Wage gaps according to skin color within Mexican labor market

(1)

Master’s Thesis

Wage Gaps According to Skin Color

Within Mexican Labor Market

Rodrigo Velasco Aguilar

Student number: 11376740 Date of final version: July 15, 2017 Master’s programme: Econometrics

Specialisation: Free Track

Supervisor: Prof. dr. Marco J. van der Leij Second reader: Prof. dr. Eleni Aristodemou

FACULTY OFECONOMICS AND BUSINESS

Faculty of Economics and Business

Amsterdam School of Economics

Requirements thesis MSc in Econometrics.

1. The thesis should have the nature of a scientic paper. Consequently the thesis is divided up into a number of sections and contains references. An outline can be something like (this is an example for an empirical thesis, for a theoretical thesis have a look at a relevant paper from the literature):

(a) Front page (requirements see below)

(b) Statement of originality (compulsary, separate page) (c) Introduction (d) Theoretical background (e) Model (f) Data (g) Empirical Analysis (h) Conclusions

(i) References (compulsary)

If preferred you can change the number and order of the sections (but the order you use should be logical) and the heading of the sections. You have a free choice how to list your references but be consistent. References in the text should contain the names of the authors and the year of publication. E.g. Heckman and McFadden (2013). In the case of three or more authors: list all names and year of publication in case of the rst reference and use the rst name and et al and year of publication for the other references. Provide page numbers.

2. As a guideline, the thesis usually contains 25-40 pages using a normal page format. All that actually matters is that your supervisor agrees with your thesis.

3. The front page should contain:

(a) The logo of the UvA, a reference to the Amsterdam School of Economics and the Faculty as in the heading of this document. This combination is provided on Blackboard (in MSc Econometrics Theses & Presentations).

(b) The title of the thesis

(c) Your name and student number (d) Date of submission nal version

(e) MSc in Econometrics

(f) Your track of the MSc in Econometrics 1

(2)

i

Statement of Originality

This document is written by Rodrigo VELASCOAGUILAR, who declares to take

full responsibility for the contents of this document. I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

ii

“If you’re white, you’re right, If you’re yellow, you’re mellow, If you’re brown, stick around, If you’re black, get back.”

(4)

iii

Acknowledgements

I want to thank my family for supporting me; on every step I walk further and with any challenge that comes my way, I am always counting on them to cheer me up despite being thousands of kilometers away. Thanks to my parents and grand-parents, the first ones that believe in me and in the goals I can reach. Finally, all my gratitude to the one that is not among us anymore, but that would be more than happy with this achievement.

Thanks to the University of Amsterdam and to the Orange Tulip Scholarship, without it, I am sure I will not be writing this lines.

To all the friends I made during the Master, both from mine and other pro-grams. It has been amazing to share this moments with you, some of them where we thought we would not make it. Luckily, with a simple look to any side, we found anyone willing to give us a hand in those tough times.

Special mention to Tim. I cannot thank enough the time you spent reviewing this work and the advice you gave me to improve it. You also witnessed the process to do it and of course, deserve my recognition and gratitude for it.

Last but not least, thank you to Prof. dr. Marco J. van der Leij for all the meetings to discuss the development of this work, your valuable insights and interest towards this thesis were vital to finish it.

(5)

iv

List of Figures

4.1 Average Income according to skin color . . . 13 A.1 Average Income According to Skin Color By Region . . . 26 A.2 Faces Showed During Interview Process . . . 26

(7)

vi

List of Tables

3.1 Descriptive Statistics for Variables in the National Survey on Discrmi-nation in Mexico: Respondents Aged 18-70 . . . 12 4.1 Regression Analysis for Respondents Aged 18-70 with Positive Wages

in 2010 . . . 15 4.2 t-test for Significant Wage Differences Between Skin Color Groups

Based on Model 3 and Model 4 from Table 4.1 . . . 17 4.3 Average Tretment Effects for Respondents Aged 18-70 with Positive

Wages in 2010 . . . 18 4.4 Average Treatment Effects for Medium-Skinned and Dark-Skinned as

Potential Control Groups . . . 19 A.1 Definition of Variables: Data Source, National Survey on

Discrmina-tion in Mexico (ENADIS) . . . 23 A.2 Regression Analysis for Respondents Aged 18-70 with Positive Wages

in 2010. Dummy Variables for State Included . . . 24 A.3 Regression Analysis for Respondents Aged 18-70 with Positive Wages

(8)

1

Chapter 1

Introduction

During 2011, the first article in Mexican Constitution was modified to prohibit discrimination based on national or ethnic origin, gender, age, disability, social sta-tus, health conditions, religion, opinions, sexual orientation, marital status or any other that violates human dignity and is intended to nulify or impair the rights and freedom of individuals (Diario Oficial, 1917). However, as usually happens, there is a notable gap between theory and practice; hence, in the daily grind is common to find different actions that segregate religious, ethnic or sexual minorities and also women, all of which are considered among the most vulnerable groups to be dis-criminated.

In this sense, there is an extensive literature available nowadays dealing with different perspectives towards discrimination and also providing reliable evidence is something happening at different levels1. Nevertheless, this was not always the case. During the first half of the 20th century, there was a huge environment of denial worldwide when it came to this kind of issues already affecting diverse social groups; we may describe two main reasons why this situation arose.

One one hand, the widely accepted idea of race inferiority played an important role to strengthen intergroup differences already present during the 19th century. The superiority of White over Black races was something given for granted and attitudes of racial superiority or antipathy to Blacks were widely accepted as in-evitable and natural responses to the seemingly obvious inferiority and backward-ness of Blacks and other groups (Duckitt, 1992). This was also the case during the colonial period in Latin America, where the codification of the differences between conquerors and conquered (indigenous people) in the idea of race, assumed a dif-ferent biological structure that placed some in a natural situation of inferiority with respect to the others (Quijano, 2000). On the other hand, the lack of a solid theoret-ical background that allowed interpretations of differentials between majority and minority groups specially in Economic fields, can also explain why academia turned a blind eye to it during several decades. One of the first attemps to remedy this ne-glect is undoubtedly the pioneer work by Becker, (1957), who developed a theory of discrimination in the market, supplementing the psychologists’ and sociologists’ analysis of causes with an analysis of economic consequences.

Moreover, after recognizing it was becoming a major problem in some coun-tries, the next challenges arrived: defining and measuring it. For social sciences like Sociology or Anthropology, the task was to go out in the streets and observe intergroups behavior. Duckitt (1992), offers a historical analysis focusing on expla-nations of racial prejudice and how specific social circumstances triggered it; Sherif and Sherif, 1961, conducted an extensive and well-known experiment about the ori-gins of prejudice in social groups. Besides, according to empirical evidence, if we

1_{See for example the works by Oaxaca and Ransom, (1994), Blinder, (1973) or even an interesting}

(9)

Chapter 1. Introduction 2 look at what was happening to African-American population in United States of America (USA), for instance, it is known how many of the services (even to fulfill basic needs) were provided according to skin color or race and to the detriment of this group’s well-being. While in Economics, the first duty consisted in finding any source willing to provide numerical information that could be used to measure and explain the differences across social groups as a starting point to develop a method to measure its consequences.

Fortunately, now it is possible to give an unambiguous definition of discrimi-nation in the market place by looking at monetary aspects, i.e. income, commonly used as a measuring rod. Why, after all, should two or more groups of workers who have the same productivity receive different remmuneration? One can find a considerable number of publications that examine the influence of race on wages in the United States of America and the mechanisms through which race affects wages. Two important authors are Cain, 1986, who investigates the fundamental problems associated with income and wage differences among groups classified by sex, race, ethnicity, and other characteristics, raising the question of whether a la-bor market that pays unequal wages to equally productive workers is inefficient; and Altonji and Blank, 1999, who considered differentials by race and gender in the labor market, concluding that wage gaps have increased in the previous 15 years among blacks versus whites (particularly among women), and that although gender differences have been narrowing, they are still large.

