Education, cost of living and regional wage inequality in Brazil

Education, cost of living and regional wage inequality in Brazil Carlos R. Azzoni 1 Luciana M. S. Servo 2 Abstract The objective of this paper is to analyze wage inequality among the 10 largest metropolitan regions in Brazil in the 90 s. We assess to what extent worker characteristics (education, age, gender, race, position in the family) and job characteristics (occupational position, sector, experience) can explain wage inequality. The analysis is made both with regional-nominal and with regional-real wage data. In the first case a national price index is used to deflate the data; in the second case regional price indexes are added to the national price index in order to control for differences in cost of living among regions. Wage differentials in Brazil were slightly reduced when the controls were introduced, but the remaining inequality was still high. Considering the differing cost of living levels, the results indicate that they do have a role in explaining wage inequality in Brazil, but even after controlling for that factor, the remaining regional differentials are still important 1. Introduction 3 Brazil is well known for its high levels of regional inequality. Being a country with a large territory, that should not be surprising. The Northeast region of Brazil hosted 28% of Brazilian population in the year 2000 and produced only 13% of Brazilian GDP in the year 1998; the rich Southeast region hosted 43% of Population and produced 58% of GDP. Per capita income in the Northeast was 54% below the national average whilst in the Southeast it was 36% above that level. The poorest state, Piauí, in the Northeast region, had a per capita income level 5.6 times lower than the richest state, São Paulo, in the Southeast region 4. The level of regional income inequality is not only high but the trends are not promising. The above relative figures are not too different from the situation 1 Professor of Economics, Universidade de São Paulo, Brazil (cazzoni@usp.br) 2 Economist, IPEA Instituto de Pesquisas Econômicas e Sociais, Ministry of Planning, Brazil (lmsantos@ipea.gov.br) 3 This paper is based on Mrs. Luciana Servo master thesis, presented at the Department of Economics of the Universidade de São Paulo (Servo, 1999). The study was developed under the supervision of Professor Carlos Azzoni 4 See www.ibge.gov.br/ibge/estatistica/economia/contasregionais/, for information on regional income for Brazil, and http://www.ibge.gov.br/ibge/estatistica/populacao/censo2000/ for information on population.

half a century ago, for in 1937 the per capita income relation between Piauí and São Paulo states was 5 to 1. New sets of data have been released recently that allow for the application of econometric techniques to analyze the problem. It has been noted that inequality levels could be augmenting in recent times. A polarization reversal process starting in the early 70 s lasted until the mid 80 s, being replaced by a situation in which inequality is at most stable, depending on the variable considered (Haddad, 1999; Azzoni et all, 2000; Azzoni, 2001; Ferreira, 2000). In this study we work with ten Brazilian metropolitan areas in three different years: 1992, 1995 and 1997. The consideration of different moments in time is necessary for the Brazilian economy has experienced huge fluctuations in growth and inflation in the 90 s, as can be observed in Figures 1 and 2. The first year coincides with an inflation level around 20% per month and a negative growth rate for national GDP; the second features a situation in which inflation was lower and GDP growth rates were positive; the third year portrays an even lower inflation rate and a GDP growth rate similar to the previous year. Different macroeconomic scenarios could influence the dispersion of wage income levels across regions, for periods of growth and depression could be differentiated in space over time 5. Moreover, labor markets are heavily regulated in Brazil and a great variety of wage legislation was observed in that period. Thus, we consider three points in time in order to take care of these factors. This paper draws on a study done for the 80 s comparing the same metropolitan regions (Savedoff, 1990), allowing for the consideration of wage inequality over a two-decade period. Although the database and the methodology applied was not the same we apply here, it will be interesting to compare the results and verify if inequality has varied significantly between the two decades. The next section presents a short discussion of theoretical models that deal with the problem. Section 3 provides information on the database used and the econometric methodology applied to the data. The results are presented and commented on in Section 4. The last section provides concluding comments. 5 Azzoni (2201) shows that periods of rapid economic growth in Brazil were associated with increasing regional inequality, while periods of recession were associated with decreasing inequality.

