Isolating the Effects of Discrimination

Regional migrants—people who move to a region in which they were not born—differ on important life outcomes as compared to nonmigrants living in the same region. For example, Southern-born African Americans moving north in the early twentieth century were very different from African Americans already living in the North with respect to culture, skills, and socioeconomic origins. Differences in observed outcomes, hence, do not represent (just) the effects of region-based discrimination; they also reflect preexisting differences. How can we isolate the effects of region-based discrimination from the effects of cultural, skill, and socioeconomic factors that might confound the association between migrant status and life outcomes? In other words, if people living in the North or Midwest who are perceived to be of Southern origins do worse with respect to education, occupation, and earnings, how can we tell the degree to which this is due to discrimination on the basis of perceived regional origins?

Our strategy is to compare siblings, identical in terms of region of birth and family cultural and socioeconomic factors, who differ with respect to the degree to which their names signal Southern origins. Proper names are an important indicator of group membership, and can convey a great deal about social, economic, cultural, religious, racial, and regional origins (e.g., Arai and Skogman Thoursie Reference Arai and Skogman Thoursie2009; Biavaschi et al. Reference Biavaschi, Giulietti and Siddique2017; Busse and Seraydarian Reference Busse and Seraydarian1977; Elchardus and Siongers Reference Elchardus and Siongers2011; Laham et al. Reference Laham, Koval and Alter2012; Rubinstein and Brenner Reference Rubinstein and Brenner2014). By comparing individuals who are the same with respect to factors that might confound the association between migrant status and life outcomes but who differ with respect to the “Southern-ness” of their names, we seek to isolate the effects of discrimination against perceived regional outsiders. Among brothers who both lived in the North (or Midwest), if the brother with the more Southern-sounding name typically fared worse, then this is evidence of discrimination based on perceived regional origins. This is true regardless of actual regional origins, although we examine differences in effects across actual region of birth.

Why Might People with Southern-Sounding Names Have Faced Discrimination?

Previous studies have demonstrated that socially or culturally nonnormative names—names that identify a person as an “other” or “outsider”—subject the individuals who hold them to negative life outcomes relative to individuals with socially normative names. This literature has considered the effects of holding names that are nonnormative with respect to race, gender, immigrant status, and perceived income in both historical and contemporary contexts (e.g., Bertrand and Mullainathan Reference Bertrand and Mullainathan2004; Busse and Seraydarian Reference Busse and Seraydarian1977; Carneiro et al. Reference Carneiro, Lee and Reis2020; Cook et al. Reference Cook, Logan and Parman2016; Figlio Reference Figlio2005; Fryer Jr. and Levitt 2004; Goldstein and Stecklov Reference Goldstein and Stecklov2016). In general, because names often indicate cultural or social group membership, holding a name from a social outgroup can lead a person to experience discrimination. For example, Bertrand and Mullainathan (Reference Bertrand and Mullainathan2004) assessed the impact of racialized names on employment outcomes by submitting nearly identical resumes, differing only on the implied race of the applicant’s first name, to various employers. They found that applicants with Black-sounding first names received 50 percent fewer callbacks than those with White-sounding first names. In other words, when all other social and cultural factors were the same, an outgroup name caused individuals to experience worse outcomes compared to their otherwise similar peers.

However, no previous research has used names to identify the effects of being a (perceived) regional outsider in the United States in either historical or contemporary contexts. Our study is thus innovative in that it fills this gap in the current literature and does so in the context of one of the largest migratory movements in American history: the Great Migration. Why might people with Southern-sounding names have faced professional and social discrimination in the North or Midwest?

Given the significant economic, cultural, and political differences between the North and the South in the early twentieth century, it is fair to compare the relationship between the two regions to the relationship between two nations. In this context, we draw on previous research that has considered the effects of holding names associated with international migrants on social and economic outcomes. For example, Biavaschi et al. (Reference Biavaschi, Giulietti and Siddique2017) examined the consequences for new immigrants to America of changing their names. They found that a more “American” name provided economic benefits for immigrants living in New York City, and could even make up for the economic deficits of being an immigrant (Biavaschi et al. Reference Biavaschi, Giulietti and Siddique2017). Similarly, Arai and Skogman (2009) found that immigrants moving to Sweden who changed their surname experienced an increase in economic outcomes in comparison to their nonconforming peers (Arai and Skogman Thoursie Reference Arai and Skogman Thoursie2009; Goldstein and Stecklov Reference Goldstein and Stecklov2016).

