Immigration and the Occupational Distribution of Natives: a Factor Proportions Approach

Transcription

1 Immigration and the Occupational Distribution of Natives: a Factor Proportions Approach Javier Ortega City University London, CEP (LSE), CReAM, and IZA Gregory Verdugo Banque de France and IZA March 22, 2012 Abstract We study how immigration impacts the labor market outcomes and the occupational distribution of natives similar to immigrants in terms of their education/experience levels. Combining large (up to 25%) extracts of five French censuses and data from Labor Force Surveys for , we first show that immigration has a positive and significant impact on the wages and employment of natives, both at the national and at the regional level. We then perform a decomposition of the evolution of wages within education/experience cells, and show that an important part of the positive relation between immigration and wages comes from a reallocation of natives to better-paid occupations within the cells. Finally, we show that a higher presence of immigrants in the cell is associated with (i) a more dissimilar within-cell distribution of natives vs immigrants across three-digit occupations, and (ii) natives within the cell performing more abstract tasks relative to immigrants. JEL classification: J15, J31 We thank Herbert Brücker, Denis Fougère, Arnaud Lefranc, Claudio Michelacci, Núria Rodríguez-Planas, Patrick Sevestre, and seminar participants at the Banque de France, CEP (LSE), Université de Lille, IFN (Stockholm), LAGV 2010 (Marseille), EALE-SOLE 2010, IV Inside Conference (Barcelona), and Cergy for useful comments and helpful discussions. We also thank INSEE and the Centre Maurice Halbwachs (CMH) -in particular Alexandre Kych- for giving us access to the data and for their help. The data used in this paper can be accessed through the CMH. Javier Ortega is also affiliated to FEDEA. This paper does not necessarily reflect the views of the Banque de France. Department of Economics, School of Social Sciences, City University London, Northampton Square, UK- London EC1V 0HB, javier.ortega.1@city.ac.uk Service des Analyses Microéconomiques, Direction des Etudes Microéconomiques et Structurelles, Banque de France, 31 rue Croix-des-petits-champs, Paris Cedex 01, France. gregory.verdugo@banquefrance.fr 1

2 Introduction Recent years have seen a renewed interest in the research on the impact of immigration on the labor market outcomes of natives. Until the 1990s, the methodology exploited the variation in the share of immigrants across geographical locations. 1 After Borjas, Freeman and Katz (1997) underlined some potential problems of this approach, 2 a new strand of literature has proposed to measure the impact of immigration by relating the variation over time in the number of specific groups of immigrants with the outcomes of the natives with similar characteristics. The dimension that has attracted most attention is the education/experience dimension, as proposed initially by Borjas (2003), which assumes that natives and immigrants within education/experience cells are perfect substitutes, and uses the variation in the immigrant share at the cell level over time to identify the impact of immigration. 3 Using this approach, Borjas (2003) finds a large negative impact of immigration on natives wages. 4 A series of papers have suggested different modifications or refinements of Borjas (2003) keeping the analysis along the education/experience dimension. In particular, Card (2009) argues that, in the case of the U.S., a model with four groups of education (high-school dropouts, high-school graduates, individuals with some college, and college graduates) as in Borjas (2003) or Borjas and Katz (2007) does not fit well the data. In addition, Manacorda, Manning and Wadsworth (2012) for the U.K. and Ottaviano and Peri (2012) for the U.S. estimate a multi-level CES production function and conclude that natives and immigrants within education/experience cells are imperfect 1 See e.g. Card (1990), Altonji and Card (1991) or Hunt (1992). The consensus was that the effect of immigration on natives was small, see for instance Friedberg and Hunt (1995) or Borjas (1994). 2 Borjas et al. (1997) argues that this approach may understate the impact of immigration because (i) natives may respond to immigrant inflows by moving out to other locations, which would diffuse the impact of migration across locations but would not be captured in a spatial correlation approach, and (ii) immigrants may choose the best locations, which would result in reverse causality. For a paper treating these biases, see Pischke and Velling (1997). Recent evidence on the response of natives to immigrations flows is relatively mixed, see Card and DiNardo (2000), Card (2001), and Borjas (2006). 3 Friedberg (2001) instead studies the impact of mass-migration of Russian workers into Israel using occupations as the relevant dimension. No evidence of a negative relation between the inflow of immigrants to certain occupations and the wage growth of natives working in the same occupation is found, and this is interpreted as evidence for the existence of immigrant-native complementarity within occupations. 4 Aydemir and Borjas (2007) finds also a large negative impact for Canada, Mexico, and the U.S. However, the same approach applied to European countries (see Bonin, 2005, Steinhardt, 2011, and Bauer, Flake and Sinning, 2011, for Germany; Carrasco, Jimeno and Ortega, 2008, for Spain; and Dustmann, Fabbri and Preston, 2005 for the U.K.) has produced much smaller impacts of immigration. Some of these studies have been criticized by Aydemir and Borjas (2011) on the basis that they typically use a relatively small sample to compute the share of immigrants per education/experience cell. 2