80 70 79.1% Brazilian monthly inflation rates 60 50 50.8% 40 30 20 10 0 10 Brazilian annual GDP growth rates (% per year) 8 6 4 2 0-2 -4-6

2. Theoretical considerations If factors are free to move between regions, why should workers accept less income in one region if they could get more in another area? Capital is also expected to move, from high wage regions to low wage regions, in order to profit from this advantage, therefore contributing to wage equalization. Market imperfections, such as lack of information on wage differences, legal barriers to migration, cultural and institutional differences, etc. would be an answer: firms and individuals would not be aware of different opportunities. Another factor is the cost of movement, an aspect that could be important in a large country like Brazil. The observed wage differentials would reflect the cost of moving from one region to the other, that is, after discounting for that cost, individuals would not get any improvement from moving 6. In the Brazilian case, factors are free to move between regions. As a matter of fact, there are even tax incentives to move capital to the poor regions. Labor migration from poor to rich regions is part of the Brazilian scenario after 1940, with a large contingent of immigrants moving to the wealthiest areas of the country. It is true that distances are very high. For instance, between the city of Porto Alegre, in the South, and Belém, in the North, there is a distance of over 2.500 miles. Between São Paulo City, a traditional recipient of migrants, and Fortaleza, a traditional supplier, the distance is over 1.700 miles (see Map 1). However, the transportation system has progressed significantly in the last five decades, integrating all regions but the North (Amazon rain forest). This has reduced the cost of movement, and therefore increased the accessibility for workers to move from one region to another, but apparently that did not change significantly regional inequality in the country. Regional wage differences can be taken as indicative of the degree of integration of regional markets within a country. However, taking wage differentials to indicate market integration is complicated, for wages are subject to a variety of factors, such as differences in the composition of labor force, amenity levels in different regions, differences in cost of living etc. These differences could produce great inequality in nominal wages even in a highly integrated market, or they could indicate equalization in a non-integrated market 7. 6 Even in advanced countries regional inequalities may prevail. For example, Weiler (2000) finds evidence that persistent local unemployment in some US counties appear to be due to non-clearing labor markets. 7 Robertson (2000) studies labor market integration between US and Mexico and finds that those markets are closely integrated, in spite of a large wage differential between them. They are integrated in the sense that they respond simultaneously to common shocks.

Three different sources for regional labor differentials can be found in the literature (Maier and Weiss, 1986). According to the Theory of Human Capital, they reflect differences in the average level and in the distribution of human capital among workers, or differences in other worker characteristics, such as gender, race, etc 8. The theoretical framework was developed to analyze personal wage inequalities by Mincer (1958 and 1974), Becker (1964), and Becker and Chiswick (1966). The idea is that individuals invest in education and on-the-job-training because they expect to increase future returns from their participation in the labor market. They would only invest in education if the present value of the expected returns after education were greater than the present value of returns without education. The role of human capital expenditures and labor mobility as forces behind the process of income equalization within a country is analyzed in Razin and Yuen (1997). The Theory of Segmentation and the Theory of Polarized Development provide another interpretation for regional wage inequalities. Within this line of reasoning, inequalities are due to differences in the productive structure, institutional framework, access to technology and the level of development of regions. According to this interpretation, the Theory of Human Capital assigns no role for the demand for labor in determining wages and ignores any restrictions to labor mobility existing in segmented markets. A third line of explanation concentrates on compensating differences (Nordhaus and Tobin, 1972; Rosen, 1986). According to this interpretation, regional wage differentials are required to equalize monetary advantages and disadvantages of sectors, and regions, among workers. Factors as weather conditions, crime rates and pollution are some of the non-pecuniary factors that could explain regional wage differentials. We understand that all three sources are useful in explaining regional wage differentials, the same point of view of the previous study we are replicating (Savedoff, 1990). In order to analyze regional wage differentials we use a multivariate linear model, with wage as the dependent variable and category dummies as independent variables. This sort of model is suited for the analysis, for it identifies different average wage levels for each statistically significant category. Two models were estimated: 1) regional dummies only, with no consideration for personal and job characteristics; 2) regional dummies, worker and job characteristics. Model 1) is taken only as a point of reference, for the dummy variables coefficients indicate the wage differences among regions without controlling for any of the factors cited by the theoretical models described above. It is equivalent to a simple comparison of the wage level of each region to the average wage level for all regions. The idea is to verify how 8 Trejo (1997) shows that Mexican Americans earn low wages primarily because they possess less human capital than other workers. More than three-quarters of the wage gap is attributable to Mexican American workers relative youth, English language deficiencies and especially their lower educational attainment.

