The Labor Demand Curve Is Downward Sloping: Reexamining the Impact of Immigration on the Labor Market Author(s): George J. Borjas Source: The Quarterly Journal of Economics, Vol. 118, No. 4 (Nov., 2003), pp. 1335-1374 Published by: Oxford University Press Stable URL: http://www.jstor.org/stable/25053941 . Accessed: 09/03/2015 08:44 Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp . JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org. . Oxford University Press is collaborating with JSTOR to digitize, preserve and extend access to The Quarterly Journal of Economics. http://www.jstor.org This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions THE LABOR DEMAND CURVE IS DOWNWARD SLOPING: REEXAMINING THE IMPACT OF IMMIGRATION ON THE LABOR MARKET* George J. Borjas Immigration is not evenly balanced across groups of workers who have the same education but differ in their work experience, and the nature of the supply imbalance changes over time. This paper develops a new approach for estimating the labor market impact of immigration by exploiting this variation in supply shifts across education-experience groups. I assume that similarly educated work ers with different levels of experience participate in a national labor market and are not perfect substitutes. The analysis indicates that immigration lowers the wage of competing workers: a 10 percent increase in supply reduces wages by 3 to 4 percent. "After World War I, laws were passed severely limiting im migration. Only a trickle of immigrants has been admitted since then ... By keeping labor supply down, immigration policy tends to keep wages high." Paul Samuelson, Economics [1964] I. Introduction Do immigrants harm or improve the employment opportuni ties of native workers? As Paul Samuelson's assertion suggests, the textbook model of a competitive labor market predicts that an immigrant influx should lower the wage of competing factors.1 Despite the intuitive appeal of this theoretical implication and despite the large number of careful studies in the literature, the existing evidence provides a mixed and confusing set of re sults. The measured impact of immigration on the wage of native workers fluctuates widely from study to study (and sometimes even within the same study), but seems to cluster around zero. A widely cited survey by Friedberg and Hunt [1995, p. 42] concludes that "the effect of immigration on the labor market outcomes of * I am grateful to Daron Acemoglu, Joshua Angrist, David Autor, Richard Freeman, Daniel Hamermesh, Lawrence Katz, Michael Kremer, Casey Mulligan, and Stephen Trejo for helpful comments and suggestions, and to the Smith Richardson Foundation for financial support. 1. The historical context of Samuelson's [1964, p. 552] assertion is interest ing. He was writing just before the enactment of the 1965 Amendments to the Immigration and Nationality Act, the major policy shift that initiated the resur gence of large-scale immigration. ? 2003 by the President and Fellows of Harvard College and the Massachusetts Institute of Technology. The Quarterly Journal of Economics, November 2003 1335 This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions 1336 QUARTERLY JOURNAL OF ECONOMICS natives is small." Similarly, the 1997 National Academy of Sci ences report on the economic impact of immigration argues that "the weight of the empirical evidence suggests that the impact of immigration on the wages of competing native workers is small" [Smith and Edmonston 1997, p. 220]. These conclusions are po tentially inconsistent with the textbook model because the immi grant supply shock in recent decades has been very large, and most studies of labor demand (outside the immigration context) conclude that the labor demand curve is not perfectly elastic [Hamermesh 1993]. This paper presents a new approach for thinking about and estimating the labor market impact of immigration. Most existing studies exploit the geographic clustering of immigrants and use differences across local labor markets to identify the impact of immigration. This framework has been troublesome because it ignores the strong currents that tend to equalize economic condi tions across cities and regions. In this paper I argue that by paying closer attention to the characteristics that define a skill group?and, in particular, by using the insight that both school ing and work experience play a role in defining a skill group?one can make substantial progress in determining whether immigra tion influences the employment opportunities of native workers. My analysis uses data drawn from the 1960-1990 U. S. Decennial Censuses, as well as the 1998-2001 Current Popula tion Surveys, and assumes that workers with the same education but different levels of work experience participate in a national labor market and are not perfect substitutes. It turns out that immigration?even within a particular schooling group?is not balanced evenly across all experience cells in that group, and the nature of the supply imbalance changes over time. This fact generates a great deal of variation?across schooling groups, experience cells, and over time?that helps to identify the impact of immigration on the labor market. Most importantly, the size of the native workforce in each of the skill groups is relatively fixed, so that there is less potential for native flows to contaminate the comparison of outcomes across skill groups. In contrast to the confusing array of results that now permeate the literature, the evidence consistently suggests that immigration has indeed harmed the employment opportunities of competing native workers. This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions LABORMARKET IMPACT OF IMMIGRATION 1337 II. Measuring the Labor Market Impact of Immigration The laws of supply and demand have unambiguous implica tions for how immigration should affect labor market conditions in the short run. The shift in supply lowers the real wage of competing native workers. Further, as long as the native supply curve is upward sloping, immigration should also reduce the amount of labor supplied by the native workforce. If one could observe a number of closed labor markets that immigrants penetrate randomly, one could then relate the change in the wage of workers in a particular skill group to the immi grant share in the relevant population. A negative correlation (i.e., native wages are lower in those markets penetrated by immigrants) would indicate that immigrants worsen the employ ment opportunities of competing native workers. In the United States, immigrants cluster in a small number of geographic areas. In 1990, for example, 32.5 percent of the immigrant population lived in only three metropolitan areas (Los Angeles, New York, and Miami). In contrast, only 11.6 percent of the native population clustered in the three largest metropolitan areas housing natives (New York, Los Angeles, and Chicago). Practically all empirical studies in the literature, beginning with Grossman [1982], exploit this demographic feature to identify the labor market impact of immigration. The typical study defines a metropolitan area as the labor market that is being penetrated by immigrants. The study then goes on to calculate a "spatial corre lation" measuring the relation between the native wage in a locality and the relative number of immigrants in that locality. These correlations are usually negative, but very weak.2 The best known spatial correlations are reported in Card's [1990] influen tial study of the Mariel flow. Card compared labor market condi tions inMiami and in other cities before and after the Marielitos increased Miami's workforce by 7 percent. Card's difference-in differences estimate of the spatial correlation indicated that this 2. Representative studies include Altonji and Card [1991], Borjas [1987], LaLonde and Topel [1991], Pischke and Veiling [1997], and Schoeni [1997]. Friedberg [2001] presents a rare study that uses the supply shock in an occupation to identify the labor market impact of immigration in the Israeli labor market. Although the raw Israeli data suggest a substantial negative impact, correcting for the endogeneity of occupational choice leads to the usual result that immigra tion has little impact on the wage structure. Card [2001] uses data on occupation and metropolitan area to define skill groups and finds that immigration has a slight negative effect. This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions 1338 QUARTERLY JOURNAL OF ECONOMICS sudden and unexpected immigrant influx did not have a discern ible effect on employment and wages in Miami's labor market.3 Recent studies have raised two questions about the validity of interpreting weak spatial correlations as evidence that immi gration has no labor market impact. First, immigrants may not be randomly distributed across labor markets. If immigrants endo genously cluster in cities with thriving economies, there would be a spurious positive correlation between immigration and wages.4 Second, natives may respond to the wage impact of immigration on a local labor market by moving their labor or capital to other cities. These factor flows would reequilibrate the market. As a result, a comparison of the economic opportunities facing native