3 substitutes. However, Borjas, Grogger and Hanson (2008, 2012) argue that one cannot reject that immigrants and natives are perfect substitutes in the U.S. case. This paper contributes to the literature by explicitly identifying occupations as an important source of imperfect substitutability between natives and immigrants within education/experience cells. Indeed, we provide evidence that the presence of immigrants in education/experience cells is associated to a specialization of the natives within the cell to better-paid occupations that require a more extensive use of abstract tasks. In evaluating the impact of immigration to France for , 5 our starting point is as most of the papers above Borjas (2003) s specification. We focus on France because French data have two important advantages with respect to the data available for other European countries. First, given the long period of time under consideration, the immigrant share varies in France to an extent similar to that of the U.S., which is crucial given that the identification of the model depends on annual aggregate changes over time across population subgroups. Second, we have access to 25% extracts of the Census population for most of the censuses in that period, which is larger than Borjas (2003) s sample, and thus renders the results immune from attenuation biases as identified in Aydemir and Borjas (2011). In the baseline specification à la Borjas (2003), we find that a 10% increase in the immigrant share is associated with wages higher by 3% and employment rates higher by 2.5%. 6 The results are shown to be robust to different specifications, and in particular to accounting both at the national and regional level for the potential self-selection of immigrants to the best paid education/experience cells. Specifically, when instrumenting with the settlement pattern of immigrants across 21 French regions, the impact of immigration on wages is still found to be positive and significant, with associated estimates quite close to the national-level estimates. This indicates that endogeneity problems are unlikely to be the driving force for the positive correlation between immigration and natives labor market outcomes. To understand the positive correlation between immigration and wages, we explore the link 5 To the best of our knowledge, this is the first paper applying a factor proportion approach to France. Hunt (1992) studies the impact of immigration on natives exploiting the spatial variation in the settlement of the repatriates from Algeria in Ortega (2000) proposes a theoretical rationale for why immigration may increase native wages and lower native unemployment. 3

4 between immigration and the occupational distribution of natives within education/experience cells. To this purpose, we first decompose the evolution of wages to study whether immigration is related to changes in the distribution of workers across occupations within the cell and/or to changes in the prices of different occupations. We find that higher wages coming from the reallocation of natives towards better paid occupations are positively correlated with the presence of immigrants in the education/experience cell. Quantitatively, native reallocation across occupations accounts for an important part of the overall impact. Then, we exploit the richness of the head-counts Census data to understand the nature of this reallocation. Specifically, we first show that low-educated natives are less likely to work as blue-collar workers if there are many immigrants in their education/experience cell and that, symmetrically, highly-educated natives are more likely to work in white-collar occupations if immigration in their cell is high. In order to check whether the same type of result holds for a more detailed occupational classification, we study whether the degree of similarity of the distributions of natives and immigrants within education/experience cells across 120 three-digit level occupations depends on the presence of immigrants in the corresponding cell. To this purpose, we regress across education/experience cells the congruence index of the distribution of natives vs. immigrants on the log ratio of immigrants to natives, both at the national and regional level. In all the specifications, we show that the congruence of the two distributions is negatively related to the presence of immigrants in the cell, which goes again in the direction of concluding that the presence of immigrants favors the relative specialization of natives within the cell. Finally, we build on the task content literature 7 to study whether there are qualitative differences in the tasks performed by natives in education/experience cells depending on the number of immigrants in the cell. Specifically, we follow Autor and Dorn (2011) and Goos, Manning and Salomons (2010) and focus on the distinction between "abstract" and "routine" tasks. Both 7 See in particular Autor, Levy and Murnane (2003), Goos and Manning (2007), Black and Spitz-Oener (2010), or Acemoglu and Autor (2011) for papers on the demand for job skills, and Peri and Sparber (2009), Amuedo- Dorantes and de la Rica (2011), Prantl and Spitz-Oener (2011) or Imai, Stacey and Warman (2012) for papers specifically interested in immigration. In particular, Peri and Sparber (2009) and Amuedo-Dorantes and de la Rica (2011) provide evidence across respectively US states and Spanish regions that low-skilled natives respond to immigration by specializing in language-intensive tasks for which they have a comparative advantage and which are better remunerated than manual-physical tasks. 4

