Joni Hersch, Skin Color Discrimination and Immigrant Pay

• Jurisdiction: United States,Federal
• Publication year: 2008
• Citation: Vol. 58 No. 2

SKIN COLOR DISCRIMINATION AND IMMIGRANT PAY

Joni Hersch^*

INTRODUCTION

My article, Profiling the New Immigrant Worker: The Effects of Skin Color and Height,¹presents strong evidence that darker skin color is associated with lower wages for new legal immigrants to the United States. Taking into account education, English language proficiency, occupation in the source country, and family background, as well as Hispanic ethnicity, race, and country of birth, I found that immigrants with the lightest skin color earn, on average, 17% higher wages than comparable immigrants with the darkest skin color. This Essay addresses to what extent this pay disparity associated with skin color is evidence of employment discrimination.

Multiple regression analysis is the standard empirical methodology used to identify whether pay disparities between groups of workers may be due to discrimination. In multiple regression analysis, discrimination is a residual inference drawn after taking into account legitimate productivity-related characteristics. Pay disparities between groups of workers that remain after taking such characteristics into account in the regression analysis are frequently attributed to discrimination. However, such unexplained disparities may instead arise from omitted productivity characteristics. For example, pay disparities on the basis of race that remain after taking into account education may be due to unobserved differences in school quality or to neighborhood effects.

In this Essay, I demonstrate that the negative effect of darker skin color on wages is not due to omitted productivity characteristics. In contrast to characteristics such as race or sex that are dichotomous, skin color varies within race, country of birth, and even families. It is thereby unlikely that an omitted productivity factor such as school quality will be correlated with skin color, after accounting for other productivity characteristics. Empirically, I demonstrate the invariance of the magnitude of the skin color effect as extensive productivity-related characteristics are sequentially taken into account in the regression analysis.

Title VII of the Civil Rights Act of 1964 prohibits discrimination on the basis of color, as well as on the basis of race, religion, sex, and national origin.²My results showing that darker skin color adversely affects earnings among otherwise comparable legal immigrants provide evidence of discrimination on the basis of a protected characteristic. Furthermore, my results demonstrate that color discrimination is a form of discrimination separate from, and in addition to, any discrimination based on race or national origin. Thus, my analysis contributes generally to understanding the ethnic and racial gap in pay observed in the United States.³In addition, my analysis contributes to understanding current widespread opposition to immigrants in the United States. Specifically, the independent effect of skin color on wages of immigrants to the United States demonstrates that one source of discrimination is based on appearance.

Part I provides an overview of the multiple regression approach used to examine whether pay gaps between groups of workers may be due to discrimination. In that Part, I discuss how the possibility of omitted variables bias affects whether unexplained pay gaps can be interpreted as caused by discrimination. To demonstrate the possible importance of omitted variables bias, I provide an example that shows how the non-Hispanic-to-Hispanic pay gap changes as additional variables are included in a regression equation. Part II explains how analysis of gradations of skin color can be used to counter problems of omitted variables bias. Drawing on my 2008 article, I present empirical evidence that darker skin color is associated with lower wages among immigrants and demonstrate that this finding is unlikely to be due to omitted variables bias. Part III concludes with a discussion of the relevance of my research to employment discrimination litigation.

I. MULTIPLE REGRESSION APPROACH TO IDENTIFYING EMPLOYMENT

DISCRIMINATION

Statistical evidence of discrimination in pay between groups of workers can be provided using multiple regression analysis. Multiple regression analysis takes into account the effect on pay of differences in individual worker characteristics and allows us to isolate the contribution of each characteristic.