workers in different cities would show little or no difference because, in the end, immigration affected every city, not just the ones that actually received immigrants.5 Because the local labor market may adjust to immigration, Borjas, Freeman, and Katz [1997] suggested changing the unit of analysis to the national level. If the aggregate technology can be described by a CES production function with two skill groups, the relative wage of the two groups depends linearly on their relative quantities. By restricting the analysis to two skill groups, the "factor proportions approach" precludes the estimation of the impact of immigration?there is only one observation at any point in time (usually a Census year), giving relative wages and rela tive employment. As a result, the typical application of this ap proach compares the actual supplies of workers in particular skill groups with those that would have been observed in the absence of immigration, and then uses outside information on labor de 3. Angrist and Krueger [1999] replicate Card's study using an alternative time period, and find that a "phantom" influx of immigrants (in the sense that had it not been for a policy intervention, many immigrants would likely have arrived) had a sizable adverse effect on Miami's labor market. This result suggests that many other factors influence labor market conditions in Miami and comparison cities. At the least, one should be cautious when interpreting the spatial correla tions estimated from comparisons of specific localities. 4. Borjas [2001] presents evidence indicating that new immigrants belonging to a particular schooling group tend to settle in those regions that offer the highest return for their skills. 5. Borjas, Freeman, and Katz [1997] and Card [2001] provide the first at tempts to jointly analyze labor market outcomes and native migration decisions. The two studies reach different conclusions. Card reports a slight positive corre lation between the 1985-1990 rate of growth in the native population and the immigrant supply shock by metropolitan area, while Borjas, Freeman, and Katz report a negative correlation between native net migration in 1970-1990 and immigration by state?once one standardizes for the preexisting migration trends. This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions LABORMARKET IMPACT OF IMMIGRATION 1339 mand elasticities to simulate the consequences of immigration. The immigrant flow to the United States in the 1980s and 1990s was relatively low-skill. Not surprisingly, the Borjas-Freeman Katz [1997] simulation finds that immigration worsened the rela tive economic status of low-skill workers. Despite all of the confusion in the literature, the available evidence teaches two important lessons. First, the study of the geographic dispersion in native employment opportunities is not an effective way for measuring the economic impact of immigra tion; the local labor market can adjust in far too many ways to provide a reasonable analogue to the "closed market" economy that underlies the textbook supply-and-demand framework. Sec ond, the factor proportions approach is ultimately unsatisfactory. It departs from the valuable tradition of empirical research in labor economics that attempts to estimate the impact of labor market shocks by directly observing how those shocks affect some workers and not others. For a given elasticity of substitution, the approach mechanically predicts the relative wage consequences of supply shifts. Ideally, one would want to estimate directly how immigra tion alters the employment opportunities of a particular skill group. As noted above, by aggregating workers into groups based on educational attainment, there is just too little variation to examine how supply shocks affect relative wages. However, the human capital literature emphasizes that schooling is not the only?and perhaps not even the most important?determinant of a worker's skills. The seminal work of Becker [1975] and Mincer [1974] stressed that skills are acquired both before and after a person enters the labor market. Iwill assume that workers who have the same schooling, but who have different levels of experi ence, are imperfect substitutes in production. As a result, a skill group should be defined in terms of both schooling and labor market experience. To see how this insight can provide a fruitful approach to the empirical analysis of the labor market impact of immigration, con sider the following example. Recent immigration has increased the relative supply of high school dropouts substantially. The labor market implications of this supply shock clearly depend on how the distribution of work experience in the immigrant population con trasts with that of natives. After all, one particular set of native high school dropouts would likely be affected if all of the new low-skill This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions 1340 QUARTERLY JOURNAL OF ECONOMICS immigrants were very young, and a very different set would be affected if the immigrants were near retirement age. It is unlikely that similarly educated workers with very dif ferent levels of work experience are perfect substitutes [Welch 1979; Card and Lemieux 2001]. The definition of a skill group in terms of both education and experience provides a great deal more independent variation in the immigrant supply shock that can be used to identify how immigration alters the economic opportunities facing particular groups of native workers. III. Data The empirical analysis uses data drawn from the 1960, 1970, 1980, and 1990 Public Use Microdata Samples (PUMS) of the De cennial Census, and the 1999, 2000, and 2001 Annual Demographic Supplement of the Current Population Surveys (CPS). I pool all three of the CPS surveys and refer to these pooled data as the "2000" cross section. The analysis is restricted to men aged 18-64 who participate in the civilian labor force. A person is defined to be an immigrant if he was born abroad and is either a noncitizen or a naturalized citizen; all other persons are classified as natives. Ap pendix 1 provides a detailed description of the construction of the data extracts and of the variables used in the analysis. As noted above, I use both educational attainment and work experience to sort workers into particular skill groups. In particu lar, I classify the men into four distinct education groups: persons who are high school dropouts (i.e., they have less than twelve years of completed schooling), high school graduates (they have exactly twelve years of schooling), persons who have some college (they have between thirteen and fifteen years of schooling), and college graduates (they have at least sixteen years of schooling). The classification of workers into experience groups is bound to be imprecise because the Census does not provide any measure of labor market experience or of the age at which a worker first enters the labor market. I initially define work experience as the number of years that have elapsed since the person completed school. This approximation is reasonably accurate for most native men, but would surely contain serious measurement errors if the calculations were also conducted for women, particularly in the earlier cross sections when the female labor force participation rate was much lower. Equally important, this measure of experience is also likely This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions LABORMARKET IMPACT OF IMMIGRATION 1341 to mismeasure "effective" experience in the sample of immi grants, i.e., the number of years of work experience that are valued by an American employer. After all, a variable that roughly approximates "Age - Education - 6" does not differen tiate between experience acquired in the source country and experience acquired in the United States. I address this problem in Section VI below. I assume that the age of entry into the labor market is 17 for the typical high school dropout, 19 for the typical high school graduate, 21 for the typical person with some college, and 23 for the typical college graduate. Let AT be the assumed entry age for workers in a particular schooling group. The measure of work experience is then given by (Age AT). I restrict the analysis to persons who have between 1 and 40 years of experience. As noted inWelch's [1979] study of the impact of cohort size on the earnings of baby boomers, workers in adjacent experience cells are more likely to influence each other's labor market oppor tunities than workers in cells that are further apart. Throughout much of the analysis, Iwill capture the similarity across workers with roughly similar years of experience by aggregating the data into five-year experience intervals, indicating if the worker has 1 to 5 years of experience, 6 to 10 years, and so on. Consider a group of workers who have educational attain ment ?, experience level j, and are observed in calendar year t. The (ij,t) cell defines a skill group at a point in time. The measure of the immigrant supply shock for this skill group is defined by (1)pijt = Mijtl{Mijt + Nijt)9 where Mijt gives the number of immigrants in cell (i9j9t)9 sn?Nijt gives the corresponding number of natives. The variable pijt measures the foreign-born share of the labor