5 our national- and regional-level estimates show that a higher share of immigrants in the cell is associated with a decline in the average routine intensity of the occupations of natives and a higher specialization of natives in abstract tasks relative to immigrants. The paper is organized as follows: Section 1 describes the data, Section 2 presents education/experience regressions to assess the impact of immigration on the labor market outcomes of natives, and finally Section 3 studies the relation between immigration and the occupational distribution of natives (versus immigrants) within education/experience cells. 1 Data We use data from six successive French censuses from 1962 to 1999 (1962, 1968, 1975, 1982, 1990, and 1999) to compute the number of immigrants and natives with a given level of education and labor market experience in each year. As detailed below, the 1962 Census data are just used to construct an instrumental variable for the immigrant share. As common in the literature (see for instance Borjas, 2003, or Manacorda et al., 2012), we restrict our attention for most of our estimates to males aged Men are classified into four educational groups depending on their highest attained diploma: no education or primary education (less than six years of education), secondary education (between 6 and 9 years), high school (11 or 12 years), and college (at least 14 years). 8 Table 1 shows the evolution of the educational composition of the male French labor force over the period. The most striking feature is that the share of individuals with only primary education decreased from about 80% in 1962 to 24.5% in 1999, while the share of individuals with high-school or college diploma rapidly increased. Labor market experience is measured as the age of the individual minus the entry age into the labor market. As the entry age into the labor market is not observed, we assume that individuals with primary, secondary, high school, and college education enter the labor market respectively when 15, 16, 19, and 24 years old. In addition, we restrict the analysis to individ- 8 Appendix 1 provides a detailed match between reported diplomas in the censuses and our education groups. We follow here the diploma classification which serves as a reference for French labor relations. A category some college cannot be constructed for the entire sample period as data on uncompleted college ( Bac+2 and Bac+3 ) is only available since However, given the relatively low educational level of the French labor force at the beginning of the sixties, such distinction is not fundamental for the period of time under consideration. 5

6 Table 1: Distribution of Educational Attainment in the French Population (percentage) Primary School Secondary School High School College Notes: Tabulations include men aged between 18 and 64 years old, not enrolled in school nor in military. uals between 1 and 40 years of labor market experience. For each education level, we group individuals in 5-years experience groups. Following Borjas (2003), the immigration shock experienced by natives with education i and experience j at year t can be measured by p ijt, the relative share of immigrants among all individuals in the cell: p ijt = M ijt /(M ijt + N ijt ), (1) where N ijt and M ijt respectively denote the number of natives and the number of immigrants in the corresponding cell. Table 2 reports p ijt for the male population between 1962 and 1999 as computed from the Census data. From this table, it appears that the evolution of the share of immigrants over time greatly varies across education groups. For individuals with primary school education, the share first rises and then declines. Instead, the share of immigrants among individuals with secondary education rises over the period, although not always in a monotonous fashion. Finally, for higher educational levels (high school and college graduates), the share of immigrants generally decreases until 1982 and then rises in the 80s and the 90s, an evolution opposite to the evolution for individuals with primary education. Since the French Census does not include information on income or wages, we rely on other surveys to construct our wage sample. 9 As no information on wages is available at all for 1968 and 1975, the best approximation is the 1969 and 1976 data on annual wage income available in the 1970 and 1977 Enquête Formation et Qualification Professionnelles (FQP). For 1982, 1990, and 1999, the best available information on wages is given by the corresponding French Labor Force Survey (LFS), which provides information since 1982 on monthly wages in the 9 We are thus following Katz and Murphy (1992) in constructing separate count and wage samples. 6

7 Table 2: Percent of Male Labor Force that is Foreign Born per Education/Experience cell Education Experience Primary School Secondary School High School College Graduates Notes: For each census year, the Table reports the percentage of immigrants among workers with similar education level and labor market experience. Sources: Census of Population,

8 month preceding the survey month. 10 Using the LFS and FQP data, we compute the average log monthly wage of natives and convert it into 2007 euros using the CPI deflator from the French Statistical Institute (INSEE). Average log wages per experience and education level over the period are reported in Table 3. The picture for wages over time is quite simple and uniform across education/experience cells. Indeed, with few exceptions, wages rise during the period and then decrease throughout the period. 2 Immigration and Labor Market Outcomes of Natives This section presents regressions at the education/experience level to try and identify the relation between immigration and natives labor market outcomes. Specifically, Section 2.1 presents the estimates of the Borjas (2003) specification, Section 2.2 reports specifications with alternative measures of the immigration shock, and Section 2.3 presents the estimates of models identified using geographical variations. 2.1 Borjas model The initial specification (Borjas, 2003) relates the labor market outcomes of natives to the immigrant share across education/experience groups as follows: y ijt = θp ijt + ψ F E + ϕ ijt (2) where y ijt is a labor market outcome at period t for natives with education i and experience j, p ijt is the immigrant share, and ψ F E is a set of education, time, and experience fixed effects s with their corresponding interactions 11 i.e. ψ F E = s i +s j +s t +(s i s j )+(s i s t )+(s j s t ). 10 Given the small number of observations by education/experience cell, we do not include immigrants wages in the analysis. 11 The set of fixed effects included in the model should absorb the impact of demand shocks to different education/experience groups unrelated with changes in the immigrant share. If this is not the case, our estimates are biased if unobserved changes in labor demand are correlated with the immigrant share. The issue of which fixed effects should be included in the model has been recently debated in the literature. Borjas (2003) and Borjas et al. (2008) preferred specification includes education-by-time and experience-by-time fixed effects in addition to education, experience and time fixed effects. Instead, Ottaviano and Peri (2012) argue that such a saturated 8