Furthermore, when investigating the influence of race in wages, a distinguishing feature found in most of the research is a categorization of workers as black or white, consistent with the conventional USA racial and ethnic categorization. Nevertheless, the conventional wisdom has it that race is constructed in vastly different ways in the United States of America and throughout Latin America. Race ostensibly is under-stood as genotypical in USA, while race ostensibly is underunder-stood as phenotypical in Latin America (Darity Jr, Dietrich, and Hamilton, 2005). In this sense, in Mexico, the country which I am considering for this study, there are found 65 different indige-nous ethnic groups, all with a remarkable genetic variation, making some of them as different from each other as Europeans are from East Asians (Moreno-Estrada et al., 2014). Considering the latter, in this thesis I offer an alternative perspective, a link between wages and skin color, not only as a simple bivariate ranking with whites at the top and dark-skinned blacks at the bottom but rather a more gradational ranking with whites at the top and dark-skinned blacks at the bottom. The theory I am test-ing suggests that as skin-shade lightens, wages rise, leadtest-ing to a greater white-black gap for darker skinned blacks than for blacks with lighter skin2.

Several questions regarding discrimination based on skin color have become really relevant in Mexico due to a recently publication from Instituto Nacional de Estadística y Geografía (INEGI), the National Statistics Office, regarding this topic. They conducted a survey during 2016 that included a question in which people had to choose the skin color that most resembled theirs from a sample of 11 different skin tones that varied from light white to dark brown; this was the first time that such question appeared in one of the regular surveys they carry out. By implementing the latter, interesting findings concerning discrimination based on skin color across the country came out, as well as relevant features of Mexican society, like the fact that lower educated people were found more frequently within the darkest skin color groups, which grabbed the attention of media and society in general.

(10)

Chapter 1. Introduction 3 Social psychologists believe that human categorization is a necessary precondi-tion to the existence of stereotypes of social groups. In the past, skin tone has been seen as important only in its role in determining racial category membership, but today new studies are emerging to demonstrate a preference for whiteness among other ethnic groups. Maddox and Gray, 2002, suggest that skin tone plays a more complex role as both whites and blacks attribute a greater social status to lighter-skinned blacks. Espino and Franz, 2002, examine the issue of phenotypic discrimina-tion against Mexicans in the U.S. labor market, concluding that darker-skinned Mex-icans face significantly lower occupational prestige scores than their lighter-skinned counterparts even when controlling for factors that influence performance in the la-bor market.

Motivated by the theory of white preference, this work tests the hypothesis that sheds light on one question: are there differences in wages according to skin color within Mexican labor market? This is answered by using reduced-form specifica-tions that control for labor market indicators. The next chapters of this thesis are or-ganized as follows: in Chapter 2, I dig into the definition of discrimination when this topic became polemized in social science research and how it changed until reach-ing the one I am usreach-ing for this work, I also explain the statistical models I use to solve my research question. In Chapter 3, I introduce the data and explain the clean-ing process involved; further, I show descriptive statistics. In Chapter 4, I present the results along with a number of hypothesis tests. Finally, in Chapter 5, I draw conclusions from this work and formulate recommendations.

(11)

4

Chapter 2

Theoretical Background

Since the decision by the Supreme Court of the United States of America to out-law segregation by color in public schools on 1954, no single domestic issue occu-pied the country’s media more than race relations. However, during the first half of the 20th century, despite occurring in a great variety of situations and environ-ments, discrimination was an ambiguous topic among social science research3. Al-though other social scientists, notably the sociologists and anthropologists, had an early entrance into this field, this phenomenon was largely ignored by economists and their publications. With respect to discrimination based on race, this neglect may have arised after one of the following good reasons: (1) Usable statistics on African-Americans in the labor force were not easily assembled; (2) It is extremely difficult for a private individual or agency to gather exact information on African-American employment by the interview method; (3) The existence of color lines in the labor market possessed problems of method that would be formidable even in the best tabulated of worlds (Dewey, 1952).

2.1 The Definition of Economic Discrimination

2.1.1 Overview

Economic discrimination is a concept that defies precise definition. The work of Gary Becker, kicked off economists’ understanding on how the tastes of employers, customers, or coworkers could result in labor market discrimination; if an individual has a "taste for discrimination", he must act as if he was willing to pay something, either directly or in the form of a reduced income, to be associated with some per-sons instead of others. When actual discrimination occurs, he must, in fact, either pay or forfeit income for this privilege. This simple way of looking at the matter gets the essence of prejudice and discrimination (Becker, 1957). However, over the past 30 years, the theoretical work on discrimination has particularly emphasized the role of imperfect information about attributes, skills or behavior workers from the minority group might have. In the latter case (better known as statistical discrim-ination), employers use information on the average productivity of this group and the discrimination arises when it is perceived as less productive than the other. The strength of this models is that they are consistent with long run equilibria in which group differentials persist, while simpler models of taste-based discrimination often predict the elimination of discrimination through competition or segregation (Al-tonji and Blank, 1999).

3_{As already stated in Chapter 1, for a long time also social scientists thought other races were}

inferior. Thus, in spite of the research agenda available, the departure point was not always from out-groups as being of equal status, which is necessary to study discrimination

(12)

Chapter 2. Theoretical Background 5 Recent research has managed to merge ideas from both types of models, obtain-ing results worth knowobtain-ing. On one hand, Borjas and Bronars, 1989, investigated self-employment rates and income differences by race, finding that differentials arise in markets with consumer discrimination and incomplete information about the price of the good and the race of the seller. On the other hand, Black, 1995, constructed an equilibrium search model where some employers have a distaste for hiring minority workers, showing that this biased results in economic discrimination against minor-ity workers but that their wages increased when their proportion increases in the labor market. It should be noticed that these papers point out the role of imperfect information about the locations and prefereces of customers, employees or employ-ers as it will limit the ability of competition and segregation to eliminate the effects of prejudice on labor market outcomes.

2.1.2 Defining Discrimination

The theoretical question behind discrimination and that I will use as a starting point for defining it is: Under what conditions will essentially identical goods have dif-ferent prices in competitive markets?. Economic discrimination refers to a group rather than to an individual and within the labor market context, it takes labor services as the good in question and the wage rate as the price. For instance, if an employer feels he is incurring in a disutility or loss by hiring a minority worker solely because of his or her particular demographic characteristic, which by itself is not important to the worker’s physical productivity, then employers may said to be prejudiced. Nevertheless, if the majority group of coworkers manifest their feelings of psychic disutility by actions which curtail the minority worker’s phsysical productivity, this outcome will be considered discriminatory (Becker, 1957). Under some but not all conditions, these tastes will lead to discrimination, defined by wages to the minority group being below what they would receive if only their physical productivity was determinant. For example, at the level of a potential worker - taking the skin color as the common variable found among the minority group - discrimination based on skin color is said to arise if an otherwise identical person is treated differently by virtue of that person’s skin color or gender, and those characteristics by themselves have no direct effect on productivity. Discrimination is a causal effect defined by a hypothetical ceteris paribus conceptual experiment - varying race but keeping all else constant (Heckman, 1998).

In other words, labor market discrimination arises when persons who provide labor market services and who are equally productive in a physical or material sense are treated unequally in a way that is related to an observable characteristic such as race, ethnicity, or gender. By "unequal" it is understood that these persons receive different wages or face different demands for their services at a given wage (Altonji and Blank, 1999).