force in a particular skill group. The various panels of Figure I illustrate the supply shocks experienced by the different skill groups between 1960 and 2000 (Appendix 2 reports the underlying data). There is a great deal of dispersion in these shocks even within schooling categories. It is well-known, for instance, that immigration greatly increased the supply of high school dropouts in recent decades. What is less well-known, however, is that this supply shift did not affect equally all experience groups within the population of high school dropouts. Moreover, the imbalance in the supply shock changes This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions A. High School Dropouts B. High School 10 20 30 Yearsof experience 10 C. Some College D. College Grad 20 30 Yearsof experience 10 Figure I The Immigrant Supply Shock, 1960-2000 Within each education group, workers are aggregated into experience intervals. The figures use the midpoint of each experience interval to This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions LABORMARKET IMPACT OF IMMIGRATION 1343 over time. As Panel A of the figure shows, immigrants made up half of all high school dropouts with ten to twenty years of experience in 2000, but only 20 percent of those with less than five years. In 1960, however, the immigration of high school dropouts increased the supply of the most experienced workers the most. Similarly, Panel D shows that the immigrant supply shock for college graduates in 1990 was reasonably balanced across all experience groups, generally increasing supply by around 10 percent. But the supply shock for college graduates in 1960 was larger for the most experienced groups, while in 2000 it was largest for the groups with five to twenty years of experience. The earnings data used in the paper are drawn from the sample of persons who worked in the year prior to the survey and reported positive annual earnings, are not enrolled in school, and are employed in the wage and salary sector. Earnings are deflated to 1999 dollars by using the CPI-U series. Table I summarizes the trends in log weekly wages for the various native groups. Not surprisingly, there is a great deal of dispersion in the rate of decadal wage growth by education and experience. Consider, for instance, the sample of college graduates. In the 1970s, wage growth was steepest for college graduates with 31-35 years of experience. In the 1990s, however, the wage of college graduates grew fastest for workers with 11-20 years of experience. In sum, the data reveal substantial variation in both the immigrant sup ply shock and native labor market outcomes across skill groups. Before proceeding to a formal analysis, it is instructive to document the strong link that exists between log weekly wages and the immigrant share within schooling-experience cells. In particular, I use the data reported in Table I to calculate the decadal change in log weekly wages for each skill group, and the data summarized in the various panels of Figure I (and reported in Appendix 2) to calculate the decadal change in the group's immigrant share. Figure II presents the scatter diagram relating these decadal changes after removing decade effects from the differenced data. The plot clearly illustrates a negative relation between wage growth and immigrant penetration into particular skill groups, and suggests that the regression line is not being driven by any particular outliers. Put simply, the raw data show that weekly wages grew fastest for workers in those education experience groups that were least affected by immigration. Finally, the validity of the empirical exercise reported below hinges on the assumption that similarly educated workers who This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions 1344 QUARTERLY JOURNAL OF ECONOMICS TABLE I Log Weekly Wage of Male Native Workers, 1960-2000 Education Years of experience 1960 1970 1980 1990 2000 High school dropouts High school graduates Some college College graduates 1-5 6-10 11-15 16-20 21-25 26-30 31-35 36-40 1-5 6-10 11-15 16-20 21-25 26-30 31-35 36-40 1-5 6-10 11-15 16-20 21-25 26-30 31-35 36-40 1-5 6-10 11-15 16-20 21-25 26-30 31-35 36-40 5.535 5.920 6.111 6.188 6.201 6.212 6.187 6.175 5.940 6.257 6.392 6.459 6.487 6.478 6.450 6.435 6.133 6.412 6.535 6.604 6.634 6.620 6.615 6.575 6.354 6.625 6.760 6.852 6.876 6.881 6.867 6.821 5.758 6.157 6.305 6.360 6.413 6.439 6.407 6.377 6.132 6.476 6.587 6.639 6.664 6.677 6.674 6.622 6.322 6.633 6.752 6.805 6.832 6.841 6.825 6.728 6.612 6.891 7.032 7.109 7.158 7.146 7.095 7.070 5.722 6.021 6.166 6.286 6.364 6.368 6.419 6.418 6.090 6.343 6.497 6.609 6.638 6.662 6.667 6.657 6.237 6.472 6.641 6.762 6.764 6.789 6.781 6.718 6.432 6.702 6.923 7.043 7.087 7.085 7.079 6.985 5.494 5.839 6.006 6.087 6.180 6.268 6.295 6.295 5.837 6.159 6.309 6.415 6.495 6.576 6.572 6.548 6.085 6.387 6.534 6.613 6.711 6.771 6.740 6.658 6.459 6.766 6.908 7.005 7.112 7.122 7.095 6.950 5.418 5.751 5.932 5.989 6.034 6.036 6.086 6.168 5.773 6.140 6.273 6.323 6.406 6.414 6.493 6.460 6.013 6.366 6.489 6.591 6.626 6.648 6.662 6.623 6.458 6.747 6.943 7.046 7.051 7.084 7.074 6.944 The table reports the mean of the logweekly wage ofworkers in each education-experience group. All wages are deflated to 1999 dollars using the CPI-U series. have different levels of experience are not perfect substitutes. Studies that examine this question, including Welch [1979] and Card and Lemieux [2001], find less than perfect substitutability across experience groups. Nevertheless, it is of interest to docu ment that (for given education) immigrants and natives with similar levels of experience are closer substitutes than immi grants and natives who differ in their experience. This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions LABORMARKET IMPACT OF IMMIGRATION 1345 -0.1 -0.05 0 0.05 0.1 0.15 0.2 Decadal change in immigrantshare Figure II Scatter Diagram Relating Wages and Immigration, 1960-2000 Each point in the scatter represents the decadal change in the log weekly wage and the immigrant share for a native education-experience group. The data have been adjusted to remove decade effects. The regression line in the figure weighs the data by (non^/inQ + nx), where n0 is the sample size of the cell at the beginning of the decade, and nx the sample size at the end. The slope of the regression line is -.450, with a standard error of .172. I use Welch's [1999] index of congruence to measure the degree of similarity in the occupation distributions of immigrants and natives. The index for any two skill groups k and I is defined by c Sc (qhc - qc)(qlc - qc)/qc where qhc gives the fraction of group h (h k9l) employed in occupation c, and qc gives the fraction of the entire workforce employed in that occupation. The index Gkh which is similar to a correlation coefficient, equals one when the two groups have identical occupation distributions and minus one when the two groups are clustered in completely different occupations. I calculate the index of congruence in the 1990 Census. I use the three-digit Census Occupation Codes to classify male workers into the various occupations, and restrict the analysis to workers in nonmilitary occupations. To minimize the problem of having many occupation-experience cells with few observations, I aggre gate workers into ten-year experience bands. Table II reports the calculated indices for each of the education groups. The occupa tion distributions of immigrants and natives with the same ex perience are generally more similar than the distributions of This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions 1346 QUARTERLY JOURNAL OF ECONOMICS TABLE II Index of Congruence in Occupation Distributions within Education Groups, 1990 Experience of corresponding immigrant group Education-experience of native groups: 1-10 years 11-20 years 21-30 years 31-40 years High school dropouts 1-10 years 0.709 0.714 0.671 0.619 11-20 years 0.525 0.631 0.628 0.585 21-30 years 0.410 0.527 0.567 0.566 31-40 years 0.311 0.435 0.496 0.518 High school graduates 1-10 years 0.682 0.611 0.498 0.405 11-20 years 0.279 0.379 0.387 0.338 21-30 years 0.030 0.184 0.297 0.272 31-40 years -0.035 0.126 0.276 0.311 Some college 1-10 years 0.649 0.571 0.474 0.291 11-20 years 0.147 0.401 0.492 0.336 21-30 years -0.052 0.230 0.432 0.407 31-40 years -0.066 0.217 0.458 0.489 College graduates 1-10 years 0.756 0.710 0.639 0.531 11-20 years 0.561 0.673 0.674 0.593 21-30 years 0.430 0.597 0.661 0.619 31-40 years 0.422 0.599 0.688 0.691 Equation (2) defines the index of congruence. The index is calculated separately for each pair of native and immigrant groups. immigrants and natives with different levels of experience. More over, the congruence index falls, the larger the disparity in work experience between the two groups. Consider the group of native workers who are high school dropouts and have eleven to twenty years of experience. The index of congruence with immigrants who have the same experi ence is 0.63. This index falls to 0.53 for immigrants who have 1 to 10 years of experience, and to 0.59 for immigrants with 31 to 40 years. Similarly, consider the native workers who are college graduates and have fewer than ten years of experience. The index of congruence with immigrants who have the same experience is 0.76, but this index falls to 0.71 for immigrants who have 11 to 20 years of experience, to 0.64 for immigrants who have 21 to 30 years, and to 0.53 for immigrants who have more than 30 years. In sum, the occupation distributions of immigrants and natives (for a given level of education) are most similar when one com This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions LABORMARKET IMPACT OF IMMIGRATION 1347 pares workers who have roughly the same level of work experience. IV. Basic Results Let yijt denote the mean value of a particular labor market outcome for native men who have education i (i = 1, . . . 