9 Table 3: Log Monthly Wage of Full Time Male Native Workers Per Education/Experience Education Years of Experience Primary Education Secondary Education High School College Graduates Notes: The table provides the average log monthly wage of native men, working full time, per education/experience group. See text for details. The population excludes self-employed and civil servants. Wages are in 2007 euro, using the CPI computed by the INSEE. Sources: FQP 1964, 1970, 1977 and LFS 1982, 1990,

10 The upper panel of Table 4 presents estimates of (2) using OLS or WLS. In the baseline case (row 1), and in contrast with Borjas (2003), the immigrant share is found to be positively and significantly correlated with the average log monthly wage (column 1), the employment to population ratio (column 2), and the employment to labor force ratio (column 3). Quantitatively, the estimated impact is quite large: a 10% increase in the immigration share is estimated to raise native s wages by 3.0%, the employment/population ratio by 2.5%, and the employment/labor force ratio by 2.2%. 12 Figure 1 illustrates how well the model fits the data. The figure plots the residuals from the regression of wages on the fixed effects and the residuals of the immigrant share on the fixed effects, together with the GLS regression line. From the figure, the results do not appear to be driven by specific observations. For instance, if one excludes the observations with log wage residuals above 0.1 or below -0.1, we obtain similar results with a parameter estimate (standard deviation) of (0.118). A first concern for the validity of these initial estimates is that changes in participation rates and wages across demographic groups over this period might be spuriously correlated with variations of the immigrant share. Although France and the U.S. experienced similar employment-population ratios during the 60s, the employment-population ratio in France dramatically fell after that period both for both young workers (under 25) and old workers (above 55). Even if our model controls for interactions between experience and year which should absorb the effect of this change across education groups over time, our results may potentially reflect these changes if immigration was lower for some cells within education groups. As a check of the robustness of our findings, Row 3 eliminates from the sample the cells of less than 11 years of experience and more than 30 years of experience. In that case, the estimated effects model might absorb most of the identifying variations necessary to estimate the model. We remain conservative and include the full set of fixed effect and interactions between education, experience and time in all the models estimated in this paper. This guarantees that our estimates are not biased by potential changes in the returns to education or experience over time. Another hypothesis necessary for the correct identification of the model is that natives do not change their education level over time in response to actual or anticipated future variations of immigrants flows. Following Hunt (2010), which shows that immigration has a relatively small effect on educational levels in the U.S., this may be a reasonable assumption. 12 When the endogenous variable y ijt represents the wage, the parameter θ can be interpreted as an elasticity giving the percentage change in wages associated with a percentage change in labor supply. As in Borjas (2003), we define the "wage elasticity" as log w/ m = θ/(1 + m) 2 Over the period, the mean value of the relative number of immigrants (m) is about 9%. The wage elasticity evaluated at the mean value can therefore be obtained by multiplying θ by

11 Table 4: Impact of Immigrant Share per Education/Experience Cells Specification Av. Log Employment Employment N Monthly wage Population Labor Force A. WLS/OLS 1. Basic Estimates 0.366*** 0.308*** 0.260*** 160 (0.117) (0.060) (0.045) 2. Unweighted Regression 0.342** 0.449** 0.302*** 160 (0.161) (0.156) (0.080) 3. Experience 0.277** 0.187*** 0.208*** 80 between 11 and 30 (0.119) (0.027) (0.027) 4. Only Primary 0.418*** 0.272*** 0.258*** 80 and Secondary Education (0.120) (0.048) (0.043) 5. Estimates without 0.344*** 0.315*** 0.253*** years exp. (0.117) (0.054) (0.033) 6. Women and men in p ijt and 0.453** *** 160 in dependent variable (0.194) (0.137) (0.098) 7. Women and men 0.171* *** 160 in dependent variable (0.100) (0.074) (0.046) B. 2SLS 8. Basic Estimates 0.404*** 0.390*** 0.313*** 160 (0.130) (0.077) (0.050) 9. Without weights 0.320* 0.582*** 0.379*** 160 (0.166) (0.206) (0.098) 10. Experience 0.335** 0.193*** 0.201*** 80 between 11 and 30 (0.145) (0.043) (0.048) 11. Only Primary 0.490*** 0.310*** 0.285*** 80 and Secondary Education (0.134) (0.051) (0.043) 12. p i,(j,t) 2 as IV 0.393** 0.267*** 0.154*** 72 (0.153) (0.071) (0.028) 13. p i,(j,t) 3 as IV 0.259* 0.352*** 0.162*** 40 (0.136) (0.105) (0.012) Notes: The table reports the coefficient of the immigrant share from regressions for the above mentioned dependent variables for the period For rows 8 to 13, the model is estimated using 2SLS taking p i,(j,t) 1 as an instrument. Robust heteroscedastic standard errors reported in parenthesis are adjusted for clustering within education/experience cells. Controls (fixed effects) are added for education, experience, year, and for interactions between education and experience, year and experience, education and year. When the dependent variable is the employment to population or the employment to labor force rate, weights are the number of natives per cell divided by the total number of natives in the census year. When the dependent variable is the average log wages, weights are the number of observations per cell used to compute the average wage with the LFS or FQP divided by the total number of observations used to compute average wages per year. *, ** and *** denotes significant at respectively 10%, 5% and 1% level. Sources: Census of Population , FQP 1970, 1977 and LFS 1982, 1990,