2.2 Statistical Models for Discrimination

Two general types of statistical models may be used to analyze discrimination. In model 1, the outcome variable of interest - income or wages - is compared for the two (or more) groups, holding constant certain variables that are believed to either affect the outcome variable (or be relevant to the interpretation of the outcome variable) or be exogenous to the process of discrimination under study. Whereas, in model 2, all included characteristics are considered endogenous, and any difference across

(13)

Chapter 2. Theoretical Background 6 groups is attributed to the process of discrimination under study. The practical-and-theoretical problem of differences in wages for equally productive workers is generally examined by model 1 (Cain, 1986). Therefore, in this work we will focus on this type of model to answer the research question.

2.2.1 Regression Analysis

Also following Cain (1986), let Yi = the outcome of the process, such as the in-come, earnings, of wage for the ith person; Xi = a vector of productivity character-istics of the ith person that are presumed exogenous in that they do not depend on Y nor on the particular form of economic discrimination under study; Zi = 1 if the person is in the majority group and 0 if in the minority group; ei = a random error term; and let A and B be coefficients representing the effects on Y of Z and X. As-suming a linear and additive model for convenience and suppressing subscripts to avoid clutter, we have

Y = X’B + AZ + e (2.1)

Then a regression in which we find A>0 would be evidence of discrimination. The contrary case is assumed to be A = 0; "positive discrimination" (A<0) is not considered. In this type of model, the two groups designated by Z are assumed to provide "essentially identical" labor services, conditional on (holding constant) X. Equivalently, we could define market discrimination, D, as

D = ( ˆY |X, Z = 1) − ( ˆY |X, Z = 0)

where ˆY is the predicted value of Y conditional on X, so in Equation 2.1, D = A (Cain, 1986).

2.2.2 Average Treatment Effects

Suppose we have observed a sample of subjects, some of whom received a treat-ment and the rest of whom did not. In social science applications, a treattreat-ment could be participation in a job-training program or inclusion in a classroom or school in which a new pedagogical method is being used (StataCorp., 2013). Estimating aver-age treatment effects (ATE) has become important in the program evaluation litera-ture, and as an additional approach used in this work to study wage gaps according to skin color, we will consider the skin colors available as a different treatment each individual in the survey has been exposed to4. As an extension of this estimators and due to the solution it provides for the omitted variables problem our model might be having5we are going to use it to dig into the wage diferentials when con-sidering groups with very similar demographic characteristics.

Literature on treatment effects begins with a counterfactual, where each individ-ual has an outcome with and without treatment. Let y1 denote the outcome with treatment and y0the outcome without treatment; because and individual cannot be in both states, we cannot observe boty y0and y1. To mesure the effect of a treatment, we are interested in the difference in the outcomes with and without treatment, y1

4_{We stress that each individual chose the skin color he or she identified himself or herself with} 5_{When using log wage as dependent variable, there are diverse factors that have an influence on it}

and that for several reasons we can’t include in the model. From linear effects like IQ or Work Attitude to cuadratic effects such as Age2_.

(14)

Chapter 2. Theoretical Background 7 - y0. Because this is a random variable, we must be clear about what feature of its distribution we want to estimate (Wooldridge, 2010).

According to Rosenbaum and Rubin, 1983, the quantity of interest is the ATE, τate= E(y1− y0) (2.2) which is the expected effect of treatment on a randomly drawn person from the population. If we let wi = 1 if unit i is assigned to the experimental treatment, and wi = 0 if unit i is not assigned to the experimental treatment, a second quantity of interest is the average treatment effect on the treated (ATT),

τatt = E(y₁− y₀|w = 1) (2.3) which is the mean effect for those who actually received one of the treatments. As noted previously, the difficulty in estimating equations 2.2 or 2.3 is that we observe only y0or y1, not both, for each person. More precisely, along with w, the observed outcome is

y = (1 − w)y₀+wy₁=y₀+w(y₁− y₀) (2.4) Therefore, the question is: How can we estimate ATE or ATT with a random sam-ple on y and w (and usually some observed covariates)? (Wooldridge, 2010) We can-not estimate them by simply talking the difference between the sample means for the treated and untreated subjects, because there are covariates that are related to the po-tential outcomes and the treatment. To start dealing with this question, Rosenbaum and Rubin, 1983, introduced the following assumption, which they called ignorabil-ity of treatment (given observed covariates x): Conditional on x, w and (y0,y1) are independent, this implies that

E(y₀|x, w) = E(y₀|x) E(y₁|x, w) = E(y₁|x)

The idea underlying this assumption is that if we can observe enough informa-tion (contained in x) that determines treatment, then (y0,y1) might be mean indepen-dent of w, conditional on x. Loosely, even though (y0,y1) and w migh be correlated, they are uncorrelated once we partial out x (Wooldridge, 2010). Next, assuming that ignorability holds, estimating ATE will require being able to observe both treated and non-treated units for every outcome on x. This assumption is typically called the overlap assumption and it states that for any setting of covariates in the assumed population, there is a chance of seeing units in both the treatment and non-treatment groups. The probability of treatment, as a function of x, plays a very important role in estimating average treatment effects. It is usually called the propensity score and it is usually denoted by:

p(x) = P (w = 1|x) (2.5) Thus, we also need to estimate the propensity score. In practice, except when

x takes on only a few values, this estimation will be either zero or one for most values of our sample. Consequently, in order to estimate propensity scores, some moedelling will be required. The propensity score can often be modelled using an appropiate logit model or discriminant score (Rosenbaum and Rubin, 1983).

(15)

Chapter 2. Theoretical Background 8 Given the previous ignorability and overlap assumptions, one can establish that estimators for ATE and ATT are identified. According to Wooldridge, 2010, one way to do it (and that motivates the estimation method we will apply to our data later), is to use inverse propensity score weighting. If we mantain the ignorability of treatment assumption, notting that wy = w1, we have, by iterated expectations,

E[ wy p(x)|x] = E[ wy₁ p(x)|x] = = E{E[wy1 p(x)|x, w]|x} = E{ wE(y₁|x, w) p(x) |x} = = E{wE(y1|x) p(x) |x} = E{ w p(x)|x}µ1(x) = µ1(x)

because E(w|x) = p(x), with µ1(x) = E(y | x,w = 1). A similar argument shows that

E[(1 −w)y

[1 −p(x)]|x] = µ0(x)

with µ0(x) = E(y | x,w = 0). Combining these two results and using simple algebra gives

E{ [w − p(x)]y

p(x)[1 − p(x)]|x} = µ1(x) − µ0(x) = τate(x) (2.6) Moreover, if we also mantain the overlap assumption, we can use iterated expec-tations to write

τate= E{

[w − p(x)]y

p(x)[1 − p(x)]} (2.7) which, because w, y, and x are all observed, establishes identification of τate us-ing the propensity score (Wooldridge, 1999c). Usus-ing a slightly different argument, Dehejia and Wahba, 1999, arrived to the following expression for τatt:

τatt= E{[w − p(x)]y

ρ[1 −p(x)]} (2.8)

with ρ ≡ P(w = 1), the unconditional probability of treatment. The latter identi-fication results in equations 2.7 and 2.8 can be directly turned into estimating equa-tions for τateand τatt.

2.2.3 Matching Methods

Recalling the definition of descrimination we provided earlier, it is a causal effect from a conceptual experiment - varying some characteristic but keeping everything else constant. In this sense, we will address matching estimators as a feasible mea-sure of possible discrimination based on skin color. Matching estimators are based on the idea of comparing the outcomes of subjects that are as similar as possible with the sole exception of their treatment status (StataCorp., 2013). In particular, for each observation, we impute values for the counterfactuals, yi0and yi1. Matching estima-tors use the observed outcomes when possible. In other words, if we let ˆyi0 and ˆyi1 denote the imputed values, ˆyi0 = yi when wi = 0 and ˆyi1= yi when wi = 1. Generally, matching estimators thake the forms

(16)

Chapter 2. Theoretical Background 9 ˆ τate,match= N−1 N X i=1 (ˆyi1− ˆyi0) (2.9) ˆ τatt,match= N−1 N X i=1 wi(yi− ˆyi0) (2.10) where the latter formula uses the fact that yi1 = yi for the treated subsample (Wooldridge, 2010).