94), experience./ (j = 1, . . . , 8), and are observed at time t (t = 1960, 1970, 1980, 1990, 2000). Much of the empirical analysis reported in this paper stacks these data across skill groups and calendar years and estimates the model:6 (3) Jijt = QPijt + Si + Xj + 17t+ (st X Xj) + (S?X TTt)+ (Xj X 7T,)+ (p#, where s? is a vector of fixed effects indicating the group's educa tional attainment, Xj is a vector of fixed effects indicating the group's work experience, and nt is a vector of fixed effects indi cating the time period. The linear fixed effects in equation (3) control for differences in labor market outcomes across schooling groups, experience groups, and over time. The interactions (s? X T?t) and (xj X nt) control for the possibility that the impact of education and experience changed over time, and the interaction (s? X Xj) controls for the fact that the experience profile for a particular labor market outcome differs across schooling groups. The dependent variables are the mean of log annual earn ings, the mean of log weekly earnings, and the mean of fraction of time worked (defined as weeks worked divided by 52 in the sample of all persons, including nonworkers). Unless otherwise specified, the regressions are weighted by the sample size used to 6. The generic regression of wages on some measure of immigrant penetra tion is used frequently in the literature. Suppose that the labor demand function in the preimmigration period is log wkt = Dkt + e \ogNkt + tp,where k is a skill group. The wage change resulting from an exogenous influx of immigrants is A log wkt = ADkt + E log [(Nkt(l + nkt) + Mkt)INkt\ + ? ? ADkt + e(nkt + mkt) + g, where nkt gives the percent change in the number of natives, and mkt = MktINkt. The rate of change nkt is determined by the native labor supply function, nkt = Skt + a A log wkt + |x. The reduced-form wage equation is A log wkt = Xkt + e*mkt + ?*, where Xkt = (ADkt + eSkt)/(l ?0-) and e* = e/(l ?ct). Equation (3) is a transformation of this reduced-form equation that approximately uses log mkt, rather than mkt, as the measure of immigrant penetration. In particular, log m ^ (M iV)/(0.5(M + N)) = 2(2p - 1). I opted for the immigrant share specification because the relation between wages and m is nonlinear and m has a large variance both over time and across groups. This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions 1348 QUARTERLY JOURNAL OF ECONOMICS TABLE III Impact of Immigrant Share on Labor Market Outcomes of Native Education-Experience Groups Dependent variable Log annual Log weekly Fraction of Specification: earnings earnings time worked 1. Basic estimates -0.919 -0.572 -0.529 (0.582) (0.162) (0.132) 2. Unweighted regression -0.725 -0.546 -0.382 (0.463) (0.141) (0.103) 3. Includes women in labor force counts -0.919 -0.637 -0.511 4. Includes log native labor force asregressor -1.231 -0.552 -0.567 (0.661) (0.159) (0.148) -1.231 -0.552 -0.567 (0.384) (0.204) (0.116) The table reports the coefficient of the immigrant share variable from regressions where the dependent variable is themean labormarket outcome for a native education-experience group at a particular point in time. Standard errors are reported in parentheses and are adjusted for clustering within education-experi ence cells. All regressions have 160 observations and, except for those reported in row 2, are weighted by the sample size of the education-experience-period cell.All regression models include education, experience, and period fixed effects, as well as interactions between education and experience fixed effects, education and period fixed effects, and experience and period fixed effects. calculate yijt. The presence of the education-experience interac tions in (3) implies that the impact of immigration on labor market outcomes is identified from changes that occur within education-experience cells over time. The standard errors are clustered by education-experience cells to adjust for possible se rial correlation. The first row of Table III presents the basic estimates of the adjustment coefficient 6. Consider initially the results when the dependent variable is the log of weekly earnings of native work ers. The coefficient is -0.572, with a standard error of 0.162. It is easier to interpret this coefficient by converting it to an elasticity that gives the percent change in wages associated with a percent change in labor supply. Let mijt = MijtINijt, or the percentage increase in the labor supply of group (i,j,t) attributable to immi gration. Define the "wage elasticity" as7 7. As noted above, the immigrant share approximates log m. Because there are no cells with zero immigrants in the data used in Table III, the results are virtually identical (once properly interpreted) if log m is used as the regressor. In the next section, however, where I categorize workers by state of residence, education, and experience, 15.7 percent of the cells have no immigrants, and using log m would create a serious selection problem. This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions LABORMARKET IMPACT OF IMMIGRATION 1349 d log Wijt = e dmijt (1+ mijtf ' By 2000, immigration had increased the number of men in the labor force by 16.8 percent. Equation (4) implies that the wage elasticity?evaluated at the mean value of the immigrant supply increase?can be obtained by multiplying 6 by approximately 0.7. The wage elasticity for weekly earnings is then -0.40 (or -0.572 X 0.7). Put differently, a 10 percent supply shock (i.e., an immigrant flow that increases the number of workers in the skill group by 10 percent) reduces weekly earnings by about 4 percent. Table III indicates that immigration has an even stronger effect on annual earnings, suggesting that immigration reduces the labor supply of native male workers. A 10 percent supply shock reduces annual earnings by 6.4 percent and the fraction of time worked by 3.7 percentage points. Note that the difference in the coefficients from the log annual earnings and the log weekly earnings regressions gives the coefficient from a log weeks worked specification. A simple supply-demand framework implies that the labor supply elasticity for workers can be estimated from the ratio of the immigration effect on log weeks worked and log weekly earnings. The point estimate for this ratio is 0.6. This estimate lies above the range reported by Juhn, Murphy, and Topel [1991], who report labor supply elasticities between 0.1 and 0.4.8 The remaining rows of Table III conduct a variety of specifi cation tests to determine the sensitivity of the results. The coef ficients reported in the second row, for example, indicate that the results are similar when the regressions are not weighted by the 8. The variable piJt gives the immigrant share among labor force partici pants. The labor force participation decision may introduce some endogeneity in this variable. The problem can be addressed by using an instrument given by the immigrant share in the population of all men in cell (ij,t). The IV estimates of 0 (and standard errors) are -0.803 (0.586) for log annual earnings, -0.541 (0.153) for log weekly earnings, and -0.493 (0.125) for the fraction of time worked. These coefficients are similar to those reported in the first row of Table III. The immi grant share may also be endogenous in a different sense. Suppose that the labor market attracts foreign workers mainly in those skill cells where wages are relatively high. There would be a spurious positive correlation between pijt and the wage. The results in Table III should then be interpreted as lower bounds of the true impact of immigration. Finally, the 2000 Census was released while this paper was in press. I reestimated the basic models to determine the sensitivity of the results when the 2000 CPS cross-section was replaced with the 2000 Census. The coefficients for the key specification reported in the first row are quite similar: -0.924 (0.462) for log annual earnings, -0.514 (0.203) for log weekly earnings, and -0.468 (0.077) for the fraction of time worked. This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions 1350 QUARTERLY JOURNAL OF ECONOMICS sample size of the skill group. In the third row the regression redefines the measure of the immigrant share pijt to include both male and female labor force participants. Despite the misclassi fication of many women into the various experience groups, the adjustment coefficients remain negative and significant, and have similar values to those reported in the first row. The last row of the table addresses the interpretation problem that arises be cause a rise inpijt