12 Log Wage Residuals Immigrant Share Residuals Figure 1: Residuals of log Wages versus Immigrant Share Men with experience between 1-40 The figure depict the residuals of the immigrant share on fixed effects with the residuals of the log average wages on fixed effects. of immigration on log wages and employment rates remain positive, although non significant for wages. A second issue is that the impact of immigration could differ across education groups in which case our estimates may reflect the simultaneous increase in immigration and wages for the most educated groups. However, the estimates in Row 4 for low education levels (primary or secondary education only) still display a positive and significant correlation between immigration and wages, and a positive (although of lower magnitude) correlation with employment rates. 13 Row 5 shows that the coefficients remain similar when we exclude from the analysis the individuals with less than 10 years of experience, for which the prevalence of the minimum wage is very important, especially after The minimum wage is thus unlikely to be the 13 The estimated impact of immigration for the individuals with more than secondary education only (not reported) is negative but never significantly different from zero. 14 The proportion of natives paid at the minimum wage plus 5% peaks at 87.3% in 1999 for the individuals with primary education and experience level 1-5, increases rapidly over time, and is generally non negligible among the least educated for all experience levels and the least experienced for the all education levels. Instead, the share is systematically below 5% for individuals with at least high school education and more than ten years of experience. A table presenting the share of workers paid at the minimum wage across education/experience cells is available upon request. 12

13 main factor behind the positive correlation between immigration and natives labor market outcomes. As men and women clearly do not constitute two disjoint segments of the labor market, Rows 6 and 7 provide estimates when data for women are included in the regressions. Specifically, Row 6 measures both the immigrant share and the dependent variable including data for women, while Row 7 keeps an immigrant share defined only for men but includes women in the endogenous variable. On the whole, including females does not alter qualitative results, although the coefficient on the employment to population ratio becomes insignificant. However, including women in the measure of the immigrant share may be problematic due to the lower participation rate of immigrant women and the important increase over the period of the participation rate of female natives. Similarly, changes in the selection of women that participate to the labor market (see Mulligan and Rubinstein, 2008, and Olivetti and Petrongolo, 2008) may render difficult the interpretation of the coefficient associated to the wage when the wages of women are included in the computation of the native s wage. For these reasons, we follow the rest of the literature and stick to regressions for males aged only. Another issue is that the variations in the immigrant share can also be endogenous if the immigrant share depends on the particular outcomes of a cell i.e. if immigrants with education/experience levels associated to relatively higher wages are more likely to enter France. We attempt to deal with this issue by using an instrumental variable strategy based on the persistence of the immigration decision over time. Assuming there is no autocorrelated cohort effect in the error term, a plausible instrument for p ijt is the immigrant share for the same cohort of workers in the preceding census, i.e. the i-level educated immigrants with j k to j k + 5 years of experience in census year t k (p i,(j,t) 1 ). 15 Indeed, given that most immigrants stay in France independently of changes in labor market conditions, 16 p i,(j,t) 1 is likely to be correlated with the contemporary immigrant share p ijt and uncorrelated to contemporary outcomes in the cell conditional on the inclusion of other covariates. The lower panel of table 4 shows that the 15 Table 12 in Appendix 2 defines the cohorts. It follows for instance that the immigrant share in 1968 for primary education workers with years of experience is instrumented by the immigrant share in 1962 for primary education workers with years of experience 16 For example, Schor (1996) shows that subsidies to the return of immigrants to their country of origin during economic downturns were never successful in the French case. 13

14 estimates using that instrument are similar to those of the upper panel. In addition, the instrument proves to be strong for all regressions, with a first-stage F-stat typically above 40. In case the distance between two consecutive censuses would not be sufficient to purge the potential endogeneity of the immigrant share, which may be particularly the case if the error terms are serially correlated at the cohort level, rows 12 and 13 use as instruments the immigrant share in increasingly distant censuses, with the estimates remaining qualitatively unchanged. 2.2 Alternative measures of the immigration shock As pointed out by Bohn and Sanders (2007), the use of the immigrant share as a dependent variable can be problematic. Clearly, if the number of natives within given education/experience cells is stable over time, changes in the immigrant share will essentially capture variations in the number of immigrants over time in these cells. However, if the number of natives is not stable, using the immigrant share constrains the effect of an increase in the number of immigrants and a decrease in the number of natives to be the same. Given that France experienced a large increase in general high-school graduation rates in the period under consideration (see Table 1), it appears necessary to test for a specification where both the number of natives and the number of immigrants are included as regressors. The first column of table 5 presents a regression controlling for both the immigrant share and the log of natives in the cell. As expected, the coefficient of the log of natives is negative and measured relatively precisely. Controlling for the log of natives in the regression reduces by a third the magnitude of the estimated effect of the immigrant share with respect to the baseline model, but the correlation is still positive. This suggests that controlling separately for the variations in the supply of native workers might be important. To further investigate this issue, column (2) introduces separately the coefficient of the log of natives and immigrants. 17 Once again, the coefficient of the log of natives is negative, while the coefficient of the log of immigrants is positive. The results indicate that similar variations in the number of natives have a much stronger (negative) effect on wages than variations in the number of immigrants. 17 This last alternative is not used by Borjas (2003) because in its data 15.3% of the education/experience cells by state of residence have no immigrants. Instead, with our very large sample extracts, all education/experience cells at the regional level have immigrants. 14