(17)

10

Chapter 3

Data

3.1 About the Information

The data used in this work comes from the National Survey on Discrimina-tion in Mexico (ENADIS) made by the NaDiscrimina-tional Council to Prevent DiscriminaDiscrimina-tion (CONAPRED). From October 2014 to November 23, 2010, 13,751 homes were vis-ited. That number granted information coming from 52,095 persons. Homes were selected in each of the 32 federated entities of the country along 301 counties and 1,359 departing points. The sample used for the selection was random, multi-staged, stratified, conglomerated and the primary sample units were generally selected ac-cording to probability proportional to the size of the population. The following in-struments, specifically designated for survey research, were applied in the ENADIS: 1. A home questionnaire to learn the characteristics of the selected households and

their conditions of living.

2. A questionnaire of opinion applied to one of the home members, randomly se-lected, to learn values, attitudes and practices regarding discrimination. 3. Ten questionnaires for vulnerable groups oriented to persons pertaining to groups

susceptible to be discriminated in order to gather their perceptions, attitudes and values about discrimination and the conditions of their group of vulnera-bility.

4. A victimization questionnaire oriented to record the experiences of discrimina-tion against the populadiscrimina-tion pertaining to any of the vulnerable groups.

For the purpose of this thesis, I only used information available in the first two questionnaires. The home questionnaire contained all the demographics I included as covariates in the model as well as the income respondents reported. In this sense, I stress that the only income found in the data set was the total household monthly in-come. Therefore, I made the assumption that this number represents the wage of the person answering the survey6. From the questionnaire of opinion, I used information regarding the skin color. The respondents were shown 9 different faces that varied according to their skin color, from a ligther to a darker one and they had to choose the one they identified themselves with7. The latter constitutes a notable drawback from the information. Due to the way it was collected, we face an important bias from both income and skin color. The former because people might have increased

6_{There are, however, some equivalence scales that assign each household type in the population}

an income fraction in proportion to its needs. For more information see Atkinson, Rainwater, and Smeeding, (1995) for a wide range of equivalence scales.

7_{A sample of the images shown during the survey can be found in Figure A.2 from the Appendix}

(18)

Chapter 3. Data 11 or decreased the real household income; the latter, because although people only had 9 different predefined colors, it was still selfchosen8. Finally, the information contained an expansion factor calibrated with census data to allow the results be-come national representative. I used the latter as a sampling weight for obtaining the statistics and results.

As stated before, one randomly chosen person from the visited house answered both questionnaires containing the information I am using in this thesis, therefore I am only taking into account the 13,751 respondents. Nevertheless, many of them either did not report any skin color or answered I do not know, hence, because this answer is crucial in my research I decided not no consider those observations. Fur-thermore, I bounded the age limit of my sample to individuals between 18 and 70 years old in order to ensure those ones were fully or partial employed. After this, I ended up with 10,322 observations that I used to run the models presented.

In Table 3.1, descriptive statistics of the information reported can be found. Col-umn 1 shows information about the lightest skin color available, and colCol-umn 9 does it about the darkest skin color available. The first thing to notice is that there is an income decrease of around 2,000 Mexican Pesos from lightest to darkest skin; fur-thermore, it is interesting to notice that within each of the three subsamples, the same effect is present: as skin darkness the income seems to decrease. Another fea-ture worth noticing is the years of education. We can see that the lightest people with the lightest skin color are reaching around 12 years of education (which repre-sents a High School level), whereas people with the darkest skin color reach 10 years of education (representing JR High School). Hence, as we can see, this latter group seems to have less education than the former, this interesting feature will become important on our further analysis.

Regarding the other demographic characteristics, we see that across groups Mar-ried indicator variable has a similar behavior, meaning that half of responents are married. The same happens for the number of members in the house, which moves around 4. According to Gender variable, there were more women that answered the survey, this may be perhaps due to the fact that in Mexico is common to find that men go to work while women stay at home, meaning that more of them were there when the interviewers visited the houses. Very few people declared to speak an Indigenous Language, in contrast, many people identified themselves as Catholic, which also shows the huge presence of this religion within Mexican society. Finally, although is not the majority, there were an important presence of migrants among respondents. Overall, besides Income and Years of Education, all other variables seem to behave very similar across skin color groups.

8_{According to UNAM Anthropological Institute, there are 36 different skin colors in the world. For}

the purpose of this survey, they developed a sample of 9 male and female most representative skin colors in Mexico, which was used to answer this question on the survey

(19)

Chapter

3.

Data

12

TABLE3.1: Descriptive Statistics for Variables in the National Survey on Discrmination in Mexico: Respondents Aged 18-70

Variables Light Medium Dark

1 (n=1619) 2 (n=6420) 3 (n=2760) 4 (n=1517) 5 (n=1025) 6 (n=588) 7 (n=950) 8 (n=807) 9 (n=414) Income 8149.65 (345.59) 8278.47 (550.83) 6807.38 (223.99) 6504.61 (223.99) 6533.76 (315.60) 6064.45 (420.13) 6235.48 (325.97) 5747.46 (339.72) 5914.75 (422.12) log Income 8.65 (0.04) 8.70 (0.06) 8.52 (0.03) 8.48 (0.04) 8.45 (0.05) 8.41 (0.06) 8.41 (0.05) 8.28 (0.06) 8.34 (0.07) Age 39.65 (0.67) 39.64 (0.87) 36.14 (0.45) 38.35 (0.65) 36.65 (0.83) 38.30 (0.95) 38.20 (0.85) 40.52 (0.98) 41.20 (1.16) Gender 0.71 (0.02) 0.67 (0.03) 0.73 (0.02) 0.66 (0.02) 0.66 (0.03) 0.67 (0.04) 0.50 (0.03) 0.60 (0.03) 0.63 (0.04) Schooling Years 12.19 (0.26) 12.54 (0.39) 11.87 (0.16) 11.48 (0.22) 11.54 (0.28) 10.84 (0.31) 11.27 (0.23) 10.44 (0.34) 9.72 (0.33) Married 0.47 (0.02) 0.46 (0.04) 0.45 (0.02) 0.46 (0.02) 0.43 (0.03) 0.53 (0.04) 0.50 (0.03) 0.51 (0.03) 0.49 (0.04) Indigenous Language 0.01 (0.005) 0.04 (0.01) 0.02 (0.006) 0.01 (0.01) 0.05 (0.02) 0.09 (0.02) 0.11 (0.02) 0.12 (0.02) 0.15 (0.03 Family Members 4.02 (0.15) 3.91 (0.14) 4.45 (0.11) 4.49 (0.17) 4.47 (0.25) 4.73 (0.27) 4.29 (0.18) 4.35 (0.15) 4.00 (0.18) Migrant 0.23 (0.02) 0.25 (0.04) 0.22 (0.02) 0.17 (0.02) 0.16 (0.02) 0.27 (0.04) 0.17 (0.02) 0.22 (0.02) 0.20 (0.04) Work 0.55 (0.01) 0.04 (0.01) 0.05 (0.01) 0.07 (0.01) 0.05 (0.01) 0.05 (0.02) 0.09 (0.02) 0.14 (0.02) 0.13 (0.03) Catholic 0.88 (0.02) 0.82 (0.04) 0.89 (0.01) 0.90 (0.02) 0.88 (0.02) 0.89 (0.02) 0.88 (0.02) 0.91 (0.02) 0.85 (0.03)

Note: Data Source: National Survey on Discrmination in Mexico (ENADIS). Means are reported, with their standard errors in parenthesis, for the subsamples of each skin color reported. Variables are described in Appendix Table 1.