regional wage differentials, as measured by the differences in regional dummies, vary, as controls variables are included. If the introduction of control variables reduces the coefficients for the regional dummies, part of the wage differentials detected in model 1) is in fact due do the different regional distributions of the control variables. 3. Data base and econometric model This study is based on samples of workers of the ten largest metropolitan regions of Brazil, living in urban areas, age 18-65, employed, working more than 20 hours a week, for the years 1992, 1995 and 1997. The surveys are conducted yearly by the Brazilian official statistics agency (IBGE) and are highly recognized in the country as a sound and consistent data source 9. The ten metropolitan regions account for almost 50 million people in 1997, out of a national population around 160 million (see Map 1). The samples used in this study, in each of the three years, are over 37.000 people, corresponding to a universe of 14 million people (Table 1 provides detailed data on the samples). We work with the expanded sample, adopting the sampling weights used by IBGE to design the survey. For this study we take the hourly wage in the main job of the worker, controlling for the number of hours worked (always over 20h/week). We estimate traditional mincerian wage equations, in which the logarithm of wage is to be explained by the variables suggested by the above theoretical explanations 10. Two models are estimated: model 1: log( Yh ) = α + β. RD + ε model 2: log( Yh ) = α + β. RD + φ. PD + ϕ. JD + ε With Yh being the hourly wage, RD the regional dummies, PD the set of dummies for the personal characteristics of the workers and JD the set of dummies for job characteristics. The dummy variables utilized are: a) 10 metropolitan regions the estimated coefficients for each region indicate their labor income level in relation to the average of the ten metropolitan regions; 9 PNAD Pesquisa Nacional por Amostra de Domicílios (National Survey on a Sample of Households), provided by the Brazilian Statistics Institute, IBGE Instituto Brasileiro de Geografia e Estatística. For more details, visit www.ibge.gov.br/ibge/estatistica/populacao/trabalhoerendimento/pnad99/ 10 Two factors recommend the use of such a specification. Within the Theory of Human Capital framework, in a very simplified model, let E o be the yearly income of a person without any schooling, E s be the income of a person with s years of schooling and r be the rate of return to schooling. Then, E s = E 0 e rs, and ln E s = E 0 + rs. Besides that, since the wage distribution is log normal, the residuals present a log linear function, what makes the semi logarithm specification the most adequate for wage studies. If a linear specification were chosen, the estimator would be inefficient.

b) 7 education levels (less than 1 year; 1-3; 4; 5-7; 8; 9-11; 12 or more); c) 5 age levels (18-24; 25-34; 35-44; 45-54; 55-65) according to the Theory of Human Capital, it is expected that income levels will increase as the number of years of education increase; d) 5 age levels (18 to 24; 25-30; 35-44; 45-50, 55-65) according to the Theory of Human Capital, it is expected that increases in income occur up to a certain age only, with a decline thereafter; e) 2 genders; f) 2 races (white-non white); e) 2 roles in the household (head-non head); g) 5 job situations (formal employee - with a labor contract and full benefits; informal employee - no labor contract signed; public worker; autonomous worker business owner without employees; employer); h) 7 sectors (manufacturing; construction; commerce; services; transportation and communication; public administration; social and other); i) 2 sector situations (public and private); j) 2 experience situations (less than 2 years in the job; two years and more) 11. We estimate the models using Ordinary Least Squares. In order to grant full-rank, we could either take one metropolitan region as a reference or estimate the model subject to a number of restrictions that corresponds to the number of linear dependencies. If we used the first alternative, the regional dummy coefficients would indicate differences in relation to the dropped region. That usually makes it complicated to make comparisons of different regions at the same time. On the other hand, the use of Restricted Least Squares requires an a priori choice of a restriction (or of restrictions), therefore encompassing some arbitrariness. In the present case we use a restriction frequently employed in the econometric literature, that is, we center the estimated coefficients, forcing that each coefficient indicates deviations around the average for that variable 12. In summary, if the model is correctly specified, the dummy coefficients, including those for the metropolitan regions, indicate an estimate of the labor income gain (loss) if the worker moved from one category to another. In the regional case, they show how much the worker gains (loses) in moving from one region to another. As mentioned before, there are ten dummy coefficient sets, one for each explanatory variable; each set is composed of different categories 13. The restriction for each dummy coefficient set i is given by 11 See table 1 for a breakdown of the samples among the different categories. 12 Greene and Seaks (1991) demonstrate that the Restricted Least Squares estimation applying the same restriction we use in this study is sufficient for the dummy coefficients to represent deviations around the mean. Centering the coefficients also solves the problem of the adequate estimation of the coefficient standard deviations. For a discussion, see Haisken-DeNew and Schmidt (1997). 13 For example, for the explanatory variable Gender we have a two-category set: Male and Female.