15 Table 5: Impact of the Log of Immigrants and the Log of Natives per Education/Experience Cells Alternative Measures of the Immigrant shock Dependent variable: Log Wages (1) (2) Immigrant Share 0.271** (0.122) Log natives * *** (0.025) (0.024) Log immigrants 0.028* (0.015) Log (immigrants/natives) 0.040*** (0.014) Notes: The table reports the results from regression of the average log wages on the indicated variables using observations for Controls (fixed effects) are added for education, experience, year, and for interactions between education and experience, year and experience, education and year. Robust heteroscedastic standard errors reported in parenthesis are adjusted for clustering within education/experience cells. Weights are the number of observations per cell used to compute the average wage with the LFS or FQP divided by the total number of observations used to compute average wages per year. *, ** and *** denotes significant at respectively 10%, 5% and 1% level. The number of observations is 160. Sources: Census of Population , FQP 1970, 1977 and LFS 1982, 1990, While the coefficient associated to the log of immigrants is much smaller than the coefficient found in Table 4, once one considers that a 10% increase in the immigrant share is equivalent to a 122% increase in the log of immigrants (for a given number of natives), 18 the associated increase in wages with the estimates of column (2) in Table 5 becomes 3.41%, which is quite close to the 3.7% baseline estimate in Table 4. In addition, this specification allows us to test for the implicit restriction made by model (2) that an increase in the number of immigrants has a symmetric effect with respect to a a decrease in the number of natives in the cell. Denoting by θ I the parameter of the log of immigrants in the regression of column 2 and by θ N the parameter of the log of natives, a simple test for θ I = θ N provides a chi-square of 3.16 and a p-value of 7.56%. This suggests that the implicit restriction is not too strongly rejected by the data. Finally, the last column shows that a 10% increase in the log of immigrants to natives is associated with wages higher by 0.4%. 18 The percentage increase of the number of immigrants necessary to increase the immigrant share by 10 percentage points is given by the formula 0.1 l (1 l) where l is the initial immigrant share. With an average 0.9 immigrant share in the sample, the percentage increase is 122%. 15

16 2.3 Regional estimates As the identification of aggregate model might be biased if the immigrant share is endogenous, we now report estimates based on geographical variation across 21 French metropolitan regions. 19 The main interest in using the geographical approach is that we can use the proportion of co-nationals in the region as an instrument for future immigrant inflows, as e.g. in Card (2001) and Cortes (2008). Specifically, let C denote the immigrant s country of origin and t 0 a reference year. Then, one can predict the region-r number of immigrants with education/experience (i, j) for year t ( ˆM ijtr ) as: ˆM ijtr = C λ CR M C,ijt (3) where λ CR = M CR,t 0 M C,t0 is the share of country-c immigrants in region R in the reference year t 0 and M C,ijt is the total number of immigrants with education/experience (i, j) residing in France in year t. 20 Given our large sample size, we distinguish groups of immigrants by using the maximum number of nationalities available, namely the 54 different countries of birth which are always reported separately across censuses. 21 Our instrument is then simply computed as: ˆp ijtr = ˆM ijtr ˆM ijtr +N ijtr. Importantly, the use of up to 25% census extracts guarantees sampling errors in the immigrant share to be relatively small even for regions with few immigrants, thus reducing the extent of potential attenuation biases from sampling error. Table 6 reports WLS and 2SLS regional level estimates. Following Borjas (2003), we esti- 19 We exclude Corsica from the sample, as it represents only 0.5% of the French population. Even if they are biased by native outflows in response to immigrant inflows, estimates of the impact of immigration using spatial correlations can be used as a lower bound for the magnitude of the effects. Empirically, most studies have concluded that displacement effects of immigrants by natives are not very large: for the U.S., Card and DiNardo (2000), Card (2001) and Cortes (2008) have found no evidence of natives outflows in response to immigrant inflows. One notable exception is Borjas (2006) who finds a strong displacement effect. Peri and Sparber (2011) argues that Borjas (2006) s results might be explained by the particular empirical specification that is used in that paper. 20 Following Hunt and Gauthier-Loiselle (2010), we construct λ CR with data from immigrants of all education and experience levels in order to allow for a greater role of geography and ethnic networks. In doing so, we hope to avoid the results being driven by economic factors that might attract workers with specific education/experience levels to specific regions. 21 We use two versions of this instrument. The first version uses 1968 as the reference year for all censuses, while our second version uses the lagged census year as a reference year. The two versions are likely to be quite different given that the stock of immigrants in 1968 comes mainly from Europe and the Maghreb, while post- 1970s immigration comes also from Asia and Sub-Saharan Africa. Therefore, in some sense, the first version of the instrument uses traditional long run immigrant flows while the second instrument is related with more recent immigrant waves. 16