(20)

13

Chapter 4

Results

4.1 Regression Analysis

We may first take a look at the distribution of the income according to the nine different skin colors people could identify themselves with. In Figure 4.1, a bar graph is displayed with the average income for each skin color, from the darkest (A) to the lightest (I). We can see there is an upward trend as skin color lightens, not only within subgroups from each of the three skin color groups, but also if we consider all nine groups available, which shows and increment of 30% on the average income from group A to group I.

FIGURE4.1: Average Income according to skin color

In Table 4.1 several models are presented. All of them share the same explained variable, natural logarithm of wage, but vary with respect to the explanatory vari-ables included. In the first of this models, eight dummy varivari-ables to control for skin color are included; the lightest one is used as a reference. As expected, by looking at the coefficients’ sign, we notice almost all of them have an inverse relation with the dependent variable, that is, people with non-white skin color are having a lower in-come. Moreover, as the coefficient in absolute value is increasing, we can say that on average, the darker the skin, the lower the income. It is also noticed that seven out of the eight indicator variables are significant at 1%; just the WhiteMedium variable, which is the indicator for the second lighter skin color, besides showing a positive coeffcient, has no statistical significance, meaning we can not really stare there is any difference between this income’s group and the white-skinned one. Finally, is

(21)

Chapter 4. Results 14 also interesting to see that the smallest coefficient is not found in the darkest skin color available, but one color before, meaning this group, on average, has the lowest income compared with the white-skin one, 35% lower.

In model 2, besides controlling for skin color, I am adding dummy variables to control for geographical regions in Mexican Republic using the North as reference9. Although, this is something easily found in the literature, the main reason that mo-tivates the inclusion of such variables in this work is that the paternal ancestry es-timated in western Mexico is mainly European (60-64%). Significant genetic hetero-geneity was established between Mexicans from western (Jalisco State) and northern Mexico (Chihuahua State) compared with Mexicans from the center of the Mexican Republic (Mexico City), this attributable to higher European ancestry in western and northern than in central and southeast populations, where higher Amerindian an-cestry was inferred (Rangel-Villalobos et al., 2008). Therefore, one might think the region not only will affect the job opportunities available as the north part of the country is well known for having higher industrial development than the south10, but also inhabitants’ skin color due to their different inheritage. If we look at the coefficients estimated for skin color dummies, there is a remarkable decrease (in ab-solute value) compared to the ones from the previous model. The reduction goes from 22% for the DarkDark variable, to almost 70% for the MediumDark. Most of them are significant at 5% at least, and also have a negative sign. Next, is interesting to see how the coefficients for the region indicator variables are behaving. As one could have expected, compared to the North, on average, people in any other region of the country are perceiving less income, with the East and South, the ones where the wage decreases the most, around 50%. It is also good to notice that only the indicator variable for West region is not significant at 1% as all the others, meaning that compared with North, there is not much differences in wages to explain.

In a third model, I added a number of demographic variables. The first thing to notice is that the indicator variables for Migrant and Catholic are not significant, un-like every other demographic variable that is significant at 1% level. Next, we should notice that both Type of Work and Indigenous Language indicator variables, have the highest coefficients (in absolute value), the latter variable being related to having an indigenous inheritage. Special attention deserves the former one as it shows a 58% of decreasing on the wage if the survey respondent works for outdoors related activ-ities which may influence their sun exposure and hence their skin color. However, it should also be stressed that this type of activities (farming, fishing, hunting, etc.) are usually in the lowest percentiles of the income distribution. Finally, there is evidence of a decrease in wages for women as the Gender variable coefficient is -20.85.

At this point, there is room for the reader to ask: Why isn’t Education included so far?. As you may know, it is undeniable the importance of schooling to explain wages and there is many research that confirms it and deals with different and in-teresting approaches11_{. To fulfill this requirement, in model four I finally included} Schooling Years variable. To begin with, it makes sense that the variable has a pos-itive sign, meaning that by an additional year of schooling, the wage received will increase 7.2%, besides it shows statistical significance at 1% level. However, the most important effect of including this variable into the model is to notice what happens with the dummy variables for Skin Color. Not only the estimate coefficients have

9_{An additional model was performed using dummy variables for 31 of the 32 states in Mexican}

Republic. It is shown in Appendix A, Table A.2

10_{A bar chart of the average income per region is provided in Figure A.1 of the Appendix A as a}

reference

(22)

Chapter 4. Results 15

TABLE 4.1: Regression Analysis for Respondents Aged 18-70 with Positive Wages in 2010

Dependent Variable: log wage (1) (2) (3) (4)

Constant 8.653* (.0383) 8.804* (.0403) 9.100* (.0888) 7.688* (.0934) Skin Color DarkDark - .3091* (.0784) - .2528* (.0762) - .1776* (.0754) - .0343 (.0713) DarkMedium - .3730* (.0712) - .2974* (.0682) - .2409* (.0638) - .1451** (.0591) DarkLight - .2415* (.0630) - .1664* (.0620) - .1749* (.0594) - .0973*** (.0570) MediumDark - .2395* (.0764) - .1413** (.0728) - .1631* (.0684) - .0699 (.0650) MediumMedium - .2007* (.0642) - .1564* (.0604) - .1833* (.0613) - .1157** (.0548) MediumLight - .1690* (.0562) - .1255** (.0558) - .1273** (.0568) - .0729 (.0531) WhiteDark - .1354* (.0483) - .0995** (.0473) - .1223* (.0471) - .0793*** (.0435) WhiteMedium .0518 (.0702) .0934 (.0696) .0789 (.0665) .0531 (.0560) Region Center • - .1152* (.0371) - .1258* (.0357) - .1445* (.0322) East • - .5317* (.0487) - .4723* (.0450) - .4282* (.0404) West • - .0829** (.0411) - .0720** (.0414) - .0156 (.0359) South • - .4579* (.0444) - .3848* (.0442) - .3566* (.0417) Demographics Married • • .0918* (.0297) .0931* (.0272) Gender • • - .2085* (.0321) - .1009* (.0272) Migrant • • - .0284 (.0371) - .0020 (.0348) Catholic • • - .0551 (.0474) - .0217 (.0463) Family Members • • .0245* (.0096) .0373* (.0093) Indigenous Language • • - .3917* (.0802) - .2705* (.0755) Age • • - .0053* (.0012) .0035* (.0011) Type of Work • • - .5823* (.0624) - .3932* (.0581) Schooling Years • • • .0720* (.0033) N 10,322 10,322 10,322 10,322 Adjusted R Squared .0180 .0742 .1402 .2698 Note: Coefficient estimates using OLS are reported and robust standard errors are shown in parentheses. Variables are described in Appendix A, Table A.1. The reference cate-gories are White Skin for the Skin Color and North for the Region. *, **, *** indicates significance at the 1, 5 and 10 percent levels, respectively.

(23)

Chapter 4. Results 16 changed compared to the previous model, but also the statistical significance has been influenced. One can see that none of them is significant at 1% level anymore, just the Dark Medium and Medium Medium indicator variables are significant at 5% level and White Dark and Dark Light also are but at 10% level. Overall, it is easy to notice how after I included a variable regarding Education in the model, Schooling Years, the strength of the relationship between Skin Color and Wages has dropped, in other words, Education is mediating this effect.

A mediator is a third variable that links a cause and an effect; its purpose is to enhance a deeper and more refined understanding of a causal relationship between an independent and a dependent variable (Wu and Zumbo, 2008). A variable is said to function as a mediator when a previously significant relation between the independent and dependent variables is no longer significant, with the strongest demonstration of mediation occurring if they become independent from each other (Baron and Kenny, 1986).