' R. = 0, where i B i R i ri,1 ri,2 =... ri, j and B i bi,1 bi,2 =... bi, j, with ni, j ri, j = and r i, j = 1 (1) n j i, j j R i and B i are, respectively, a column vector (j x 1) of weights for that dummy set and a column vector (j x 1) of the estimated coefficients for the same dummy set. The weights take into account the importance of each category within the set (n ij is the number of observations of category j in dummy group i). Summarizing, the models are estimated using Restricted Least Squares with i restrictions, and the estimated coefficients indicate deviations from the mean. In order to analyze the dispersion of the coefficients within each set in different years, we use the Weighted Adjusted Standard Deviation WASD, given by equation (2) below. It indicates the dispersion of the estimated dummy coefficients within each set, adjusting for their variance, and weighting by the weights of the variable categories. WASD ( Bi ) = R i. Φ ( Bi ) R i. Ψ ( Var ( Bi ) (2) In equation (2), R i and B i are the same as in restriction (1), Φ(B i ) is a (j x j) diagonal matrix, the main diagonal elements being the elements of vector B i and Ψ(Var(Bi)) is a column vector (j x 1), whose elements are the same as in the main diagonal of the coefficients variance-covariance matrix Var (B i ). This method has been widely used in this literature (Savedoff, 1990; Haisken, De New and Schimdt, 1997; Arbache, 1998). The econometric software SAS, version 6, was used for the estimations. 4. Econometric results 4.1. Regional-nominal wage differentials The results for the regional dummies are presented in table 2. The first set of columns presents the results for Model 1, indicating the gross regional wage differentials; each coefficient indicates a deviation from the average for the ten regions considered in the study 14. It can be seen that only São 14 Halvorsen and Palmquist (1980) have demonstrated that the relative effect of a dummy variable on the independent variable is given by [exp(estimated coefficient) 1] and not by the estimated coefficient itself. The percentage effect is equal to 100*[exp(estimated coefficient) 1] and not by 100*(estimated coefficient). In order to facilitate the interpretation of the results, Tables 1 and 3 present the adjusted version of the coefficients.

Paulo and Brasília (the Federal District) presented wage levels above average in 1992 (19.6% and 16.8% respectively). In the other two years, 1995 and 1997, Curitiba joins the group, with +11.3% and +8.5%. The very poor regions are Recife and Fortaleza, both in the poor Northeastern part of Brazil, with labor income levels at least 34% below average for Recife and at least 40% for Fortaleza. Belém and Salvador, also in the North/Northeast region, appear right above the two mentioned cities, at least 17% below average. Belo Horizonte, Rio de Janeiro (in the rich Southeast region) and Porto Alegre (not so rich South region) are closer to the average. The second column of table 2 displays the results of the complete model, introducing worker and job characteristics. The first thing to notice is that no important changes in positions are present: the very rich and the very poor cities are the same. These results illustrate the high levels of wage inequality in Brazil, even for people in similar conditions (age range, employed, living in metropolitan areas, working in the same sector, etc.). Some minor modifications are preset, though: São Paulo looses the first place in the rank to Brasília in 1997; Belo Horizonte moves up two steps in the ladder in all three years after the controls are included. The Spearman correlation coefficient for each model in different years is always significant and above.8, indicating that the results within each model do not change importantly between the years considered. The same correlation coefficient for the results of the two models is above.9 and significant for all three years, indicating that the introduction of the control variables does not produce important changes in the regional ranking of labor income levels. This stability can also be measured by the dispersion of the dummy coefficients, as indicated by the Weighted and Adjusted Standard Deviation - WASD: its values, before and after introducing the controls, are: 19% and 16.6% for 1992; 25.6% and 21.1% for 1995; 24.9% and 23.2% for 1997. Another way to see this is to compute the distances between the highest and the lowest coefficients with and without controls: these are 61% (regional dummies only) to 49% (all controls included) for 1992; 72% to 57% for 1995, and 74% to 61% for 1997. That is, although labor income inequality is slightly lower after controlling for worker and job characteristics, the reduction is limited and the remaining inequality levels are still very high. Comparing these results to the ones displayed in Savedoff (1990), the similarity is striking, even considering sampling, control variables and econometric method differences. As can be observed in table 3, after controlling for job and worker characteristics, São Paulo is always over 15% above the average income. The three poorest metropolitan areas, Belém, Recife and Fortaleza show always the three lowest incomes.

