Combining Micro Data Sets to Study Immigration from Mexico to the US.

Transcription

1 Combining Micro Data Sets to Study Immigration from Mexico to the US. Alfredo Cuecuecha Abstract Current empirical studies on immigration of Mexicans to the US have failed to obtain whether Mexicans are positively or negatively selected from the Education distribution from Mexico, either due to lack of representative data or because of the use of wrong econometric techniques. This paper uses consistent and efficient methods to combine two micro data sets representative of Mexicans living both in the US and Mexico: the Current Population Service and the Encuesta Nacional de Ingreso de los Hogares. The sample includes males between the ages of 25 to 65 years old, which are in the labor force, or are discouraged workers, and unpaid familymembers. TheUSsampleonlyincludes individuals that enter the US at 15 years old or more. By selecting this sample, the study focuses in individuals that are less likely to have obtained their education in the US, after immigrating, and that migrated to obtain a job, even if they end up being unemployed. The results show that there is a positive relation between education and migration between 0 and 11 years of education. After which, the relation became negative. These findings reject the notion of negative selection, and support the idea that individuals that migrate from Mexico have medium levels of education by Mexican standards, but still low levels of education by US standards. 1 Introduction Despite the importance of the Mexican immigration to the US 1, basic information about the individuals entering the US from Mexico has not been clearly Many thanks to Kevin Lang for his dedicated advises in the preparation of this paper. The author also thanks Dilip Mookherjee, John Harris, Bob Lucas, Claudia Olivetti, Victor Aguirregabiria, Oriana Bandiera, Todd Idson, Rasmuz Lentz and all the participants in the Monday Micro Lunch. Any mistakes are my own. 1 Mexicans represent an important fraction of the individuals entering both legally and illegally to the US every year. In terms of legal immigrants, data from the INS reveals that almost 1

2 identified in the literature. The main reason for this failure in the literature has been the lack of data representative of the entire Mexican population, or the use of inadequate techniques to combine data sets from Mexico and the US. Early studies of the educational characteristics of the Mexicans relied on data not representative of the entire population moving into the US, and concluded that Mexicans were negatively selected from the education distribution 2.Thisnotion has been recently challenged by Chiquiar and Hanson (2002). Combining samples from the 1990 US and Mexican censuses, Hanson and Chiquiar concluded that while Mexicans in the US are certainly below the average level of education of the Native population, they have an average level of education greater than the population in Mexico. Furthermore, they claim that the group of Mexicans more over represented in the US is comprised by individuals with college education or more. There are three main problems in the work of Chiquiar and Hanson, that make their conclusions suspicious. The first of them is the possibility that Mexicans upon entering in the US may acquire education, the second is that the two data bases have different sampling rates, and the third is the inclusion of women in their sample. Chiquiar and Hanson attempted to avoid problems with education by focusing in individuals that were between 25 and 65 years old at the moment of the censuses. However, what is really important in determining the type of Mexicans migrating to the US is the age at which the individuals migrated to the US. If the individuals migrated at early ages, it is likely that their education was acquired in the US and not in Mexico. Consequently, their study would wrongly conclude that educated Mexicans are migrating to the US. This problem seems to be present, since Chiquiar and Hanson argue that high wage earners are more likely to migrate to the US. It is plausible that individuals that acquired education in the US, will also become bilingual individuals, develop networks in theus,andsimplybecomemoreassimilatedtotheamericanjobmarket,which may plausible make them a group of high wage earners, among the population of Mexicans. To understand the intuition for the need of an adjustment for different sampling rates, consider the following example. If the US sample of the census has 1000 Mexicans, while the sample of the Mexican census has 2000 individuals, we will conclude that the migration rate equals 30% of the Mexican population, when it could be the case that the true migration rate is just 10%. Finally, including women in the sample adds the problem that some women may decide to migrate not by choice, but simply to follow their spouses in the US, besides the 28% of all legal immigrants during the 1989 to 2001 period were nationals from Mexico (INS, 2001). As for illegal immigration, the INS estimates that in the year 2000, 69 % of all the illegal immigrants are Mexicans(INS, 2003). 2 The study of Taylor (1986), used data from rural villages in the central part of Mexico. Later studies have attempted to increase the number of villages and regions, but still lack representativity of the entire country (see for example the work of Massey and Espinosa (1997) ). 2