With all this in mind, it is reasonable to hypothesize that having a Southern-sounding name may have had detrimental effects on life outcomes among people living in the North or Midwest—regardless of their actual origins. People with more Southern-sounding names may thus have had worse life outcomes than their siblings with less Southern-sounding names. Given the unique racial dynamics of Northern and Midwest areas that received large numbers of in-migrants during the Great Migration, it is also reasonable to suppose that the effects of having a Southern-sounding name may have been more pronounced for African Americans than for Whites. That is, in the hostile context faced by African Americans, discrimination against men with Southern-sounding names may have been even more severe.

Why Having a Southern-Sounding Name Might Not Have Mattered

Our hypothesis—that having a Southern-sounding name had detrimental effects on life outcomes among people living in the North or Midwest—is based on extensive research that has found name-based discrimination on the basis of race, nativity, gender, and other attributes. However, the bulk of this research has been conducted in contemporary settings. In modern America, decisions about things like college admissions and hiring often happen without face to face interactions. In such bureaucratic and impersonal contexts, those with the power to make admissions or hiring decisions often make guesses—consciously or unconsciously—about applicants’ personal attributes and group memberships (e.g., race, regional origin, gender, nationality). Names can be one important source of information for making such assumptions and guesses. In that space, names—and thus assumptions about personal attributes—can lead to biases and discrimination. However, in the early twentieth century, most such decisions were made after in-person meetings. Job seekers applied in person at the factory or farm, for example, and college applicants more often relied on family and personal connections to gain admission. In these contexts, those with the power to make admissions or hiring or other consequential decisions did not have to make guesses about applicants’ regional origin (or other attributes) based on their names. They met applicants in person, heard their accents, saw their manner of dress, knew their families, or observed other cultural cues directly. Consequently, even though extensive prior research conducted in more contemporary contexts implies that people with Southern-sounding names may have fared worse that their siblings with less Southern-sounding names, there is also good reason to doubt that this hypothesis will hold in the context of the early twentieth century.

Linking Records

To match records across the 1920 and 1940 censuses we first defined the universe of potential matches. To make the task computationally tractable, we restricted the population of potential matches to records that displayed identical or similar characteristics on features that should be consistent over time (e.g., year and place of birth). For example, when attempting to find the 1940 Census record for Michael Corcoran—born in 1917 in Massachusetts according to the 1920 Census—we limited the population of potential matches in 1940 to males born in Massachusetts between 1914 and 1920. Because age is reported in the 1920 and 1940 censuses rather than year of birth, and because of reporting inaccuracies, we allow for deviations of ±3 years in birth year across data sources.

To identify the correct record from among all possible records in the 1940 Census—for example, the correct “Michael Corcoran” in 1940 from among all the possible Michael Corcorans—we employed a machine learning algorithm that performed probabilistic record linking techniques. Briefly, we trained a computer algorithm to recognize patterns in a dataset of potential matches that are consistent with a true match. We used a modification of Feigenbaum’s (2016) probit regression approach, which—like other methods of supervised machine learning—required input from training data. The training data represents a subsample of the population that one wishes to link, but where links have been declared by a trained human to ascertain that confirmed links are as accurate as possible. We used these training data to calibrate the linking algorithm and to evaluate how well it performs in declaring matches and avoiding false positives matches. The goal was to train the algorithm to link as many records as possible while keeping the false positive match rate exceedingly low.

To construct the training data, we randomly selected 1,000 age-appropriate boys from the 1920 Census who we then attempted to link by hand to the relevant universe of possible 1940 matches. Here, the universe of potential matches was limited to cases in which name similarity scores (Jaro Reference Jaro1989, Reference Jaro1995; Winkler Reference Winkler1990, Reference Winkler2006) were at least 0.8. We assessed all potential matches by hand using the wealth of digitized historical information and search features available from Ancestry.com. Historical information about parents, siblings, and the focal individual’s places of residence from the time between the 1920 and the 1940 Census allowed us to make confident assessments regarding the validity of potential matches and to minimize the rate of false positive linkages. Using these procedures, and carefully guarding against false positive matches, we were able to manually declare 50.2 percent of the training data sample as uniquely matched across 1920 and 1940.