Although the analysis in this paper does not focus in the other control variables, it is interesting to mention that the results replicate other studies covering specific variables in Brazil (see Table 4 15 ). For the year 1997, labor income is higher for more educated people, with the biggest jump occurring between 9 to 11 years of education (senior high school) and 12 years or more (college). People without any education present a labor income level 46% below average and people with a college degree, 142% above average in 1997 (Reis and Barros, 1991). In the same year, an autonomous worker (business owner without employees) gets 1.4%, and an employer gets 78% above the average. Males get 8% above and females get 14% below the average; white people get 6% above and non-white people get 9% below (Lovel, 2000). Individuals with less than two years of experience receive 12% below the average, while people with 2 or more years of experience receive 11% above the average. As for the age levels, people under 24 years of age get 21% less labor income than the average in 1997; the level goes up until 50 years of age, when it reaches 16% above average. After this age a drop is observed, such that employed people over 55 years present labor income only 12% above the average, replicating the traditional life cycle behavior of income present in studies on the subject (Azzoni et all, 2000) 4.2. Total wage differentials and regional differences How much of the total wage differential can be explained by regional differences? In order to provide an answer to that question, a covariance analysis based on the R 2 was performed 16. Three models were estimated, providing for three different values of R 2 : 1) Model 1 presented before, with regional dummies only (R 2 r ). 2) Model 2 presented before, with all controls included (R 2 c); 3) A model with all controls but regional dummies (R 2 w); As it was said before, Model 1 is equivalent to a simple regional comparison of labor income levels. Its R 2 is the upper bound for the contribution of the regional differences, for it indicates how much of labor income variance can be explained by the regional dummies. The difference R 2 c - R 2 w, that is, between the model with all controls (Model 2) and a model with all controls but regional dummies, is the lower bound for the regional wage differential contribution to the global wage differential. 15 Table 4 displays the estimated coefficients. The analysis in the text however considers the adjusted coefficients [exp(estimated coefficient)-1] 16 Arbache (1998) explains this technique step by step.

This technique was applied to the data and the same procedure was replicated to all control variables, producing the results presented in table 5.The first set of columns indicate the lower and upper bonds for the marginal contribution of each variable to the global wage differential. The marginal contribution of region of residence (metropolitan region) to the global wage inequality in Brazil is restricted to the interval 3.0% to 7.4%, depending on the year. Considering the lower bound, regional dummies are second only to education, the most important variable in explaining wage differentials. Taking the upper bound, regional dummies are ranked after position in the job and with approximately the same importance as experience and race. Another way to analyze the marginal contribution of the control variables is to calculate the variation in the model s standard deviation taking out one of the control variables at a time. The results are also presented in table 5 and indicate the same relative importance for the variables as presented by the R 2 model. Overall, the results involving the marginal contribution of the non-regional control variables are very close to the ones obtained by other studies in Brazil, such as Reis and Barros (1991), on education-income profiles; Cavalieri and Fernandes (1998), on white non-white differentials; Barros et alli (1998), on position in the job and formality of the labor market. 4.3. Cost of living differentials and regional wage inequality The usual wage differential study deals with data deflated by national price indexes. However, part of the regional wage differentials can be determined by differences in cost of living levels among the regions, as well by other compensatory factors, such as crime rates, environment, etc. In this section we will deal with differences in cost of living among the metropolitan regions, using the regional multilateral price indexes calculated by Azzoni, do Carmo and Menezes (2000). They construct multilateral price indexes that allow for the comparison of cost of living levels among all ten metropolitan areas for each year in the period 1981-1999. As can be observed in Table 6, São Paulo is the most expensive metropolitan region in all years, followed by Rio de Janeiro and Brasília. The metropolitan areas located in the poor Northeast and North regions are the less expensive to leave on. Table 6 Cost of living indexes for the ten metropolitan regions