3 usual complications brought by fertility, and labor market participation decisions. In this paper, we will not attempt to model the decision of women, and we will excludethemfromthesample. This paper combines data from the Encuesta Nacional de Ingreso y Gasto de los Hogares (ENIGH) and the Current Population Survey (CPS) for the year 1994, which is the first year in which the CPS increased the sample of Hispanic households, and collected information on the country of origin of individuals, as well as on the date in which individuals entered the US. Unlike other papers about immigrants, in this paper we will apply Choice Based Sampling techniques, which help to correct the problems of the endogenous sample, and the different sampling rates on each side of the study. Combining micro-data sets without worrying about the differences in sample rates is not something particular to the Chiquiar and Hanson study, other authors have also ignored the problem altogether ( see for example Bauer et al, 1998). Other studies have developed their own corrections to the CBS problem ( see Bover and Arellano, 2002). This paper will apply econometric techniques that provide with consistent (Manski and Lerman, 1977) and efficient estimators (Imbens, 1992). The sample of the paper is comprised by males in the labor force, and discouraged workers, which were 25 to 65 years old in 1994, and that entered the US when they were at least 15 year old. By doing so, the paper focuses in individuals entering the US in a situation in which they were less likely to continue their education in the US, and in individuals that their choice of country mainly represents a choice of a place to work, even if they may end up being unemployed in any of the two countries. The results suggest that there is a positive relation between migration and education from 0 to 11 years of education, after which the relation became negative. Consequently, our results reject negative selection and agree on that respect with Chiquiar and Hanson s results. However, our result reject their claim that the 12 to 15 years of education group is the one that migrates the most, since our results suggest a negative relation between education and migration in that range of education years. The differences in results are due to two main reasons: first, unlike Chiquiar and Hanson, we use here CBS techniques that take into account differences in sampling rates in the two countries, as well as the endogeneity of the sample. Second, we are using the CPS and ENIGH and not the censuses of the two countries. Determining the nature of the selection in educational characteristics of individuals from Mexico is important for several reasons. First of all, the finding that Mexicans were negatively selected from the education distribution contradicted thenotionthatimmigrantswouldtendtobeindividualsmoremotivatedandpositively selected from the distribution of characteristics in their countries (Chiswick, 1986). The negative selection hypothesis (Borjas, 1987) that was elaborated to explain such findings, has had important implications both in the economic literature of immigration, and in the making of immigration policy analysis and 3

4 recommendations. Consequently, finding that the selection proceeds from the middle of the education distribution may have also important implications in terms of motivating new research in the economics of immigration, and plausibly in generating better knowledge for the making of immigration policy. The results of this study clearly show that the relation between education and migration for all individuals with less than high school education is positive, which is consistent with the notion that Mexicans come to the US to fill jobs that no one in the US wants to perform (Piore, 1979). On the other hand, the results of this paper suggest that among individuals with more than high school, those more educated are less likely to migrate. This result shows that while immigration of highly skilled Mexicans, certainly occur, on average individuals highly educated are less likely to migrate to the US. These findings imply that the net benefits of migration peak at medium levels of education. This information reveals that Borjas (1987) assumptions needed to derive negative selection, about the existence of migratory costs constant for all individuals are probably incorrect. However, there is not enough information to determine whether migratory costs are decreasing on education, or benefits to migrate are decreasing on education, or whether both costs and benefits to migrate are decreasing on education, as Chiquiar and Hanson (2002) argue to explain their results. The determination of this issue, however, is beyond the reach of this study 3. The paper is organized as follows: the first section presents the empirical model used to study the characteristics of the migrants, and the techniques to be implemented to obtain efficient and consistent estimators. The second section introduces the data, explains the sample, and shows evidence that some Mexicans increase their schooling years upon arrival to the US. The section also explains the source of the information used in the CBS techniques. The third section presents the results of the estimations for the characteristics of the immigrants. The fourth section concludes. Before going into the details of the paper, it is important to mention two issues regarding the data existing on Mexican immigrants. First, an important fraction of the immigration of Mexicans to the US is illegal. To date, no existing data set can have a complete description of the Mexicans in the US (Bean, et al, 1997). However, we will provide in section two with some evidence and arguments about why the data being used in this study can be said to do a good job in following the stock of Mexicans living in the US. Second, the ideal data set to study immigration from Mexico would be one with detailed labor force histories of a representative sample of individuals at both countries. Nowadays, such type of data set is non-existent, although the data set collected by the Mexican Migration Project, in the Sociology Department of the University of Pennsylvania, is perhaps the only to provide with a detailed history of employment of Mexicans 3 Cuecuecha (2003), analysis why Mexicans with medium levels of education migrate to the US. 4