Discrimination is a residual inference made after all relevant measurable variables have been included in the analysis.⁴A frequent defense in employment discrimination cases is that not all factors have been included and that the omitted variables account for the observed pay gap.⁵However, it is impossible in a multiple regression analysis to control for every possible factor that may affect pay. Indeed, doing so is neither necessary nor desirable. For the most part, omitted variables are included in the random error term that is part of every multiple regression model. Furthermore, some variables are properly excluded from a regression analysis as they may themselves be the outcome of the same discriminatory process under consideration.⁶

Exclusion of relevant variables from a regression equation can result in omitted variables bias under certain conditions. If an omitted variable is correlated with the variable of interest (e.g., sex or race), and if this omitted variable is an important determinant of the outcome (so that if this variable is included in the multiple regression model, its coefficient would be statistically significant), then the coefficient on the variable of interest will reflect both the direct effect of this factor as well as the indirect effect of the omitted variable. Failure to control for this variable results in a biased estimate of the effect of the variable of interest on the outcome.

It is important to recognize that omitted variables bias does not simply arise because a variable is left out of the equation. If the omitted variables are not correlated with the included variables, or if the omitted variables are correlated with the included variables but are not themselves statistically significant determinants of the outcome, there is no statistical problem of bias. The regression may have less explanatory power, but it will not lead to invalid inferences about the magnitude of the coefficients on included variables. In addition, if the effect of the potentially omitted variable is already largely accounted for by other variables in the equation, inclusion of this additional variable will have little effect on the coefficient on the variable of interest or on the explanatory power of the equation. In sum, one might hypothesize that the statistical results are subject to bias, but the magnitude of the bias may be small.

To demonstrate how exclusion of variables correlated with the variable of interest affects the magnitude of any estimated pay disparity between groups of workers, consider the source of pay differences between non-Hispanic and Hispanic workers. It is widely established that education is a major determinant of earnings and that those with higher education levels have considerably higher pay.⁷In addition, Hispanics have lower education levels on average than non-Hispanics.⁸If we are interested in estimating the non- Hispanic-to-Hispanic pay disparity, but do not take into account differences in education, then the estimated pay disparity will reflect not only the effect of being Hispanic rather than non-Hispanic, but also the effect of the omitted years of education. The estimated penalty to being Hispanic will be larger in wage equations that exclude education than in wage equations that include education.

Specifically, I show below how the magnitude of the pay disparity between non-Hispanic and Hispanic workers is affected by inclusion of additional variables, using data from the 2003 Current Population Survey (CPS).⁹This example will then be contrasted with a similar analysis that examines how inclusion of additional information on individuals affects the magnitude of the skin color disparity for immigrants reported in my earlier study.

Hispanic ethnicity is reported in the CPS separately from race. For the purposes of this example, I restrict my analysis to respondents who report their race as white, who are employed but not self-employed,¹⁰who are between 18 and 64 years of age, and who have an hourly wage between $1.50 and $100.

The number of observations is 134,530, with 119,000 observations on white non-Hispanic workers and 15,530 on white Hispanic workers.

Table 1, panel A presents the means or percentages of selected variables for non-Hispanic and Hispanic workers in the sample. The non-Hispanic and corresponding Hispanic values of all means or percentages reported in this table are significantly different from each other at the 1% level. The difference in hourly wage is considerable, with non-Hispanic workers averaging $18.10 per hour and Hispanic workers averaging $12.96 per hour. Also notable are the large differences in education, with non-Hispanic workers averaging 13.99 years of education and Hispanic workers averaging 11.33 years of education.

Table 1, panel B reports the coefficient on the non-Hispanic indicator variable as well as the adjusted R-squared (a measure of the goodness of fit of the regression) and the percent disparity in wages. In all equations, the dependent variable is the log of the hourly wage. I start with a minimal specification and then discuss the consequence on the magnitude of the non- Hispanic-to-Hispanic disparity and on the explanatory power of the equation as additional individual and productivity-related variables are included in the equation.

The results summarized in row 1 are based on a regression controlling only for the non-Hispanic indicator. The coefficient is 0.313, which corresponds to a 36.8% difference in wages between non-Hispanic workers and Hispanic workers.¹¹Note that the...