To sum up, according to this analysis, there are differences in wages according to skin color within mexican labor market as people with darker skins seem to receive lower wages than people with lighter skins, however it is still unclear this differences are due to discrimination based on skin color. In this sense, the role Education is playing in the model drives us to think that lower wages for dark-skinned happen not because employers pay them less, but because they are perhaps lower educated 12_.

4.1.1 Hypothesis Testing

According to the coeffcients we obtained from the regression analysis presented before, it seems that there are remarkable differences among incomes considering the skin color. However, an important part of assuring this is indeed the case, is by hypothesis testing. Therefore, in Table 4.2, I am reporting t-statistics for the test H0: βi = βjwith i6=j. I will focus on models (3) and (4) from Table 4.1 and also, the statistics I am presenting refer to the coefficients for the skin color indicator variables. According to this tests, there is no significant difference across the coeficients from most of skin color groups, in other words, from this t-tests, we can’t conclude there is difference in wage according to skin color. However, it is interesting to see that the tests do show significative differences in wages if we compare some of the darkest skin groups with the lighest ones. Furthermore, the Dark Medium group is showing significative differences at 10% and 5% level with two out of the three lightest skin colors available. It is also interesting to look at what is happening with White Medium group, the second lightest skin color, as it is the only one showing remarkable differences (at 1% level) with any other. Despite the non-significance of most of the tests, we still observe that wages between this group are different from the other groups and it makes sense to think that this difference is in favor of the lighter skin color. If we now look at the tests derived from model 4, we may find some changes. Recalling that this model included the Schooling Years variable which influenced at some point the relation found between wages and skin color, hence this effect is also visible here. Now, not only just the White Medium group is the only

12_{Suprisingly, this conclusion (drew out from a survey lifted in 2010 by The National Council to}

Prevent Discrimination in Mexico) is in line with what the National Institute of Statistics, Geography and Informatics (INEGI) found out with the results of a different survey lifted in 2016. This latter study found that from people identified themselves with the lightests skins, just 10% do not have any Schooling Years, whereas this number soars to 20.2% for those who grouped themselves in the darkest skins.

(24)

TABLE4.2: t-test for Significant Wage Differences Between Skin Color Groups Based on Model 3 and Model 4 from Table 4.1

Panel A: Model 3 Skin Color Dark

Medium Dark Light Medium Dark Medium Medium Medium Light White Dark White Medium Dark Dark 0.77 -0.03 -0.17 0.07 -0.65 -0.78 -3.03* Dark Medium -0.97 -1.03 -0.82 -1.72** -2.05** -4.29* Dark Light -0.16 0.13 -0.78 -0.99 -3.62* Medium Dark 0.27 -0.50 -0.64 -3.11* Medium Medium -0.88 -1.10 -3.61* Medium Light -0.10 -3.02* White Dark -3.33* Panel B: Model 4 Skin Color Dark

Medium Dark Light Medium Dark Medium Medium Medium Light White Dark White Medium Dark Dark 1.34 0.80 0.41 1.00 0.50 0.63 -1.03 Dark Medium -0.70 -0.99 -0.42 -1.09 -1.13 -2.66* Dark Light -0.38 0.28 -0.40 -0.34 -2.15** Medium Dark 0.61 0.04 0.15 -1.58*** Medium Medium -0.67 -0.65 -2.33** Medium Light 0.13 -1.84** White Dark -2.19**

Note: *, **, *** indicates significance at the 1, 5 and 10 percent levels, respectively.

one showing significative differences in wages compared to other groups, but also this differences have lost significance. We can say that the inlcusion of the education variable wipes out some of the differences across most of the groups, however there are still differences between the lightest skin colors and all of the other ones we found.

4.2 Matching Methods

As an additional approach to investigate wage differences according to skin color, but also to mitigate specification problems from the model derived from omitted variables bias it might have, I also estimated the average treatment effects. Accord-ing to the results in both panels of Table 4.2, we do not see significant differences in coefficients across all skin color groups, however we do discover that the coef-ficient of the lightest skin color groups (specifically light white) are different from the others. In this sense, to carry on with the analysis, we collapsed the data in two groups: light-skinned and non-light-skinned. In the former, every respondent with White Dark, White Medium or White Light was grouped and, in the latter, any other skin color. In Table A.3 from Appendix A, the result from an OLS model is shown; the dependent variable is log wage, this new categorization of skin colors is used, and all other covariates used in model remain the same as in model (4) from Table 4.1. As one can see, the results did not change much, nevertheless the main change we point out is that even with the Schooling Years as an explanatory variable, the negative relation between the wage and being part of the non-light-skinned group is still significant. Now, in Table 4.3, Treatment-effects estimations are reported.

(25)

TABLE 4.3: Average Tretment Effects for Respondents Aged 18-70 with Positive Wages in 2010

Dependent Variable: Wage

Average Treatment Effect

Average Treatment Effect of the Treated Light-Skinned (NNM) 229** (115.51) 263.17** (122.79) Light-Skinned (PSM) 215.64* (115.43) 251.48** (123.64)

Note: Coefficient estimates using Nearest-Neighbor Matching with Mahalanobis as Dis-tance Metric and Propensity Score Matching using a logit as treatment model are re-ported; robust standard errors are shown in parentheses. Schooling Years, Gender, In-digenous Language, Work, Married, and the Region were the covariates used to choose matches. Maximum matches per observation 222. *, **, *** indicates significance at the 1, 5 and 10 percent levels, respectively.

Matching estimators use an average of the outcomes of the nearest individuals to impute the missing potential outcome for each sampled individual. The difference between the observed outcome and the imputed potential outcome is an estimate of the individual-level treatment effect. These estimated individual-level treatment effects are averaged to estimate the ATE of the ATET. Nearest-Neighbor Matching de-termines the "nearest" by using a weighted function of the covariates for each ob-servation. Propensity Score Matching determines the "nearest" by using the estimated treatment probabilities, which are known as the propensity scores (StataCorp., 2013). As we may notice, results from both estimating methods are very similar. To be-gin with, both of them report that, on average, wage will increase for Light-Skinned people in 229 Mexican Pesos with the NNM and in 215.64 Mexican Pesos with the PSM. It is also interesting to look at the ATT. If we remember, this number shows the average treatment effect among those that receive the treatment, being Light-Skinned for our case. Hence, this average is 263.17 Mexican Pesos with the NNM and 251.48 Mexican Pesos with the PSM; furthermore, as the ATT is lower, we can say that just for being Light-Skinned, a person will receive a higher wage, on av-erage, than a non-light-skinned. Also, the latter results are significant at 5% and, according to how I grouped skin color for this analysis, we can say the more wage differences are not across all groups, but specially with the lighest ones and all the others. Thus, this could be an indicator of the unequal wealth distribution that is highly present within Mexican society13(Esquivel Hernández, 2015; Krozer and Moreno-Brid, 2014).

4.2.1 About the Assumptions

As discussed earlier, there are two key assumptions regarding average treatment models: ignorability of treatment and overlap. We will focus on the first one because as shown in Table 3.1, the distribution of the covariates across the 9 different skin col-ors is very similar, meaning that individuals with the same demographic characters are found in both white-skinned and non-white-skinned groups considered. How-ever, in most situations the ignorability assumption is not directly testable, however

13_{According to the OCDE, for 2014, Gini coefficient in Mexico was 0.459. The Gini coefficient takes}

values between 0 (where every person has the same income), and 1 (where all income goes to one person).