5 at both countries. Unfortunately, the MMP data set is not representative of the entire Mexican population, which lead us to settle in using two cross sections that are representative of the Mexican population living in the US and Mexico. 2 The Characteristics of the Immigrants The need for Choice Based Sampling Techniques ThestudyofthecharacteristicsoftheimmigrantsfromMexicototheUS,could be performed with a simple probit (logit) analysis, should the sample be random. However, in the context of choice based samples, the probit regression will be biased and inconsistent (Manski and Lerman, 1977) because the sampling scheme isnotentirelyrandom. Thesampleathandissaidtobechoicebased,because all individuals found in the US sample have chosen to migrate, and because the individuals are being sampled at different rates depending on the country in which they are located. Let us define the probability of choosing the US, conditional on the vector of characteristics X, P (US =1/X), as: P (US =1/X) =P (δ X) Where δ measures the contribution of characteristic X i to the score function δ X. Let us define h as the frequency with which individuals in the US are being sampled, and q as the true migration rate. If R(X) is the cdf of the characteristics of individuals, q = P (δ X) dr(x). Note that if the sample would have been random, this should be all the information we would need to construct the likelihood function. Let us now define what is the probability of finding a Mexican in the US, given their characteristics X: where P (θ X) q g(us,x)=h P (δ X) q represents the conditional probability of finding a Mexican in the US, conditional on being in the US sample. The above equation neatly summarizes the reason for the Probit to be biased and inconsistent: since the frequency rates can be different from the propensity with which the individuals choose to migrate, using a probit that does not take into account the CBS nature of the sample would specify wrongly the likelihood for each observation WESML Estimation Manski and Lerman (1977) noted that if the pair (h, q)was known we could multiplyeachobservation foundintheusbytheq/h ratio, and all the observations in Mexico by the (1 q)/(1 h) ratio. Their estimator is called the Weighted Exogenous Sampling Maximum Likelihood estimator. The likelihood function for the WESML estimator would then be equal to: 5

6 L WESML (θ) = i I (US) q h ln P (δ X i )+I (MX) 1 q 1 h ln (1 P (δ X i )) Where I (j) is an indicator variable for country j=us, MX GMM Estimation The above estimator is consistent, but inefficient and other estimators have been proposed since then. A second concern in the literature is about the fact that aperfectknowledgeof(h, q)may not be available, and that only partial or no information about them may exist. The estimator proposed by Imbens (1992) have the important characteristic that the knowledge of (h, q)is not necessary to develop the method, since in the case they are partially known or unknown, they can be treated as parameters to be estimated in the model. Furthermore, the method proposed is both consistent and efficient. Imbens proposes to use three moment conditions given by the following equations 4 : Ψ 1 (j)=h(j) I (j) Ψ 2 (j)=q(j) P (j/x,δ) Ψ 3 (j)=p δ (j/x,δ) 1 P (j/x,δ) P δ (δ X i ) h 1 h q 1 q where = hp q (δ X i )+ 1 h [1 P 1 q (δ X i )] and the index j represents the fact that for observations in the US, the sampling frequency is h, while for observations in Mexico the samplings frequency is 1 h. Similarly, q(us) represents the migration rate, while q (MX) represents the complement of the migration rate. Likewise, P (US/X,δ) represents the conditional probability of migration for individuals in the US sample, while P (MX/X,δ) represents the complement of the migration probability in the Mexican sample. Finally, P δ (j/x,δ) represents the gradient vector of the probability ofbeingobservedincountryj. Note that the second term in the Ψ 3 vector is similar for observations in the two sides of the sample. These moments have the following interpretation when evaluated at the true migration rate q, and the true sampling frequency h :thefirst moment Ψ 1 (j)simply obtains the sampling frequency, which asymptotically must be equal to the true 4 For a detailed explanation of the obtention of these three moments, the reader is referred to Imbens (1992). 6

7 h (j). 5 The second moment Ψ 2 (j)compares the marginal probability with the average of the conditional probabilities, each one of them weighted by the bias induced by the sampling design. The third moment Ψ 3 (j)represents the score of the conditional likelihood of choice j,givenx. The vector of parameters to estimate in the model is δ =[h, q, δ ] which is defined as follows: R n (δ )= 1 n i δ =argminr n (δ ) Ψ(δ,x i,us) C n 1 n i Ψ(δ,x i,us) Finally, the optimal weighting matrix C n is estimated to be: Define C n = 1 n i Ψ(δ,x i,us)ψ(δ,x i,us) 1 and 0 = EΨ(δ,x i,us)ψ(δ,x i,us) Γ 0 = E Ψ(δ,x i,us) δ Given the assumption of a probit and other regularity conditions 6,Imbens proves that δ converges almost surely to a Normal random vector with distribution: n 1/2 δ δ N 0, Γ Γ The intuition to why it should be included in the moment vector even when the true sampling frequency is known is that using sampling frequencies instead of the true frequency can increase efficiency, a result found by Cosslett(1981) in the context of the Conditional Maximum Likelihood Estimator proposed for Choice Based Samples. For other illuminating example found in the context of SUR, see footnote 3 in Imbens(1991:1199). 6 Imbens requires: 1.The space of observations be a subset of the real numbers, the choices be finite, the coefficients in the probability equation belong to the interior of a compact subset of the real numbers. 2. The probabilty function should be twice continuously differentiable with respect to the parameters, and the function and its two first derivatives should be continuous in X for all thethas. Furthermore, each choice has positive probability for all values of X and thetha in an open neighborhood of the true theta. 3.The minimizer is unique. The Deltha zero matrix is nonsingular. The Jacobian of the moment vector has full rank equal to K+2. The weigthing matrix converges almost surely to a positive semidefinite matrix. 7