can represent either an increase in the number of immigrants or a decline in the number of native workers in that skill group (e.g., the secular decline in the number of natives who are high school dropouts). Row 4 of the table reports the adjust ment coefficient when the regression adds the log of the size of the native workforce in cell (i9j9t) as a regressor. The wage elasticity for log weekly earnings is -0.39 and significant. In short, the parameter 6 in equation (3) is indeed capturing the impact of an increase in the size of the immigrant population on native labor market outcomes.9 I also estimated the regression model within schooling groups to determine whether the results are being driven by particular groups, such as the large influx of foreign-born high school dropouts. With only one exception, Table IV shows that the impact of immigration on the weekly earnings of particular schooling groups is negative and significant. The exception is the group of college graduates, where the adjustment coefficient is positive and has a large standard error. Note, however, that the regression estimated within a schooling group cannot include experience-period interactions to control for secular changes in the shape of the experience-earnings profile. As a result, the coefficient of the immigrant share variable may be measuring a spurious correlation between immigration and factors that changed the wage structure differentially within schooling groups. It is probably not coincidental that the adjustment coef ficient is positive for college graduates, the group that experi 9. The results would be roughly similar if the regressions were estimated separately using each set of two adjacent cross sections, so that the regression models would be differencing the data over a decade. The adjustment coefficients (and standard errors) for log weekly earnings are -1.042 (0.484) in 1960-1970, -0.427 (0.561) in 1970-1980, -0.277 (0.480) in 1980-1990, and -0.285 (0.270) in 1990-2000. This rough similarity contrasts with the inability of the spatial correlation approach to generate parameter estimates that even have the same sign over time; see Borjas, Freeman, and Katz [1997] and Schoeni [1997]. This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions LABORMARKET IMPACT OF IMMIGRATION 1351 TABLE IV Impact of Immigrant Share on Native Labor Market Outcomes, by Education Group High High At least school school Some College high school Dependent variable: dropouts graduates college graduates graduates 1. Log annual earnings -1.416 -2.225 -0.567 1.134 -1.184 (0.313) (0.622) (0.421) (0.436) (0.668) 2. Log weekly earnings -0.947 -2.074 -1.096 0.610 -0.335 (0.164) (0.510) (0.461) (0.440) (0.612) 3. Fraction of time worked -0.086 0.393 0.567 0.300 -1.040 (0.073) (0.251) (0.385) (0.499) (0.211) The table reports the coefficient of the immigrant share variable from regressions where the dependent variable is themean labormarket outcome for a native education-experience group at a particular point in time. Standard errors are reported in parentheses and are adjusted for clustering within experience cell (in the first four columns) and within education-experience cells (in the last column). All regressions are weighted by the sample size of the education-experience-period cell. The regressions reported in the first four columns have 40 observations and include experience and period fixed effects. The regressions reported in the last column have 120 observations and include education, experience, and period fixed effects, as well as interactions between education and experience fixed effects, education and period fixed effects, and experi ence and period fixed effects. enced perhaps the most striking change in the wage structure in recent decades.10 Finally, the last column of Table IV estimates the regressions using only the groups of natives with at least a high school education. The coefficients generally suggest that the sample of high school dropouts is not the group that is driving much of the analysis. Although the adjustment coefficients remain negative for all the dependent variables, it is insignificant for log weekly earnings. In the case of log annual earnings, however, the wage elasticity is around -0.8, suggesting that immigration had an adverse impact on native workers even when the regression ig nores the information provided by the workers who experienced the largest supply shock in the past few decades.11 10. I also estimated the regression model within experience groups. The adjustment coefficients (and standard errors) for log weekly earnings were 1-5 years of experience, -0.403 (0.470); 6-10 years, -0.358 (0.286); 11-15 years, -0.475 (0.285); 16-20 years, -0.555 (0.244); 21-25 years, -0.568 (0.244); 26-30 years, -0.634 (0.193); 31-35 years, -0.495 (0.288); and 36-40 years, -0.147 (0.228). Although these regressions only have twenty observations, the point estimate of 6 is negative and significant for many groups. 11. It is of interest to use the labor market outcomes of immigrants as the dependent variable. I used the sample of immigrants with fewer than 30 years of experience because there are relatively few observations in the cells for older workers in 1970 and 2000, and did not use data from the 1960 Census because This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions 1352 QUARTERLY JOURNAL OF ECONOMICS V. A Comparison with the Spatial Correlation Approach In contrast to the studies that calculate spatial correlations between wages in local labor markets and measures of immigrant penetration, the evidence presented in the previous section indi cates that immigrants have a sizable adverse effect on the wage of competing workers. This discrepancy suggests that itmight be instructive to examine how the results of the generic spatial correlation regression would change if that analysis defined skill groups in terms of both education and experience. Suppose that the relevant labor market for a typical worker is determined by his state of residence (r), education, and expe rience.12 I use the 1960-2000 Census and CPS files to calculate both the immigrant share and the mean labor market outcomes for cell (r,?j,?). I then use these aggregate data to estimate regressions similar to those presented above, but the unit of analysis is now a state-education-experience group at a particular point in time. Table V reports the estimated coefficient of the immigrant share variable from this regression framework. The first column of the table presents the coefficient from the simplest specifica tion, which includes the state, education, experience, and period fixed effects, as well as interactions between the state, education, and experience fixed effects with the vector of period fixed effects, and interactions between the state and education fixed effects. This regression, in effect, estimates the impact of immigration on the change in labor market outcomes experienced by a particular education group in a particular state. The adjustment coefficients for the various dependent variables are negative and mostly significant. The adjustment coefficient in the log weekly earnings regression is -0.124, with a standard error of 0.042. Note that the implied adverse impact of immigration resulting from this speci that survey does not provide information on the immigrant's year of entry into the United States. The estimates are imprecise, but the results resemble those found for native workers once I control for cohort and assimilation effects. If the regres sion is estimated on the sample of immigrants who have been in the United States for fewer than ten years, the adjustment coefficients (and standard errors) are -0.506 (0.398) for log annual earnings, -0.290 (0.350) for log weekly earnings, and -0.192 (0.105) for the fraction of time worked. 12. I use states to define the geographic boundary of the labor market because a worker's state of residence is the only geographic variable that is consistently coded across the entire 1960-2000 span. The 1960 Census does not report the person's metropolitan area of residence, and the metropolitan area identifiers for the 1970 Census differ substantially from those reported in later surveys. This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions LABORMARKET IMPACT OF IMMIGRATION 1353 TABLE V Impact of Immigrant Share on Labor Market Outcomes of Native State-Education-Experience Groups Dependent variable: (1) (2) (3) (4) 1. Log annual earnings -0.115 -0.276 -0.253 -0.217 (0.079) (0.053) (0.046) (0.068) 2. Log weekly earnings -0.124 -0.217 -0.203 -0.183 (0.042) (0.039) (0.038) (0.050) 3. Fraction of time worked -0.038 -0.100 -0.078 -0.119 (0.030) (0.015) (0.015) (0.021) Controls for: (State X period), (education X period), (experience x period), (state x education) fixed effects Yes Yes YesYes (State x education x experience) fixed effectsNo Yes Yes Yes (Education X experience X period) fixed effectsNo No Yes Yes (State x education x period), (state X experience X period) fixed effects No No NoYes The table reports the coefficient of the immigrant share variable from regressions where the dependent variable is themean labormarket outcome for a native state-education-experience group at a particular point in time. Standard