(26)

Chapter 4. Results 19 Imbens and Wooldridge, (2007), discuss a number of indirect ways to assess it. These methods typically rely on estimating a causal effect that is known to equal zero. If based on the test we reject the null hypothesis that this causal effect varies from zero, the unconfoundedness assumption is considered less plausible. One interpre-tation of this tests (and the one I am following in this thesis), is to compare average treatment effects estimated using two potential control groups. This is equivalent to estimate an ATE using only the two control groups, with the treatment indicator now a dummy for being a member of the first group. In that case the treatment effect is known to be zero, and statistical evidence of a non-zero effect implies that at least one of the control groups is invalid. As I had grouped respondents in 9 skin colors at the beginning, I will use the three darkest skin colors as the first potential control group, and the three medium skin colors as the additional control group. In Table 4.4 results regarding these estimations are shown.

TABLE 4.4: Average Treatment Effects for Medium-Skinned and

Dark-Skinned as Potential Control Groups

Dependent Variable: Wage Panel A: NNM Average Treatment Effect p-value Medium-Skinned -59.01 (120.9) 0.625 Dark-Skinned -262.17 (164.3) 0.111 Panel B: PSM Average Treatment Effect p-value Medium-Skinned -71.58 (121.1) 0.554 Dark-Skinned -241.05 (157.4) 0.126

Note: Coefficient estimates using Nearest-Neighbor Matching with Mahalanobis as Dis-tance Metric and Propensity Score Matching using a logit as treatment model are re-ported; robust standard errors are shown in parentheses. Schooling Years, Gender, In-digenous Language, Work, Married, and the Region were the covariates used to choose matches. Maximum matches per observation 301. p-values for the hypothesis that the estimated coeffcient is zero are also reported.

There is an expected negative relationship between the wage and being part of either Medium-Skinned or Dark-Skinned group, meaning that, on average, their wages are lower. Nevertheless, our main concern should be the p-values reported, as they do not show any statistical significance even at 10% level, which supports the idea that ignorability of treatment holds in the original estimations on Table 4.3. I should also add, that this results are also in line with what I found in the Hypoth-esis Testing section (Table 4.2); we cannot really conclude there are wage differences across all skin color groups but they arise when comparing the lightest skin color with any other.

(27)

20

Chapter 5

Conclusions

Discrimination is a topic undoubtedly present (and perhaps more visible ev-eryday) in today’s society. Unofortunately, this was not always the case; either a "natural" superiority of one group to another or little availability of data and methods, there are many reasons described in the literature that led to different and diverse perspectives when it came to talking about differences across several groups. Gradually, more and more scholars uncovered what segregation was about and made it more relevant with important publications communicating what people lived but barely talked about as an increasing number of investigations concerning this themes during the second half of the 20th century were done. With the work by Becker, 1957, a clear pioneer in the field that started defining and measuring the inequal circunstances African-American population was living in the United States of America; not only concerning their jobs, but also suffering of unequal treatments by other people in the streets and public services.

The way I measured discrimination within Mexican labor market in this work is by looking at the income people are receiving from their job, digging in the possible gaps that can be found between individuals sharing similar demographic character-istics, years of education or places of origin, but that differ in their skin color. While most of the work regarding wage gaps is concerned about race, the reason of choos-ing skin color for this thesis follows the understandchoos-ing of the latter concept through Latin America and its relationship with the concept of race. Besides, the question I am answering in this work comes from the hypothesis known as preference for white-ness, widely known to be present in several countries. The data from the National Survey on Discrimination in Mexico, 2010, one of the few instruments (maybe the only one at that time) that provided information about both skin color and income in Mexico, was used to perform the analysis presented.

Using an Ordinary Least Squares regression model with log wage as dependent variable and 8 indicator variables to control for skin color (using the lightest one as a reference category), I found that wage seemed to gradually decrease, on average, as the skin tone darkens until reaching a 31% less for the darkest tone available. How-ever, the biggest drop was not found in this color but one slighlty lighter, and found immediatly before on the scale, as this showed a fall of 37% compared to the lightest skin color available in our sample. As I included covariates like region and other demographic characteristics, the differences across skin colors began to become less present until I included the variable Schooling Years, to control for the level of educa-tion the respondents had at the moment of the survey, making 4 out of the 8 dummy variable for skin color not statistically significant even at 10% level, Dark Dark vari-able, indicator for the darkest skin tone availvari-able, among them. The other 4 variables also lost significance but not above the 10% level. Therefore, the first conclusion that arose from this thesis is that education is acting as a mediator variable in the model; that is, people with the darkest skin colors are indeed receiving a lower income than

(28)

Chapter 5. Conclusions 21 people with lighter skin colors, nevertheless, those differences are present due to the fact that the first group seems to be lower educated than the second and, as we may know, schooling has a huge impact on wages. The latter statement is confirmed with recent information that the National Statistics Office in Mexico has published about Education and Skin Color earlier this year.

After obtaining estimation coefficients for variables used in the model, it was now important to test whether they were statistically different from each other, spe-cially the ones related to skin color. I performed different hypothesis tests using t-statistics derived from the estimation processes. Although, from the regression analysis one could conclude there was, on average, wage differences from lighter skin color to darker ones, the hypothesis testing did not confirm this idea at all. Suprinsingly, the null hyphothesis of coefficients being equal was not rejected on most of the time when comparing coefficients across skin color groups, not even at 10% level. However, when testing the difference between one of the three lighter skin colors and any darker one, the null hypothesis was rejected several times, and if the White Medium coefficient was the one compared with any darker other, this happened in every test performed even at 1% significance level. Thus, the latter led to the conclusion that the most relevant differences in wages arise when comparing the lightest skin color with all the other ones, and not across any other skin color group and this differences were in favor of the former group, representing a gain in wage for them.

Motivated by the latter results and following one of the definitions of discrim-ination provided in Chapter 2, I decided to use an Average Treatment Effect Esti-mation, in particular one that allowed me to match similar individuals found in the sample according to demographic characteristics but different in skin color, hence I grouped the sample in two groups: White-Skinned (the three lightest skin colors available) and Non-White-Skinned (any other group) and obtained Average Treat-ment Effect and Average TreatTreat-ment Effect of the Treated Estimations using both Nearest-Neighboor and Propensity Score Matching. The results strength what I ob-tained before as report an increase of 229 and 216 Mexican Pesos, respectively if an individual belongs to the Light-Skinned group. Furthermore, to verify the ignorabil-ity of the treatment assumption, which is essential for this model, using the same es-timation methods, I digged in the possible wage differences using Medium-Skinned and Dark-Skinned groups, finding no statistic significant differences between them. This also follows the idea that the most important wage gaps are found between lightest skin color groups and all the others.

It is important to add that the results of this work mainly show the presence of statistical discrimination in the labor market as it would explain lower incomes for darker skin color people due to the lower educational level concentrated in poorer regions, which is actually one of the conclusions derived from this work. What it is left as unexplained in the model, might be the result of taste-based discrimination, which would be discrimination based on factors that I did not include in the model, variables like performance or work behavior.

To sum up, there are evidence of wage gaps within Mexican labor market ac-cording to skin color. This differences arise mainly because white-skinned groups are reaching higher education levels than dark-skinned people resulting in a higher income for the former group according to both estimation methods used in this anal-ysis. Nowadays, several techniques to discover possible discrimination arising in the market place are available; from Becker (1957) market discrimination coefficient to spot proportional differences between wage rates, to wage descomposition meth-ods suggested by Oaxaca-Blinder (Blinder, (1973); Oaxaca and Ransom, (1994)) for

(29)

Chapter 5. Conclusions 22 finding out any unexplained differentials caused by discrimination, and even the innovative use of vignette studies to dig into gender differences, for instance, in the labor market aimed to find possible discrimination indicators (Deschacht, De Pauw, and Baert, 2017). Hence, a possible extension of this work is to choose a different pathway with respect to the theoretical and technical approach of the estimations, as almost no statistical work dealing with discrimination based on skin color has been done in Mexico, there is plenty of room for further research.