8 3 Data In this paper, we will use complementary micro data sets that provide with measures of the number of Mexicans in the US and the Mexicans in Mexico. Most importantly, the two databases provide with information on education level, education years, age, year of entry to the US, and labor force status. The source of information for the immigrants in the US is the Current Population Survey (CPS), while the source of information for the non-migrant population is the Encuesta Nacional de Ingreso y Gasto de los Hogares (ENIGH). The sample to be included in the study comprises all males in the age range of 25 to 64 years old in The sample focuses in these age range to minimize the individuals who have retired, and also to focus attention in individuals who have finished their schooling. The sample comprises only males since it is possible that females may be tied migrants, i.e. they migrated to follow their husbands not by their own choice. The sample will include males in the labor force, and discouraged workers. By doing so, we want to focus attention to individuals that in 1994 chose their country of location motivated mainly with the purpose of finding a job. We will now discuss two potential issues that arise only in the CPS. 3.1 The CPS Undercounting, Occupations, and Geographic Locations The Current Population Survey is an almost ideal data set to study the stock of Mexicans because it constitutes a sample representative of the entire US population, and because, starting in 1994, the March CPS rounds increased the sample of Hispanic households, based on the November CPS results, to reduce the undercounting of the Hispanic population. Passel and Fix (2001) estimate that in 1995, the rate of undercounting in the CPS of illegal immigrants reached as most as 33%, while for the CPS of the year 2000, the rate of undercounting reduced to almost 7% because the CPS surveys may have been affected by the 2000 census outreach effort. To verify the severity of the undercounting, I calculated the proportion of Mexicans that have migrated to the US by combining the ENIGH and CPS with information from the Census Bureau of each country about the total size of the population. Basically, I obtain the proportion of males that were born between 1936 and 1969 for each of the micro data sets in 1994 and Ithen multiply this proportion by the total size of the male population in each country to obtain an estimate of the entire Mexican male population born between 1936 and By calculating the percentage of them that are in the US, I obtain a measure of the migration probability. Table 1.1 compares my estimation of the migration rate, with estimations that can be made using information from different sources. The total stock of 7 Note that the years analyzed in the CPS and ENIGH are 1994 and The reason is that the ENIGH is only available every two years, starting in

9 Mexicans in the US is estimated to have been 13.4 million in 1990, 16.1 million in 1995, and million in the year 2000 (Bean et al:2000). These numbers include the illegal immigrants that are estimated to live in the US. Regarding the number of illegal immigrants, Warren(2000) estimates that 1.85 million of Mexicans were unauthorized immigrants in Passel and Fix(2001) estimated that in 1995, 5 million unauthorized immigrants lived in the US, while the same figure for the 2000 is 7.0 million individuals. Combining this information with information on the population observed in Mexico, available from the Instituto Nacional de Estadistica Geografia e Informatica (INEGI), we can calculate the migration rate which is shown to be around 15% during the 1990 s, and pretty close to my estimation based on the CPS and the ENIGH, which is based only in the males 25 to 65 years old. 8 There are different explanations for this result. The first is that the CPS did a good coverage of the stock of Mexican living in the US, perhaps much better of what Passel and Fix estimated. Second, it is possible that the greater undercounting in the CPS is that of the individuals that enter seasonally to perform agricultural activities. If such group is constituted mainly by very young individuals, our focus in individuals which are at least 25 years old in 1994, make us to follow a group of individuals in which the rate of undercounting is small. To analyze this argument we review occupations and ages of the Mexican individuals in the 1994 and the 2000 CPS rounds, and compare to other sources of data. Table 1.2 shows that in our 1994 sample of individuals born between 1935 and 1969, aged in 1994 between 25 to 59 years old, 17.02% of individuals were engaged in such activities, while in the 2000 CPS round among the same cohort of individuals 17% are engaged in agriculture, fisheries and forestry. The figure for individuals born between 1968 and 1979, whose ages range between 15 to 24 years old in 1994 is 13.87%, while it is 17% in the 2000 CPS. Piore(1979) estimates that 17.2 % of all illegal apprehended immigrants in 1976 were in farming, fishery, and forestry activities. Recent estimates of the proportion of illegal immigrants engaged in farming, fisheries and forestry activities put that figure in about 19.1% for 1999 (CIS, 2001). It is obvious that a lot of time has occurred since Piore s estimation, but it is remarkable that the more recent figures are not quite far from that measure. Most importantly, the proportion obtained in the 1994 and 2000 CPS for the 1935 to 1969 cohort is also quite close to those reported in other sources. Note that the proportion for the younger cohort in the 1994 CPS 8 Concerning the flows of individuals going into the US, the INS 2000 yearbook shows that the average flow between was of about 415 thousand individuals legally entering the US, while for the flows are of about 135 thousand individuals. The illegal flow is also important: Passel and Fix(2001) put the size of the illegal flow in about 275 thousand individuals in the early 1990 s, while the same figure for the late 1990 s is of about 400 thousand individuals. On the other hand, the return flow is also important: Reyes(2002) shows that from those individuals that enter in 1995, 15% returned within six months; while for those entering in the year 2000, only 7 percent had returned. 9