errors are reported in parentheses and are adjusted for clustering within state-education experience cells. All regressions areweighted by the sample size of the state-education-experience-period cell and include state, education, experience, and period fixed effects. The regressions on log annual earnings or log weekly earnings have 8153 observations; the regressions on the fraction of time worked have 8159 observations. fication is far smaller than the effects reported in the previous section. The second column of Table V adds a three-way interaction between the state, education, and experience fixed effects. This specification, therefore, examines the impact of immigration on the wage growth experienced by a particular education-experi ence group living in a particular state. The adjustment coeffi cients are more negative (-0.217 in the log weekly wage specifi cation) and statistically significant. In short, defining a skill group in terms of both education and experience implies that immigration has a more adverse impact than a specification that ignores the experience component. The third column of the table further expands the model by allowing for period effects to vary across education-experience cells, while the fourth column presents the full specification of the regression that allows for all possible three-way interactions be tween the state, education, experience, and period fixed effects. This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions 1354 QUARTERLY JOURNAL OF ECONOMICS This regression specification effectively identifies the wage im pact by using only variation in immigration at the (state X education X experience X period) level. The coefficient is negative and significant (-0.183 in the log weekly wage specification), and it is numerically much smaller than the coefficients reported in the previous section. In fact, it is instructive to contrast the difference in the results reported in the last column of Table V with the evidence reported in Table III. The key difference between the two sets of estimates is the assumption made about the geographic boundary of the labor market. The estimated wage elasticity for log weekly earnings is -0.13 when a state's geographic boundary limits the size of the market, and -0.40 when the worker participates in a national market. One interesting interpretation of this discrep ancy is that there is sufficient spatial arbitrage?perhaps due to interstate flows of labor and capital?that tends to equalize op portunities for workers of given skills across regions. The spatial arbitrage effectively cuts the national estimate of the impact of immigration by two-thirds.13 Put differently, even though immi gration has a sizable adverse effect on the wage of competing workers at the national level, the analysis of wage differentials across regional labor markets conceals much of the impact. VI. Refining the Definition of Skills VI.A. Measuring Effective Experience Up to this point, labor market experience has been defined as the time elapsed since entry into the labor market for both im migrants and natives. The evidence indicates that U. S. firms 13. The smaller wage effects estimated at the state level could also be due to attenuation bias from the measurement error that arises when I calculate the immigrant supply shock at such a detailed level of disaggregation. I reestimated the model using the nine Census regions (rather than states) as the geographic unit. The region-level regression coefficients corresponding to the last column of Table V are -.346 (.096) in the log annual earnings regression, -.289 (.070) in the log weekly earnings regression, and -.057 (.023) in the fraction of time worked regression. Even though the coefficients in the annual and weekly earnings regressions are numerically larger than those obtained in the state-level analysis, the coefficient in the log weekly earnings regression is still only half the size of the one reported in Table III. Moreover, it is unclear if the relatively larger effects estimated at the region level result from the partial elimination of attenuation bias or from the possibility that some of the native flows induced by immigration are intraregional, and hence the region is a slightly better conceptual represen tation of the "closed market" required for measuring the local impact of immigra tion; see Borjas, Freeman, and Katz [1996] for related evidence. This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions LABORMARKET IMPACT OF IMMIGRATION 1355 attach different values to experience acquired abroad and expe rience acquired in the United States [Chiswick 1978]. These findings suggest that one should use the "effective experience" of an immigrant worker before assigning that worker to a particular schooling-experience group, where effective experience measures the years of work exposure that are valued in the U. S. labor market. Let A denote age, Am the age of entry into the United States, and AT the age of entry into the labor market. The years of effective experience for an immigrant worker are given by m v- ? AT (5) A-\7(A-AT), ifAm AT) into the equivalent value of experience acquired by a native worker, ? rescales the value of a year of U. S. experience acquired by these adult immigrants, and 7 rescales the experience acquired by immigrants who migrated as children (i.e., Am < AT). The parameters a, ?, and 7 can be estimated by using the standard model of immigrant assimilation, a model that also accounts for differences in immigrant "quality" across cohorts [Borjas 1985]. Suppose that we pool data for native and immi grant workers in two separate cross sections (such as the 1980 and 1990 Censuses). A generic regression model that can identify all of the relevant parameters is (6) \ogw = si + $cIc+ $DID+ \NN{A-AT) + \CIC(A AT) + \D0ID(Am AT) + \D1ID(A Am) + k7 + P7T+ cp, where w gives the weekly wage of a worker observed in a particu lar cross section, s? gives a vector of education fixed effects, Ie indicates whether the immigrant entered the country as a child, ID indicates whether the immigrant entered as an adult, N indi cates whether the worker is native-born (N = 1 - Ie ID), Y gives the calendar year of entry into the United States (set to zero for natives), and tt indicates whether the observation is drawn from the 1990 Census. The coefficient KN gives the market value of a year of expe rience acquired by a native worker; Kc gives the value of a year of experience acquired in the United States by a "child immigrant"; and XD0 and XD1 give the value of a year of source country experience and of U. S. experience acquired by an adult immi This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions 1356 QUARTERLY JOURNAL OF ECONOMICS grant, respectively. The weights that define an immigrant's effec tive experience are (7) a = T_'? = Y_'^ = \~\N \N Ajv Although the generic regression model in (6) is pedagogically useful, it ignores the curvature of the experience-earnings profile, and also ignores the possibility that the returns to education differ among the various groups. Further, it is preferable to define the calendar year of an immigrant's arrival as a vector of dummy variables indicating the year of arrival, rather than as a linear time trend. I estimated this more general model using the pooled 1980 and 1990 data. Table VI reports the relevant coefficients from this regression. The experience coefficients for natives and for immigrants who migrated as children have almost identical numerical values, so that a marginal year of experience is valued at the same rate by employers (although the tiny numerical difference is statisti cally significant). This implies that the weight 7 is estimated to be 1.0. In contrast, the value of an additional year of source country experience for adult immigrants (evaluated at the mean years of source country experience) is 0.006, while the value of an addi tional year of U. S. experience for these immigrants is 0.024. The value of a year of experience for a comparable native worker is 0.015. The implied weights are a = 0.4 and ? = 1.6. I used these weights to calculate the effective experience of each immigrant, and then reclassified them into the schooling experience cells using the predicted measure of effective experi ence.14 The top row of Table VII reports the estimated adjustment coefficients. The effects are roughly similar to those reported in the previous section. For example, the weekly earnings regression implies that the wage elasticity is -.30, and the effect is statis tically significant. 14. Neither the Census nor the CPS reports the exact year in which immi grants entered the United States, but instead reports the year of entry within particular intervals (e.g., 1980-1984). I used a uniform distribution to randomly assign workers in each interval to each year in the interval. Because the immi grant's year of arrival is not reported in the 1960 Census, the analysis is restricted to data drawn from the 1970 through 2000 cross sections. This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions LABORMARKET IMPACT OF IMMIGRATION 1357 TABLE VI Impact of Different Types of Labor Market Experience on the Log Weekly Earnings of Natives and Immigrants Group Child Adult Coefficient of: Natives immigrants immigrants Source country experience ? ? 