17 mate models with two different sets of fixed effects which should absorb the impact of regional labor demand shocks. Specifically, Panel A s regressions include controls for each possible two way interactions and fixed effects between education, experience, region and year, while Panel B s regressions also include controls for an effect of region by education and by experience. In addition, the model is estimated using both the immigrant share and the log of natives over immigrants as measures of the immigration shock. Interestingly, the positive correlations between the immigrant share and both the wages and the employment rates of natives still hold at the regional level. As in Borjas (2003), however, we find that the value of the estimated coefficients is greatly attenuated, as the estimates for the immigrant share are almost halved. The coefficient of the impact of the immigrant share on wages falls from 0.36 in the aggregate model to 0.18 and 0.19 in WLS estimates for respectively Panels A and B. The coefficient of the effect of log(m/n) is also attenuated from at the national level to at the regional level estimated with WLS. The first stage of the instrumentation with the settlement pattern of immigrants is strong (with F stats typically above 50), while over-identifying restrictions do not reject the exogeneity of the instruments. Interestingly, the 2SLS estimates are systematically higher than their WLS counterparts and close to national-level estimates. 3 Immigration and Occupations within Education/Experience Cells We next investigate whether changes in the distribution of natives across occupations related with variations of the immigrant share can explain the positive correlation between natives wages and immigration systematically obtained in the preceding section. 22 To this purpose, Section 3.1 proposes a decomposition of the evolution of wages to study whether immigration is related to changes in the distribution of workers across occupations within the cell and/or to changes in the prices of the different occupations. After that, section 3.2 studies whether the 22 Because models using the log of natives over immigrants seem more easily interpretable, we use in this section this specification instead of the immigrant share. However, results are qualitatively unaffected by this choice and estimation results from similar models using instead the immigrant share as a dependent variable are available upon request. 17

18 Table 6: Regional Models Dependent variable: Log Wages Employment Rates to Population Model A Fixed effects of Region x year ; educ x exp ; exp x year ; educ x year ; reg x educ ; exp x reg Immigrant Share 0.175** 0.235*** 0.190*** 0.241*** (0.075) (0.082) (0.025) (0.031) Log (immigrants/natives) 0.028*** 0.036*** 0.017*** 0.031*** (0.008) (0.010) (0.002) (0.003) Test of overidentifying restriction (p-value) (0.58) (0.84) (0.99) (0.35) Model B Previous Fixed effects and reg x educ x exp Immigrant Share 0.190** 0.240** 0.188*** 0.231*** (0.093) (0.099) (0.029) (0.036) Log (immigrants/natives) 0.026*** 0.026** 0.022*** 0.032*** (0.009) (0.011) (0.002) (0.003) Test of overidentifying restriction (p-value) (0.76) (0.79) (0.99) (0.35) Estimation Method WLS 2SLS WLS 2SLS WLS 2SLS WLS 2SLS Notes: The number of observations is The table reports the coefficients from regressions with the indicated dependent variables using observations from the period Controls (fixed effects) are added as indicated. Robust heteroscedastic standard errors reported in parenthesis are adjusted for clustering within region/education/experience cells. Weights are the number of natives per cell divided by the total number of natives used to compute average wages per year. *, ** and *** denotes significant at respectively 10%, 5% and 1% level. Models in columns 2, 4, 6 and 8 are estimated using 2SLS with the predicted immigrant inflows instruments based on 1968 settlement patterns. See text for details. Sources: Census of Population , FQP 1970, 1977 and LFS 1982, 1990,

19 presence of immigrants in education/experience cells is related to changes in the occupations or tasks performed by natives in those cells. 3.1 Decomposition of Natives Wages Estimates of the impact of immigration on natives might be biased if natives adjust for immigrant inflows by changing their location or capital investment decision as argued by Borjas et al. (1997). However, natives may also adjust to an immediate impact of immigration in an education/experience/occupation cell by moving to another occupation within the same education/experience cell. If this is the case, estimates of the impact of immigration are complicated by the fact that immigration simultaneously affects the wages in education/experience/occupation cells and the distribution of natives across occupations within education/experience cells. To explicitly account for this, let wijt k denote the average wage of individuals from cell (i, j, t) working in occupation k and s k ijt denote the share of workers from that cell working in occupation k. The average wage in cell (i, j, t) can be written as: w ijt = k s k ijtw k ijt (4) Then, changes in the wages of natives over time in an education/experience cell following an immigration inflow can be decomposed into changes in the distribution of natives across occupations and changes in their wages within occupations, i.e. w ijt w ijt 1 = k (s k ijt s k ij,t 1)w k ijt + k s k ij,t 1(w k ijt w k ijt 1) (5) If immigrants have an impact on the occupational distribution of natives, then s k ijt s k ij,t 1 s k ijt may be a function of the immigrant share. Therefore, even if immigrant inflows lowered wages in all occupations i.e. w k ijt < 0 k, w ijt would be positive if: s k ijtwijt k > k k s k ij,t 1 w k ijt (6) 19