10 is smaller, but such changes could be do simply to the fact that the sample size of 15 to 24 years old is considerably smaller, and not necessarily to a better counting of illegal immigrants in the 2000 CPS. In any case, it is reassuring that for the cohort followed in this study the miscounting seems to be less severe. We finally show the geographic location of the Mexican sample for both 1994 and 2000, using the census divisions (see Table 1.3). It is shown that almost 85% of Mexicans are spread over the Pacific, South and Mountain Census divisions in both 2000 and The majority of the Mexicans are located in the Pacific census division: almost 50% in 1994, and 43% in the year The reduction observed in the year 2000 in the relative number of Mexicans observed in the Pacific census division seems to be produced by an increase in the absolute number of individuals located in the Mountain census division during the year Measurement of Education A second potential problem with the use of the CPS is that if individuals migrate to the US at an early age and they get their education there, then we will conclude that more educated individuals come to the US, when in fact these individuals are migrating and finishing their education in the US. We calculate the potential years of education to analyze this problem. Potential years of education is defined as the total years of education that the individual attained plus five, minus the age at which the individual entered the US. We obtained the total years of education of the individual by assigning a given number of years to each of the levels of educationinwhichthecpsclassifies the education of the individuals (see Table 1.4). The data shows that 80% of the sample in 1994 had zero potential education. From those that potentially obtained education in the US, approximately half of them obtained up to six years of education, a quarter more obtained up to 12 years of education, the rest more than that. In fact, the average years of education for individuals with positive potential education in the US is 12 years, while for individuals with no potential education in the US is just 7.4 years. These results suggest that upon entering the US, some individuals still decide to continue their schooling. Note, however, that there are two other possible explanations for these results. The first relates to the citizenship status of the individuals, and the second refers to whether the individuals came to the US or were brought by their parents at an early age. If the individuals acquired their citizenship by their parents or relatives in the US, it is possible that they can have less restrictions to educate. If the individuals were brought to the US at an early age, it is also possible that they can have more opportunities to attend school. The mean years of education among individuals born in Mexico that are citizens of the US is 9.4, while for those non-citizens is of 8 years of education. Note that the difference between noncitizens and citizens is not that big, which probably could be explained if the individuals became citizens at a late age, if this is the case, then citizenship is obviously a choice made by the individuals. 10

11 The mean years of education among individuals who entered prior to 15 years old is 10 years, while the mean years of education among individuals who entered after that age is 8.4. These results suggest that the acquisition of education by immigrants occurs in some cases, and it also suggest that we can use whether the individual was brought to the US as a child to focus the sample to individuals that arrive in situations in which they are less likely to invest in education Labor Force Status, and Sample Table 1.5 shows how the entire number of observations can be classified according to their labor force status. The table shows the number of individuals excluded due to the fact of being out of the labor force, for either being students, disabled, or retired. The table also shows the number of individuals that are excluded because they enter the US before they were 15 years old. Finally, we can observe that the sample comprises 1039 individuals in From which, 128 individuals are unemployed. The unemployed individuals include individuals looking actively for a work, unpaid workers, and individuals who are unemployed that give up searching. The observed unemployment is of 12.31%. Table 1.5 also shows that the US sample of Mexicans has that the average individuals has 8 years of education, it is 36 years old, earns 14.5 thousand dollars per year,works 41 hours per week, and has 13 years of being in the US. 3.2 The ENIGH Beginning in 1992, the Encuesta Nacional de Ingreso y Gasto de los Hogares (ENIGH) started to be done every two years, providing with a data set that by definition provides with micro-data of the individuals that chose to stay in Mexico, unfortunately this data set does not include information on previous migratory behavior or current household members living in the US Education The education information on the 1994 ENIGH provides with a qualitative variable that classifies education of the individuals. Table 1.6 shows the correspondence between those education levels and the years of education used in the continuous variable. The table also shows the equivalencies between the Mexican levels of education and the grades that are obtained in the US system Sample As with the CPS, we will exclude from the sample disabled, students, and retired individuals, and the sample includes only males between 25 and 65 years old. Unemployment will also be measured as comprised of all individuals looking for a job, working unpaid, or that give up searching for a job. Table 1.7 shows that the total number of observations in the Mexican sample is of individuals, fromwhich625areunemployed. Theobservedunemploymentrateisof4.36%. The average individual in the Mexican sample is 38 years old, has 6.4 years of 11