0.012 (0.001) Source country experience squared - 10 ? ? -0.003 (0.000) U. S. experience 0.056 0.058 0.032 (0.000) (0.001) (0.002) U. S. experience squared - 10 -0.010 -0.010 -0.004 (0.001) (0.000) (0.001) Mean value of: Source country experience ? ? 10.6 U. S. experience 16.7 13.0 10.8 Marginal value of an additional year of experience for immigrants: Source country experience ? ? 0.006 (0.001) U. S. experience ? 0.033 0.024 (0.001) (0.001) Marginal value of an additional year of experience for natives, evaluated at mean value of relevant sample of immigrants ? 0.031 0.015 (0.000) (0.000) Standard errors are reported in parentheses. The regression pools data from the 1980 and 1990 Census and has 1,141,609 observations. The dependent variable is the log ofweekly earnings. The regressors include dummy variables indicating whether the worker is an adult immigrant or a child immigrant; a vector of variables indicating the worker's educational attainment, interacted with variables indicating whether the worker is an adult or a child immigrant; experience (and its square) for native workers; experience (and its square) for immigrants who arrived as children; source country experience (and its square) for immigrants who arrived as adults; experience in theUnited States (and its square) for immigrants who arrived as adults; dummy variables indicating the calendar year in which the immigrant arrived (1985-1989, 1980-1984, 1975-1979, 1970-1974, 1965-1969, 1960-1964, 1950-1959, and before 1950), and the interaction of this vector with a dummy variable indicating whether the immigrant arrived as an adult; and a dummy variable indicating whether the observation was drawn from the 1990 Census. VLB. Measuring Effective Skills The notion of effective experience raises amore general ques tion about the overall comparability of the skills of immigrants and natives. The U. S. labor market differentiates the value of human capital embodied in immigrants and natives along many dimensions. For example, the value that firms attach to schooling This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions 1358 QUARTERLY JOURNAL OF ECONOMICS TABLE VII Impact of Immigrant Share on Labor Market Outcomes of Native Skill Groups, Using Effective Experience and Effectpte Skills Dependent variable Log Log annual weekly Fraction of Specification: earnings earnings time worked 1. Effective experience -1.025 -0.422 -0.611 (0.506) (0.210) (0.118) 2. Using quantiles of wage distribution - 0.562 - 0.606 - 0.048 (0.329) (0.158) (0.167) The table reports the coefficient of the immigrant share variable from regressions where the dependent variable is themean labormarket outcome for a native skill group (defined in terms of education-experience in row 1 or education-quantile in row 2) at a particular point in time. The quantile definition of skill groups is based on theworker's placement in each of twenty quantiles of the (within-education) native weekly wage distribution. Standard errors are reported in parentheses and are adjusted for clustering within education experience cells (row 1) orwithin education-quantile cells (row2).All regressions areweighted by the sample size of the education-experience-period cell (row 1) or the education-quantile-period cell (row 2). The regres sions reported in row 1have 128 observations; those reported in row 2 have 400 observations. The models in row 1 include education, experience, and period fixed effects, as well as interactions between education and experience fixed effects, education and period fixed effects, and experience and period fixed effects. The models in row 2 include education, quantile, and period fixed effects, as well as interactions between education and quantile fixed effects, education and period fixed effects, and quantile and period fixed effects. will probably differ between the two groups, as well as among immigrants originating in different countries. It is of interest, therefore, to devise a simple way of summarizing the differences in "effective skills" that exist between immigrants and natives within a schooling category. It seems sensible to assume that similarly educated workers who fall in the same general location of the wage distribution have roughly the same number of effi ciency units because employers attach the same value to the entire package of skills embodied in these workers. To conduct this classification of workers into skill groups, I restrict the analysis to workers who have valid wage data. In each cross section and for each of the four schooling groups, I sliced the weekly wage distribution of native workers into twenty quantiles. By construction, 5 percent of natives in each schooling group fall into each of the quantiles. I then calculated how many of the immigrant workers in each schooling group fall into each of the twenty quantiles. The immigrant supply shock is defined by (8)pikt = Miktl{Mikt + Nikt), where Mikt and Nikt give the number of foreign-born and native This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions LABOR MARKET IMPACT OF IMMIGRATION 1359 born workers in schooling group i9 quantile k (k = l9 . . . 920)9 at time t. Consider the regression model: (9) Jikt = Qpikt+ Si + qk + >nt+ (qkX Si) + (s?X irt) + (qkX it,) + yikt9 where qk is a vector of fixed effects indicating the quantile of the cell. The second row of Table VII reports the adjustment coeffi cients estimated from this specification of the model. Despite the very different methodological approach employed to define the skill groups, the estimated coefficient in the log weekly earnings regression is similar to those reported above. The estimate of 0 is -0.606 (with a standard error of 0.158), implying a wage elastic ity of -0.42. In sum, the evidence suggests that the clustering of immigrants into particular segments of the wage distribution worsened the wage outcomes of native workers who happened to reside in those regions of the wage distribution.15 VII. A Structural Approach to Immigration and Factor Demand VILA. Theory and Evidence Up to this point, I have not imposed any economic structure in the estimation of the wage effects of immigration. As inmost of the studies in the spatial correlation literature, I have instead attempted to calculate the correlation that indicates whether an increase in the number of immigrants lowers the wage of com peting native workers. An alternative approach would impose more structure by specifying the technology of the aggregate production function.16 This structural approach would make it possible to estimate not only the effect of a particular immigrant influx on the wage of 15. The fraction of time worked variable used in the regression reported in the second row of Table VII has a different definition than elsewhere in this paper. To simplify the sorting of persons into the quantiles of the wage distribution, I restricted the analysis to working men. One could classify nonworkers into the various quantiles by using a first-stage regression that predicts earnings based on a person's educational attainment, experience, and other variables. For native men this approach leads to results that are similar to those reported in the text. 16. Early empirical studies of the labor market impact of immigration [Gross man 1982; Borjas 1987] actually imposed a structure on the technology of the local labor market, such as the translog or the Generalized Leontief, and used the resulting estimates to calculate the various substitution elasticities. Although this approach fell out of favor in the early 1990s, the evidence reported by Card [2001] and the results presented in this section suggest that the structural approach may be due for a timely comeback. This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions 1360 QUARTERLY JOURNAL OF ECONOMICS competing native workers, but also the cross effects on the wage of other natives. An empirically useful approach assumes that the aggregate production function can be represented in terms of a three-level CES technology: similarly educated workers with dif ferent levels of work experience are aggregated to form the effec tive supply of an education group; and workers across education groups are then aggregated to form the national workforce.17 Suppose that the aggregate production function for the na tional economy at time t is (10) Qt = [\KtK?t + \LtLvt-\v\ where Q is output, K is capital, L denotes the aggregate labor input; and v = 1 - 1/a^x, with vKL being the elasticity of substitution between capital and labor (- < y < 1). The vector X gives time-variant technology parameters that shift the produc tion frontier, with X^ + \Lt = 1. The aggregate Lt incorporates the contributions of workers who differ in both education and experience. Let (ID Lt = 2 e*L& i/p where Lit gives the number of workers with education i at time t, and p = 1 l/o-#, with vE being the elasticity of substitution across these education aggregates (-<*> < p < 1). The Qit give time-variant technology parameters that shift the relative pro ductivity of education groups, with X? d?t = 1. Finally, the supply of workers in each education group is itself given by an aggrega tion of the contribution of similarly educated workers with differ ent experience. In particular, -ii/n (12) Llt= Sa^LJ, j where Lijt gives the number of workers in education group i and experience group,/' at time t9 and r\ = 1 l/ax, with gx being the elasticity of substitution across experience classes within an edu cation group (-o? < r\ < 1). Equation (12) incorporates an impor 17. The three-level CES technology slightly generalizes the two-level ap proach used in the labor demand context by Bowles [1970] and Card and Lemieux [2001]. This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions LABORMARKET IMPACT OF IMMIGRATION 1361 tant identifying assumption: the technology coefficients atj are constant over time, with 2, a?J = 1. The marginal productivity condition implies that the wage for skill group (i,j,t) is (13) log wijt = log \Lt + (1 v) log Qt + (v- p) log Lt + log Qit + (p ti) log Lit + log ax + On - 1) log Lijt. As Card and Lemieux [2001] show in their recent study of the link between the wage structure and cohort size, it is straightfor ward to implement this approach empirically. In particular, note that the marginal productivity condition in (13) can be rewritten as (14) log wijt = bt + 8??