Including the most influential control variables separately or jointly does not significantly affect the casualty rate coefficient. All of the top row coefficients from column 2 to 9 are within half a standard error of the uncontrolled estimate in column 1. Given that the coefficient of interest remains stable in all of these specifications, this is evidence that omitted variable bias is of little concern. However, this line of reasoning makes the assumption that coefficient stability with respect to the observables is informative on the relationship between the casualty rate and the unobservables. In order to test whether unobservables might be a threat to identification, I compute Oster s (2016) δ statistic. Oster (2016) considers a standard 20

21 Table 3: County Difference-in-Differences Sensitivity Analysis Outcome: share of blacks in semi-skilled jobs (1) (2) (3) (4) (5) (6) (7) (8) (9) Casualty rate 1.669*** 1.675*** 1.473*** 1.722*** 1.706*** 1.670*** 1.677*** 1.693*** 1.573*** (0.367) (0.368) (0.371) (0.370) (0.360) (0.367) (0.374) (0.391) (0.383) Draft rate ** *** (0.068) (0.080) Adjacent cas. rate 1.139*** 1.467*** (0.169) (0.175) % agriculture emp *** *** (0.034) (0.031) % black males 0.200** 0.368*** (0.087) (0.102) % female manufact (0.036) (0.031) Tractors per acre (0.458) (0.525) Rosenwald schools (0.592) (0.493) Observations ,766 2,765 2,771 2,771 2,771 2,771 2,725 2,716 R-squared Number of id 1,387 1,387 1,387 1,387 1,387 1,387 1,387 1,376 1,376 Note: Difference-in-differences regressions of the share of blacks in semi-skilled occupations in each county on the county casualty rate interacted with the 1950 dummy. The estimation sample contains up to 1,387 counties in Southern states in 1940 and Coefficients are expressed in terms of a one standard deviation increase in the casualty rate. All regressions include county fixed effects and an indicator for Standard errors are clustered at the county level. Significance levels are denoted by p < 0.10, p < 0.05, p <

22 linear regression model Y = βx +W 1 +W 2 +ɛ, where W 1 = Ψw o is a vector of observable controls and W 2 is an index of unobservables. She then defines the selection relationship as δ Cov(W 1,X) V ar(w 1 ) = Cov(W 2,X) V ar(w 2 ) and solves for δ (the degree to which selection on unobservables is less than or larger than selection on observables) which would be required to produce β = 0. Assuming that W 1 and W 2 can fully explain variation in the casualty rate, i.e. R max = 1 in a regression of the casualty rate on W 1 and W 2, a reasonable threshold for the previous results in table 2 to be considered robust is δ 1, meaning that the selection on unobservables would need to be at least as important as selection observables in order to yield a coefficient of zero for the casualty rate. The δ statistics in table 2 are all larger than this threshold even though it decreases in columns 3 and 4 given that the interactions of the controls with the 1950 dummy mainly add noise to the regression. The LASSO procedure on the other hand produces a sparse model by construction yielding a lower R 2 value that is then used in computing the δ. The assumption in Belloni et al. (2014) is that all relevant variables are available in a larger set of controls but that the econometrician does not know which these variables are. The LASSO selection and the Oster test do not contradict each other but rather operate under different assumptions. As a final check on the parallel trends assumption, figure 7 shows the coefficient plot from a DiD regression including also the year 1930 keeping 1940 as the base year for comparison. Like in the Census data regressions there are no systematic differences in the share of blacks in semi-skilled occupations across counties with differing WWII casualty rates among semi-skilled whites before the war. Also it does not alter the estimated casualty rate effect after the war. Another point to notice is that the estimated casualty rate effect is not exaggerated in the Southern context given the evidence provided in figure 6 and table 1 panel C. Southerners were not more patriotic or enthusiastic with regards to service, hence the similarities in the death probabilities is also reflected in the voluntary enlistment rates (see figure A1). 22