12 education, works 40 hours per week, and earns 11.8 thousand pesos per year. At the exchange rate of 1994, that is about 3.47 thousand dollars Occupation Table 1.8 shows the occupations of the individuals in the sample of ENIGH. 37 % of the sample is dedicated to agricultural activities, the rest is distributed between services, manufacturing, and professional occupations. 3.3 Weighting Both the ENIGH and the CPS sample individuals at different rates. However, such sample design is independent to the choice based sampling. Therefore, the different sample rates will only be used to determine how many number of individuals each observation represent. By using this information we can obtain the sample frequency in our North American sample. It is straightforward to show that the sampling frequency will be given by the following expression: h = i US factor i i MX factor i + i US factor i Where factor is the inflation factor for individual i. The intuition for this result is that the inflation factor for each individual represents the number of observations that each individual represents. The inflation factors for the CPS are the individual March supplement weights.. For the ENIGH, the inflation factors provided by the survey are at the household level and need to be adjusted to represent individual level sampling probability. The adjustment is done by multiplying the number of individuals in each household per the household inflation factor, and then dividing this number by the total number of individuals in the Mexican sample. The number just found is the probability of sampling individual i, and hence the inverse of the individual inflation factor: factor i = factor hi N hi hi MX factor hi N hi where the factor hi represents the household level inflation factor, and N hi represents the number of household members in the concern household. Note that the inflation factor is the same for each household, since conditional on belonging to sampled household hi, the probability of being sampled is one. 3.4 The Combined Sample Before analyzing the results of the estimation, table 1.9 provides with the mean age and education for the combined sample. The sample has observations, mean age of 38.7 years, and mean education of 6.45 years. 1 12

13 4 Estimation 4.1 Results for the Migration Choice Table 2.1 shows the results for the estimation of the characteristics of the migrants. Column1 shows the results from using a probit without the Manski and Lerman (1977) correction, while column 2 shows the results of applying the Manski Lerman correction, both estimations include all individuals in the sample. The parameters of the first are known to be biased, so comparing them to the WESML coefficients show us the nature of the bias in the coefficients. The table shows that the coefficients of the Probit are greater than or equal to the coefficients of the WESML estimation. The two linear terms in age and education arelargerundertheprobitthanusingthewesml.bothestimationspredicta positive relation between education and migration from zero years of education up to 11 years of education, although the unweighted predicts more migration propensity below years of education, and consequently a lower marginal effect of education at the sample mean (see Table 2.2). We now turn to compare the estimations of the WESML with those of a logit estimation without the Manski Lerman correction. The logit estimations predicts that the positive relation between education and migration occurs between 0 and 9.8 years of education. Not surprisingly, the marginal effect of education is also estimated to be smaller using the logit estimation. In any case, unweighted probit estimations seem to overstate the effect of education on immigration. Finally, the GMM estimations converged to the parameters obtained with WESML, although the precision of the estimation increased considerably. In particular, the GMM estimation suggests that experience increases the probability of migration between 25 and 27.7 years of education, after which the relation is negative. So, the more experienced individuals will be less likely to migrate. On the other hand, the GMM estimation confirmedthatbetween0and11years of education, there is a positive relation between education and immigration. 4.2 Implications of Findings Figure 1 shows the predicted migration probability for the sample divided in threecohorts: 25to34yearsold,35-44yearsold,and45to54years. Consider the cohort aged 25 to 34 years old. At any point in the curve, the individuals from this cohort are more likely to migrate than individuals in the cohort, and the cohort. Similarly, the cohort is also more likely to migrate than the oldest cohort. Now within any one of the given cohorts, the individuals above the mean of education in Mexico (6.5 years of education) and 16 years of education constitute the group with the largest migration probability. Within this group of individuals, those with 11 years of education are at the highest risk of migration. This figure summarizes the results of this paper, since they confirm the rejection of the negative selection hypothesis, and the inverted U relation 13