Another intersting point to describe discrimination in the market would be to consider the perspective of the employer. While in this work we focused in what the consequences for employees are, it is also useful to gather information in or-der to find if there is indeed discrimination by companies, different techniques and methods should also be used.

(30)

23

Appendix A

Information about Variables and

Models

TABLEA.1: Definition of Variables: Data Source, National Survey on

Discrmination in Mexico (ENADIS) Variable Definition

Dark Dark 1 if respondent has dark skin with a dark tone, 0 otherwise. Dark Medium 1 if respondent has dark skin with a medium tone, 0 otherwise. Dark Light 1 if respondent has dark skin with a light tone, 0 otherwise. Medium Dark 1 if respondent has medium skin with a dark tone, 0 otherwise. Medium Medium 1 if respondent has medium skin with a medium tone, 0 otherwise. Medium Light 1 if respondent has medium skin with a light tone, 0 otherwise. White Dark 1 if respondent has white skin with a dark tone, 0 otherwise. White Medium 1 if respondent has white skin with a medium tone, 0 otherwise. White Light 1 if respondent has white skin with a light tone, 0 otherwise. North

1 if respondent comes from Baja California, Baja California Sur, Chihuahua, Durango, Sinaloa, Sonora, Coahuila, Nuevo León or Tamaulipas States, 0 otherwise.

Center

1 if respondent comes from Aguascalientes, Guanajuato, Querétaro, San Luis Potosí, Zacatecas, Mexico City, Mexico or Morelos State, 0 otherwise.

East 1 if respondent comes from Hidalgo, Puebla, Tlaxcala or Veracruz States, 0 otherwise.

West 1 if respondent comes from Colima, Jalisco, Michoacán or Nayarit States, 0 otherwise.

South 1 if respondent comes from Chiapas, Guerrero, Oaxaca, Campeche, Quintana Roo, Tabasco or Yucatán States, 0 otherwise.

Married 1 if respondent is married, 0 otherwise. Gender 1 if respondent is a woman, 0 otherwise.

Migrant 1 if respondent comes from any place outside Mexican Republic, 0 otherwise.

Catholic 1 if respondent identifies as Catholic, 0 otherwise. Family Members Number of persons living at the house when interview. Indigenous Language 1 if respondent speaks and indigenous language, 0 otherwise. Age Respondent age at survey date.

Type of Work 1 if respondent does "outdoors" activities for living (fishing, haunting, farming), 0 otherwise.

(31)

Appendix A. Information about Variables and Models 24

TABLE A.2: Regression Analysis for Respondents Aged 18-70 with

Positive Wages in 2010. Dummy Variables for State Included

Dependent Variable:

log wage A.3

Constant 7.533 (.1166) DarkDark -0.0203 (.0692) DarkMedium -0.1055*** (.0564) DarkLight -0.0779 (.0566) MediumDark -0.0133 (.0624) MediumMedium -0.0873 (.0530) MediumLight -0.0581 (.0522) WhiteDark -0.0606 (.0425) WhieMedium 0.0549 (.0552) Married 0.1096* (.0263) Gender -0.0971* (.0299) Migrant -0.0103 (.0360) Catholic 0.0225 (.0457) Family Members 0.0342* (.0091) Indigenous Language -0.1511** (.0699) Age 0.0034 (.0011) Type of Work -0.3461 (.0564) Schooling Years 0.0685 (.0032) N 10,322 Adjusted R Squared 0.3131

Note: Coefficient estimates using OLS are reported and robust standard errors are shown in parentheses. Variables are described in Appendix Table 1. *, **, *** indicates signifi-cance at the 1, 5 and 10 percent levels, respectively.

(32)

TABLE A.3: Regression Analysis for Respondents Aged 18-70 with Positive Wages in 2010. Non-Light Skinned Groups Collapsed

Dependent Variable:

log wage (A.2) Constant 7.6450* (.0875) Non-Light-Skinned -0.0551** (.0267) Center -0.1519* (.0320) East -0.4336* (.0405) West -0.0186 (.0358) South -0.3615* (.0419) Married 0.0926* (.0273) Gender -0.1016* (.0302) Migrant -0.0016 (.0345) Catholic -0.0255 (.0456) Family Members 0.0367* (.0092) Indigenous Language -0.2672* (.0751) Age 0.0037* (.0011) Work -0.3979* (.0586) Schooling Years 0.0725* (.0033)

Note: Coefficient estimates using OLS are reported and robust standard errors are shown in parentheses. Variables are described in Appendix Table 1. *, **, *** indicates signifi-cance at the 1, 5 and 10 percent levels, respectively.

(33)

FIGUREA.1: Average Income According to Skin Color By Region

FIGUREA.2: Faces Showed During Interview Process

(34)

27

Bibliography

Altonji, Joseph G and Rebecca M Blank (1999). “Race and gender in the labor mar-ket”. In: Handbook of labor economics 3, pp. 3143–3259.

Atkinson, AB, L Rainwater, and TM Smeeding ((1995)). Income Distribution in OECD Countries.k OECD Social Policy Studies.

Baron, Reuben M and David A Kenny (1986). “The moderator–mediator variable distinction in social psychological research: Conceptual, strategic, and statistical considerations.” In: Journal of personality and social psychology 51.6, p. 1173. Becker, Gary S (1957). The economics of discrimination. University of Chicago press. Black, Dan A (1995). “Discrimination in an equilibrium search model”. In: Journal of

labor Economics 13.2, pp. 309–334.

Blinder, Alan S ((1973)). “Wage discrimination: reduced form and structural esti-mates”. In: Journal of Human resources, pp. 436–455.

Borjas, George J and Stephen G Bronars (1989). “Consumer discrimination and self-employment”. In: Journal of political economy 97.3, pp. 581–605.

Cain, Glen G (1986). “The economic analysis of labor market discrimination: A sur-vey”. In: Handbook of labor economics 1, pp. 693–785.

Darity Jr, William A, Jason Dietrich, and Darrick Hamilton (2005). “Bleach in the rainbow: Latin ethnicity and preference for whiteness”. In: Transforming Anthro-pology 13.2, pp. 103–109.

Dehejia, Rajeev H and Sadek Wahba (1999). “Causal effects in nonexperimental stud-ies: Reevaluating the evaluation of training programs”. In: Journal of the American statistical Association 94.448, pp. 1053–1062.

Deschacht, Nick, Ann-Sophie De Pauw, and Stijn Baert (2017). “Do gender differ-ences in career aspirations contribute to Sticky Floors?” In: International Journal of Manpower 38.4.

Dewey, Donald (1952). “Negro employment in southern industry”. In: Journal of Po-litical Economy 60.4, pp. 279–293.

Diario Oficial, de la Federación (1917). “Constitución Política de los Estados Unidos Mexicanos”. In: México, recuperado de http://www. diputados. gob. mx/LeyesBiblio/doc/1. doc (2007, 27 de julio).

Duckitt, John H (1992). “Psychology and prejudice: A historical analysis and inte-grative framework.” In: American Psychologist 47.10, p. 1182.

Edelman, Benjamin G and Michael Luca ((2014)). “Digital discrimination: The case of Airbnb. com”. In:

Espino, Rodolfo and Michael M Franz (2002). “Latino phenotypic discrimination re-visited: The impact of skin color on occupational status”. In: Social Science Quar-terly 83.2, pp. 612–623.

Esquivel Hernández, Gerardo (2015). Extreme Inequality in Mexico. Concentration of Economic and Political Power.

Goldsmith, Arthur H, Darrick Hamilton, and William Darity (2007). “From dark to light: Skin color and wages among African-Americans”. In: Journal of Human Re-sources 42.4, pp. 701–738.

Wage gaps according to skin color within Mexican labor market

Master’s Thesis