+ 8y (I/o*) log Lijt, where 8? = log \Lt + (1 v) log Qt + (v p) log Lt, and is absorbed by period fixed effects; 8?? = log dit + (p j]) log Lit, and is absorbed by interactions between the education fixed ef fects and the period fixed effects; and 8?J = log a?j, and is absorbed by interactions between education fixed effects and experience fixed effects. The regression model in (14), therefore, identifies the elasticity of substitution across experience groups. Moreover, the coefficients of the education-experience inter actions in (14) identify the parameters log a?J. I impose the restriction that S7 a77 = 1 when I estimate the a77 from the fixedJ lJu effect coefficients. As indicated by equation (12), the estimates of ay and crx permit the calculation of Lit9 the CES-weighted labor aggregate for education group i. I can then move up one level in the CES technology, and recover an additional unknown parameter. Let log wit be the mean log wage paid to the average worker in education group i at time t. The marginal productivity condition determining the wage for this group is (15) log wlt = 8, + log Qit (l/dE) log Lit. This equation is closely related to the model estimated by Katz and Murphy [1992, p. 69] that examines how the wage differen tial between college and high school graduates varies with rela tive supplies. Note that vE cannot be identified if the regression included interactions of education-period fixed effects to capture 18. If log a?J is an estimated fixed effect coefficient, then a?j = expdog a^)/ ?j expdog ?ij). This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions 1362 QUARTERLY JOURNAL OF ECONOMICS the term log 6??. There would be twenty such interaction terms, but there are only twenty observations in the regression (four education groups observed at five different points in time). To identify aE9 I adopt the Katz-Murphy assumption that the tech nology shifters can be approximated by a linear trend that varies across education groups. It is important to note that ordinary least squares regres sions of equations (14) and (15) may lead to biased estimates of vx and dg because the supply of workers to the various education groups is likely to be endogenous over the 40-year period spanned by the data. The economic question at the core of this paper, however, suggests an instrument for the size of the workforce in each skill group: the number of immigrants in that group. In other words, the immigrant influx into particular skill groups provides the supply shifter required to identify the labor demand function. This instrument would be valid if the immigrant influx into particular skill groups were independent of the relative wages offered to the various skill categories. It is likely, however, that the number of immigrants in a skill group responds to shifts in the wage structure. Income-maximizing behavior on the part of potential immigrants would generate larger flows into those skill cells that had relatively high wages. This behavioral response would tend to build in a positive correlation between the size of the labor force and wages in a skill group. The regression coeffi cients, therefore, understate the negative wage impact of a rela tive supply increase.19 The three-level CES technology offers a crucial advantage for estimating the impact of immigration within a structural system of factor demand. My analysis defines 33 factors of production: 32 education-experience skill groups plus capital. A general specifi cation of the technology, such as the translog, would require the estimation of 561 different parameters (or n(n + l)/2). The 19. Consider the regression model given by log it; = ? log L 4- u. The IV estimate of ? has the property: cov (log M, u) p P P cov (logM, logL) ' where logM is the instrument. The total number of workers in a skill group is, in fact, positively correlated with the number of immigrants in that group, so that cov (log M, log L) > 0. Further, cov (log M, u) > 0 because skill cells with favorable demand shocks will probably attract larger numbers of income-maxi mizing immigrants. The IV regression coefficient then provides a lower bound for the wage reduction resulting from a supply increase. This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions LABORMARKET IMPACT OF IMMIGRATION 1363 three-level CES approach drastically reduces the size of the pa rameter space; the technology can be summarized in terms of three elasticities of substitution. Obviously, this simplification comes at a cost: the CES specification restricts the types of sub stitution that can exist among the various factors. The elasticity of substitution across experience groups takes on the same value for workers in adjacent experience categories as for workers who differ greatly in their experience; the elasticity of substitution between high school dropouts and high school graduates is the same as that between high school dropouts and college graduates; and the elasticity of substitution between capital and labor is the same for all the different types of workers. Finally, note that the empirical implementation of the three level CES technology described above does not use any data on the aggregate capital stock, making it difficult to separately iden tify the value o? (jKL.20 Iwill discuss below a plausible assump tion that can be made about this parameter to simulate the impact of immigration on the labor market. The first step in the empirical application of the model is to estimate equation (14) using the sample of 160 (ij,t) cells. The IV estimate of this regression equation is21 (16) log wijt = 8, + bit + 8^ - 0.288 log Lijt. (0.115) The implied elasticity of substitution across experience groups is 3.5. This estimate of vx is similar to the Card-Lemieux [2001] estimate of the elasticity of substitution across age groups. The Card-Lemieux estimates for U. S. data range from 3.8 to 4.9. 20. In principle, the elasticity vKL could be estimated even without direct information on the aggregate capital stock by going up an additional level in the CES hierarchy. This exercise yields the marginal productivity condition for the average worker at time t. This marginal productivity condition depends on a time fixed effect and on Lt, the CES-weighted aggregate of the workforce. The coeffi cient of Lt identifies -Hvkl- However, this regression would only have five observations in my data, and Iwould need to find a variable that could proxy for the movements in the period fixed effects. 21. The instrument is log Mijt and the standard errors are clustered by education-experience group. To avoid introducing errors due to composition ef fects, the regressions reported in this section use the mean log weekly wage of native workers as the dependent variable. The results would be very similar if the mean log wage was calculated in the pooled sample of natives and immigrants. The relevant coefficients (and standard errors) in equations (16), (17), and (17') would be -0.281 (0.059), -0.676 (0.518), and -0.680 (0.462), respectively. The regressions estimated in this section are weighted by the size of the sample used to calculate the cell mean on the left-hand side. This content downloaded from 147.251.185.127 on Mon, 9 Mar 2015 08:44:18 AM All use subject to JSTOR Terms and Conditions 1364 QUARTERLY JOURNAL OF ECONOMICS I use the implied estimate of the elasticity of substitution and the (transformed) coefficients of the education-experience fixed effects to calculate the size of the CES-weighted labor aggregate for each education group. I then estimate the marginal produc tivity condition for the education group given by (15). The IV regression estimate is22 (17) \ogwit = ?t + linear trend interacted with education fixed effects -0.741 log L*. (0.646) Alternatively, I can bypass the calculation of the CES-weighted labor aggregate for each education group, and simply use the actual number of workers in the group (L*t). The IV regression estimate is (170 log 1^ = 8, + linear trend interacted with education fixed effects -0.759 log Lf,. (0.582) Both specifications imply that