14 between education and migration. What explains the existence of this inverted U shaped relation?. Let us first analyze why migratory costs can be decreasing on education, and why decreasing migratory costs in education could explain these findings. Migratory costs can be decreasing in education if more educated people may be more able to learn English, if they are more able to understand immigration regulations, or if they are able to assimilate faster to the American culture. On the other hand, immigration costs may actually reduce with education if the absolute number of Mexicans with high levels of education is lower than the absolute number of Mexicans with low levels of education, and social network effects exist. Consequently, depending on the level of education in which the larger absolute number of migrants exist, we can have that the social network effect can actually produce a relation between education and migratory costs that will actually be a U-shape form. In fact, if migratory costs have such U form, with minimum at 11 years of education, and benefits to migrate are uncorrelated with education, we can observe that net benefits to migrate will have a maximum at 11 years of education. Note that social networks are not the only explanation as to why migratory costs may increase at high levels of education. After all, among the migratory costs is the opportunity cost of returning to Mexico. If Mexicans are target savers, those obtaining larger wages may in fact see a rise on the value of being in Mexico. Consider now a second possibility: migratory costs are uncorrelated to education, and migratory benefits increase with little levels of education and later reduce at high levels of education. It is easy to understand why benefits may be increasing on education: if return to education in the US is larger than in Mexico, that would be enough. On the other hand, the benefits of migration can be decreasing on education for different reasons. First, Mexicans may lack the language skills to obtain a skilled job. Second, they may lack the networks to enter the skilled sector. Third, employers may face a disadvantage in hiring some one from Mexico due to immigration regulations. Fourth, Mexicans may be discriminated in the skilled sector. In any case, if benefits to immigrate have an inverted U-shape and migratory costs are uncorrelated to education, we would observe that net benefits to immigrate are maximized at 11 years of education. As the previous discussion made clear, the findings do not provide enough information to contrast between the two possibilities. All we can tell by sure, is that the net benefit to immigrate has also a U inverted relation with education. Note that this distinction is important, since they have different implications regarding the nature of the labor market that Mexicans face in the US, the reasons for Mexicans to immigrate to the US, as well as on the evolution over timeoftheimmigrantsfrommexicointheus. 14

15 Figure 1: Results Using WESML 5 Conclusion This paper combines two micro data sets from Mexico and the US that are representative of the entire population of Mexicans immigrating to the US. The main finding of the analysis is that between 0 and 11 years of education there is a positive relation between migration and education. This result rejects the negative selection hypothesis sustained by Borjas(1987) to explain the results from analysis obtained using data from Mexican villages. This finding also rejects the claims made by Chiquiar and Hanson (2002), that argue the more over represented group of Mexicans is that comprised by individuals with more than high school. We conclude that our findings imply that the net benefits to immigrate have an inverted U shape. Without further analysis into the nature of the wage differentials between the two countries, and other considerations that individuals may take into account in their decisions, like differences in unemployment probabilities between the two countries, existence of migrant networks, return migration patters, it is difficult to establish whether the form of the net benefits is explained by returns to migration in inverted U shape form, or by migration costs in U shape, or by some combination of these two possibilities. Further analysis is required to understand the reasons behind the inverted U shaped net migratory benefits, and the possible implications of these findings. 6 Bibliography 15

16 Bauer, Thomas; Pereira, Pedro; Vogler, Michael, and Klaus Zimmermann (1998) Portuguese Migrants in the German Labor Market: Performance and Self- Selection. The Institute for the Study of Labor (IZA). Discussion Paper 20, August. Bean, Frank D. et al (1999) The Quantification of Migration between Mexico and the United States. Mexico United States commission.. Borjas, George (1987) Self Selection and the Earnings of Immigrants. The American Economic Review. 77(4): Bover, Olympia and Manuel Arellano (2002) Learning about Migration Decisions from the Migrants: Using Complementary Datasets to Model Intra-Regional Migrations in Spain. Journal of Population Economics, 15: Chiswick, Barry R. (1986) Human Capital and the Labor Market Adjustment of Immigrants: Testing Alternative Hypothesis. in Research in Human Capital and Development: Migration, Human Capital and Development. Vol. 4: JAI press. Greenwich, Connecticut. Cosslett, Stephen R. (1981) Maximum Likelihood Estimator for Choice- Based Samples Econometrica, 49 (5): Cuecuecha, Alfredo (2003) Why Mexicans with Medium Levels of Education Migrate to the US? Working Paper. Boston University. Hanson, Gordon H. and Chiquiar Daniel (2002) International Migration, Self Selection, and the Distribution of Wages. Working Paper. University of California San Diego. Imbens, Guido W. (1992) An Efficient Method of Moments Estimator for Discrete Choice Based Models with Choice-Based Sampling. Econometrica. 60 (5) : Manski, Charles F. and Steve R. Lerman (1977) The Estimation of Choice Probabilities from Choice Based Samples. Econometrica, 45 (8): Massey, D.S. and K. Espinosa (1997) What s Driving Mexico-US Migration? A Theoretical, Empirical, and Policy Analysis, American Journal of Sociology, 102 (4): Passel, Jeffrey S. and Michael F. Fix (2001), US Immigration at the Beginning of the 21st Century. Testimony Before the Subcommittee on Immigration and Claims Hearing on The US Population and Immigration, Committee on the Judiciary, US House of Representatives, Urban Institute, Washington, DC. Piore, Michael (1979) Birds of Passage. Cambridge University Press. New York, NY. Reyes, Belinda I.; Johnson, H.P., and Richard Van Swaeringen (2002). Holding the Line? The Effect of the Recent Border Build-up on Unauthorized Migration. Public Policy Institute of California, San Francisco, CA. Taylor, J. Edward (1986) Differential Migration, Networks, Information and Risk. in Research in Human Capital and Development: Migration, Human Capital and Development. Vol.4: JAIpress.Greenwich,Connecticut. 16