20 Intuitively, the previous relationship might hold if following an immigration inflow wages decrease more rapidly in low-pay than in high-pay occupations and if natives move to high-pay occupations. 23 The decomposition in (5) can be empirically implemented by constructing two counterfactual series using the reweighting procedure in DiNardo, Fortin and Lemieux (1996). Specifically, let w ijt (w 0, s t ) denote the series constructed with the actual shares holding wages constant at their reference year t 0 level, i.e. w ijt (w 0, s t ) = k s k ij,tw k ij,t 0 (7) and w ijt (w t, s 0 ) denote the series constructed with the actual wages holding shares constant, i.e. w ijt (w t, s 0 ) = k s k ij,t 0 w k ijt. (8) Given the relatively small number of observations for wages corresponding to certain education/experience cells, our counterfactual wage distributions across occupations can only be constructed if we choose a relatively parsimonious definition of occupations, namely the basic professional status whereby workers are classified as white-collar worker, technician, or bluecollar worker. Panel A in Table 7 regresses w ijt (w 0, s t ) on the immigrant share and the usual set of fixed effects for different reference years. For all reference years, higher wages coming from the reallocation of natives towards better paid occupations are positively correlated with the presence of immigrants in the corresponding education/experience cell. Quantitatively, estimates suggest an increase in 10% of the number of immigrants in a cell is associated to an additional reallocation of natives across allocations generating a 0.15% to 0.3% increase in native wages. As the overall impact of a 10% increase in log(immigrants/natives) is 0.4% (see Table 5), these 23 One can design a Roy (1951)-type model to generate such an outcome. Assume for instance a simple model with a high-pay occupation (occupation 1) and a low-pay occupation (occupation 2). Assume that the cost for an immigrant of accessing the high-pay occupation is prohibitively high, and as a result all immigrants stay in occupation 2. The cost (c) is instead affordable for natives, who move to occupation 2 so far as w 1 w 2 > c. As immigrants accrue to occupation 2, wages become lower in that occupation, and more natives move to occupation 1. With two occupations, condition (6) for wages to rise with immigration becomes s 1 c w 2, i.e. the increase in the share of high-pay natives multiplied by the wage premium of high-pay workers must compensate for the lower wages of low-pay natives. 20

21 Table 7: Decomposition of the Evolution of Wages from the Role of Occupations and Prices Reference year: A. Dependent variable: Average Log Wages Constant Prices // Occupations varies log(immigrants/natives ) 0.033*** 0.023*** 0.022*** 0.025*** 0.015** (0.009) (0.007) (0.006) (0.007) (0.006) B. Dependent variable: Average Log Wages Constant Occupations // Price varies log(immigrants/natives ) 0.022*** 0.029*** 0.025** (0.008) (0.009) (0.011) (0.008) (0.008) N Notes: The table regresses counterfactual average log native wages for education/experience/time cells against the log (immigrants/natives) and a set of fixed effects. Panel A s wages are constructed using the changes over time in the distribution of workers within education/experience cells across three basic occupations (white-collar, technician, and bluecollar) while holding wages at the education/experience/occupation level constant at the reference year level. Panel B s wages are constructed using the changes over time in education/experience/occupation wages holding constant the occupational distribution at the reference year level. Sources: Census of Population , FQP 1970, 1977 and LFS 1982, 1990, results suggest that native reallocation across occupations accounts for a large part of the overall impact. Panel B in Table 7 shows that counterfactual fixed-occupation wages (w ijt (w t, s 0 )) are also positively correlated to the immigrant share, although to a lower extent than (w ijt (w 0, s t )), as the relation is insignificant for two out of five reference years. A positive correlation of this type is the outcome predicted by a model such as Ortega (2000) but is at odds with the simple Roy-type model sketched above. In principle, this result could stem from demand shocks affecting the relative price of labor across occupations that are imperfectly absorbed by the fixed effects. 3.2 Characteristics of Natives Occupations and Immigration The preceding section provides evidence that an important part of the positive relation between immigration and natives wages is associated to a shift in the occupational distribution of natives within education/experience cells. In this section, we exploit the rich head counts Census data to understand the nature of this reallocation. Specifically, we study whether a higher presence of immigrants in the relevant cell is associated to (i) the natives being allocated to better paid professional status categories, (ii) more dissimilar distributions of natives vs immigrants 21