To calibrate our linking algorithm, we implemented a “train-test-split” procedure using our training data (in which true matches are known). In the first part of the procedure, we split our training data into two equally sized parts. To train the algorithm, we fit a probit regression model on one-half of the sample, and then evaluated its out-of-sample performance on the other half. The model specification we used is similar to the one proposed by Feigenbaum (Reference Feigenbaum2016), but we added additional individual- and household-level covariates to reduce the risk of false positives, as also illustrated by Helgertz et al. 2020. Results from the model inform the algorithm as to which, if any, of the universe of possible matches should be considered a valid link (e.g., which “Michael Corcoran,” if any, is the one we are looking for). The algorithm declared a unique link based on threshold values (1) the greatest required similarity between any 1-to-1 match (technically the predicted probability based on the probit regression estimates) and (2) the required relative difference between the best and second-best possible match. When selecting thresholds values, the precision and recall of the resulting data are fundamental metrics to consider. Recall refers to the proportion of true matches that are declared as matches by the algorithm, whereas precision indicates the proportion of the declared matches that are true matches. If the thresholds are set very conservatively, few but accurate matches will be made, translating to a low recall and a high precision. Conversely, less conservative thresholds will result in high recall but a low precision.

Choosing thresholds that favor a high precision or a high recall is not a straightforward task, which is why we used the Matthew’s Correlation Coefficient (MCC). Described in Equation 1,Footnote ² the MCC is an especially useful measurement for two-class data where the classes are not well balanced (Chicco Reference Chicco2017); MCC provides a single metric of linking performance (ranging from –1 to +1) and is based on both precision and recall. By looping multiple times over a range of realistic threshold values, we were able to choose values for (1) and (2) that optimized the overall performance of the linking algorithm.

(1) $$MCC = {{{TP \times TN - FP \times FN} \over {\sqrt {\left( {TP + FP} \right)\left( {TP + FN} \right)\left( {TN + FP} \right)\left( {TN + FN} \right)}} }}$$

Weighting

Although the rates at which we matched records across data sources are very high by the standards of historical record linkage (see, e.g., Beach et al. Reference Beach, Ferrie, Saavedra and Troesken2016; Ferrie Reference Ferrie1996), a threat to the external validity of our results is that the sample of boys we were able to match may differ systematically from the sample of boys we were unable to match. To address this potential problem, we constructed and employed poststratification weights that account for each boy’s probability of being matched across the 1920 and 1940 censuses.

We take advantage of the fact that we observe several attributes of boys in 1920 regardless of whether we could confidently match them to a record in the 1940 Census. In Table A1 we report descriptive statistics that compare boys who could and could not be matched to their 1940 Census records with respect to race/ethnicity, place of birth, mother’s literacy, father’s occupation, and homeownership. Boys who could be matched to their records in 1940 are more likely to be White, to come from somewhere other than the South, and to come from more advantaged socioeconomic origins.

To construct the weights, we estimated a probit model predicting whether a boy could be matched as a function of his race, age in 1920, state of residence in 1920, mother’s literacy, father’s literacy, father’s occupation, homeownership, and family size. We then take the inverse of the predicted probability of being linked and normalize the distribution of the resulting values to have a mean of one. Conceptually, these weights count boys who were least likely to be matched more heavily in the analyses without changing the overall sample size.

Name-Based Measures

Our key independent variable describes the percentage of boys ages 0 to 10 with each proper name who were born in a Southern state according to the 1920 Census. That is, for every boy age 0 to 10 in 1920 who had the first name “Thomas” we compute the percentage who were born in the South; we then repeat this for every first name observed among boys in this age range in the 1920 Census. As described in the preceding text, we exclude the relatively small number of boys whose first names were observed fewer than 100 times. Many of these boys had names that were abbreviations or spelling or transcription errors (e.g., “Thomqs” or “TG”). Before constructing our measure, we remove punctuation and middle names/initials (e.g., such that “Thomas A.” becomes “Thomas”). We do not account for multiple spellings or nicknames. As a result, “Thomas,” “Thos,” “Tom,” “Thom,” and “Tomas” are all treated as distinct names and appear in the data (so long as they each appear at least 100 times). We also drop names that are 0, 1, or 2 characters long (because most are simply sets of initials, like “TG”).