17 Woodrow-Lafield, Karen (1998) Estimating Authorized Emigration, in Migration Between Mexico and the United States Vol. 1. Mexico United States Binational Migration Study. Center for Immigration Studies (CIS, 2001) Labor Market Characteristics of Mexican Immigrants in the United States. /2001/mexico/labor.html Immigration and Naturalization Service (INS, 2003) Estimates of the Unauthorized Immigrant Population Residing in the United States: 1990 to Executive Summary. Immigration and Naturalization Service (INS, 2001) Immigrants Admitted by Region and Country of Birth: fiscal years Statistical Yearbook of the Immigration and Naturalization Service. IMM2001list.htm 7 Tables Table 1.1 Aggregate Migration Rates Location: Mexico a US b Illegal c Migration Rate to 65 Mexico d e US d 2.56 e 3.13 Migration Rate 14.7 e 17.3 Source: a Inegi, b Bean et al (2000), c Various Sources d own calculations with data from CPS and ENIGH. e 1994 data. 17

18 Table 1.2 Occupations Professional Administrative Service Farmers Prodn & Repair Laborers Professional Administrative Service Farmers Prodn & repair Laborers Table 1.3 Census Divisions NE, Middle At., NC South, Mountain Pacific Table 1.4 Education Years Conversion Level Used in CPS Year used Less than first grade 1 1st to 4th grade 2.5 5th or 6th 5.5 7th or 8th 7.5 9th grade 9 10th 10 11th 11 12th 12 High School 12 College no degree 14.5 Associate Degree 14 Bachelor Degree 16 Master Degree 17.5 Professional Degree 20 Doctorate Degree 21 18

19 Table 1.5 US sample 1994 Individuals 1039 Unemployed 128 Education years 8.11 Age Wages (dollars, 1994) Hours of Work Years in US Table 1.6 Education Years Conversion Level Used in ENIGH Year used No education 0 Incomplete Primary 3 Primary 6 Incomplete Secondary 7.5 Secondary 9 Incomplete Preparatoria 10.5 Preparatoria 12 College no degree 14 College 16 More than College 18.5 Equivalencies Grades Primary 1to6th Secondary 7th to 9th Preparatoria 10th to 12th Table 1.7 Mexico sample 1994 Individuals Unemployed 625 Non-Reporting 46 Education years 6.47 Age Wages (pesos, 1994) Hours of Work

20 Table 1.8 Occupations Mexico Sample Professional Administrative & Sales Service 9.09 Farmers Prodn & Repair & Laborer Table 1.9 Combined Sample Age 38.7 Education 6.45 N Table 2.1: Characteristics of the Migrants Probit Logit Unweighted WESML GMM Unweighted N Age ***.0177 (.015) (.010) (.0005) (.0329) Age *** (.0001) (.0001) ( ) (.0004) Education.1438***.0806***.0806***.2700*** (.012) (.008) (.00069) (.0294) Education *** *** *** *** (.0007) (.0004) (.00004) (.0016) Turning Pt pseudor Likelihood Significance level: *1%, **5%,*?10%,?20% 20

21 Table 2.2: Migration Choice Probit Estimation Characterictics of the Migrants, CPS and ENIGH 1994 Marginal Effects Unweigthed WESML Logit N Age (.0019) (.003) () Age (.00002) (.00004) () Education.0174***.0345***.0011*** (.0015) (.002) () Education *** -.001*** *** (.00008) (.001) () Migration Rate Predicted Rate Significance level: *1%, **5%,*?10%,?20% 21