Table 1 describes the distribution of the resulting measure. Across our 1,022,344 linked boys, the average first name is held by someone born in the South about 33 percent of the time. The average African American boy’s first name was held by someone born in the South about 46 percent of the time. Figure 1 depicts the distribution of the “Southern-ness” of each first name held by the 980,215 White boys (top half of figure) and the 41,081 Black boys (bottom half of figure). It shows that names that are predominantly held by Southerners are more often held by Blacks. Among the names appearing at least 100 times among boys matched between the 1920 and 1940 censuses, those that were most distinctively Southern were Grady (99% Southern born), Odis (95%), and Lonnie (93%). Those that were most distinctively Northern were Carmine (91% Northern born), Regis (87%), and Salvatore (85%).

Table 1. Descriptive statistics, by race

Note: Sample restricted to the 1,022,344 boys who could be linked across the 1920 and 1940 censuses; who were children, stepchildren, grandchildren, or nephews of the household head; who had at least one other age-appropriate brother in the household in 1920; whose first name appeared at least 100 times in 1920 Census; and who lived in one of the five combinations of regions of birth and regions of residence in 1940 considered in our analyses.

Figure 1. Distribution of Southern-ness of first names, by race, boys linked from 1920 to 1940.

Note: Sample restricted to the 1,022,344 boys who could be linked across the 1920 and 1940 censuses; who were children, stepchildren, grandchildren, or nephews of the household head; who had at least one other age-appropriate brother in the household in 1920; whose first name appeared at least 100 times in the 1920 Census; and who lived in one of the five combinations of regions of birth and regions of residence in 1940 considered in our analyses. All analyses are weighted by the inverse of the probability of selection into the linked sample.

Although our focus is on the effects of first names, we also constructed parallel measures for each boys’ last or surname for the purposes of supplementary analyses described in the following text.

Dependent Variables

Our outcome measures from the 1940 Census include educational attainment (measured in years), wage and salary income (in dollars, which is unavailable for many farmers and self-employed people), and occupational prestige (Siegel Reference Siegel1971). As shown in table 1, there is little missing data on educational attainment or occupational prestige (and no missing data on the name-based measures); those missing on wage and salary income were not generally wage and salary workers. As a result, in each analysis we exclude the small percentage of respondents who are missing on the dependent variable.

Analytic Methods

We begin by estimating OLS regression models that express the bivariate relationship between our measures of the Southern-ness of names and each outcome. Of course, these associations are confounded by any variable that is both associated with naming patterns and consequential for our several outcomes (e.g., geography, race/ethnicity, cultural preferences, or socioeconomic status). We thus rely most heavily on sibling fixed-effects models that account for all such attributes that are shared by siblings.

To estimate the effects of the Southern-ness of names more robustly on each outcome, we use a standard fixed-effects specification for within-pair estimation:

(2) $${M_{ij}} = \alpha + \beta {N_{ij}} + {C_j} + {\varepsilon _{ij}},$$

In Equation 2, M _ij is a continuous outcome measure (i.e., education, occupation, or income) for individual i (i = 1, 2) in sibling pair j (j = 1, 2, …, n); N _ij measures attributes of the name of individual i in sibling pair j; C _j is a measure of unobserved contextual characteristics shared by both members of the pair (e.g., geography, race/ethnicity, cultural preferences, or socioeconomic status); and ϵ _ij is a random individual-level error term that is assumed to be uncorrelated with the other explanatory variables in the model.

In such a within-sibling pair model, the unobserved shared components C _j are “controlled away” by modeling differences within pairs of brothers:

(3) $${M_{1j}}-{M_{2j}} = \Delta {M_j} = \beta \Delta {N_j} + \Delta {\varepsilon _j},$$

In Equation 3, the Δs represent differences between variables for the jth sibling pair (i.e., N _1j – N _2j ). This approach eliminates the confounding effects of unobserved contextual characteristics (C _j )—because siblings have these things in common. In our models, we control for birth order that differs between siblings, may be related to naming practices (e.g., names for firstborn sons may be systematically different than for subsequent sons), and can be associated with our outcomes (e.g., if firstborn sons disproportionately inherit property or businesses).Footnote ³

Both our OLS and within-sibling pair models employ the poststratification weights described in the preceding text to account for selectivity with respect to which boys could be linked across the 1920 and 1940 censuses.