Metropolitan Region Cost of Living Level 1992 1995 1997 Belém 0.948 0.950 0.904 Fortaleza 0.944 0.928 0.907 Recife 0.999 0.953 0.981 Salvador 0.994 0.969 0.953 B.Horizonte 0.951 0.905 0.925 R.de Janeiro 1.042 1.091 1.123 São Paulo 1.093 1.170 1.214 Curitiba 0.964 0.966 0.972 P.Alegre 0.972 0.976 0.986 Brasília 1.110 1.127 1.079 (Average for the 10 metropolitan regions = 1) Using the above cost of living indexes, we have calculated regional-real wages for the metropolitan regions and have substituted real wages for nominal wages in the left hand side of the equations presented in the previous sections. The results for the regional dummies are presented in table 7; the other control variables present similar results in both sets of equations. The first thing to notice is that the dispersion in the coefficients for the dummy variables is reduced, with both extremes approaching the average (the only exceptions are Curitiba and Brasília in specific years). This result indicates that part of the wage differential among metropolitan regions was due to cost of living differences among them. However, even after controlling for this variable, the remaining regional wage inequality is still high. The distances between the highest and the lowest wage levels, after controlling for cost of living, are: from 49% to 42% in 1992; 57% to 44% in 1995. In 1997 São Paulo lost the first place to Brasília, a city where the public sector is very important, specially the federal government. Thus, oscillations in wage levels there are highly influenced by government wage policy, an important political variable in the area. Therefore, the change of places between São Paulo and Brasília in that specific year may be artificial, in economic terms. If we take São Paulo as the point of comparison in 1997, the difference between the extremes drops from 58% to 45% after controlling for cost of living differences. These results are also similar to those presented by Savedoff (1990), although not as much as the nominal differences results. As presented in table 3, nominal differences from that study are relatively stable over time; after controlling for cost of living levels, however, a higher oscillation among years is observed, due probably to the poor cost of living index utilized. 5. Conclusions

This paper deals with one aspect of regional income inequality in Brazil, that is, regional wage differentials. The aim is to verify if wage differentials are significantly reduced when controls for worker and job characteristics and cost of living differentials are introduced. In the three years analyzed, wage differentials were slightly reduced when those controls were introduced, but the remaining inequality was still high. Considering the differing cost of living levels, the results indicate that they do have a role in explaining wage inequality in Brazil, but the remaining differentials are still important. The control variables used in the estimations presented the expected signs, with white, male, formally employed or employer, experienced, etc. workers earning higher wages. They also indicate that the metropolitan areas of São Paulo, Brasília and Curitiba present higher wage levels, even after controlling for worker and job characteristics and cost of living differences. The most important variable to explain wage differentials, as judged by the marginal contribution of the variables, is education, followed by region, experience and race, these three with approximately the same importance. After all the controls were included, regional wage differentials did decrease, and also did their dispersion, reducing the distance between the regions with the highest and the lowest wage levels. However, even after all those controls, wage inequality in Brazil is still high. In the best case, the difference between the two extremes is around 42%. Comparison of these results with other study for the years 1976-1987 revealed that the situation was similar in that period. This indicates that the rigidity in the Brazilian labor markets is strong enough to maintain such impressive wage inequality for more than two decades.

Table 7 - Nominal and real regional wage differentials 1992 Nominal Real São Paulo 0.18887 São Paulo 0.1450 Distrito Federal (Brasília) 0.10001 Distrito Federal (Brasília) 0.0138 Belo Horizonte -0.04277 Belo Horizonte -0.0031 Rio de Janeiro -0.04779 Curitiba -0.0141 Curitiba -0.05313 Porto Alegre -0.0258 Porto Alegre -0.06653 Rio de Janeiro -0.0398 Salvador -0.10707 Salvador -0.0783 Belém -0.23993 Belem -0.1798 Fortaleza -0.29670 Fortaleza -0.2539 Recife -0.29893 Recife -0.2825 1995 Nominal Real São Paulo 0.26150 São Paulo 0.1721 Distrito Federal (Brasília) 0.19630 Distrito Federal (Brasília) 0.1263 Curitiba 0.06932 Curitiba 0.0931 Belo Horizonte -0.07489 Belo Horizonte -0.0129 Porto Alegre -0.08647 Porto Alegre -0.0486 Rio de Janeiro -0.10721 Rio de Janeiro -0.0637 Belém -0.17333 Belem -0.1233 Salvador -0.25957 Salvador -0.2328 Recife -0.30446 Fortaleza -0.2536 Fortaleza -0.30853 Recife -0.2667 1997 Nominal Real Distrito Federal 0.27610 Distrito Federal 0.31109 São Paulo 0.25445 São Paulo 0.16316 Curitiba 0.04321 Curitiba 0.09643 Belo Horizonte -0.05614 Porto Alegre -0.00140 Porto Alegre -0.06695 Belo Horizonte -0.01586 Rio de Janeiro -0.12299 Belem -0.10867 Salvador -0.20872 Rio de Janeiro -0.13174 Belém -0.23982 Fortaleza -0.16509 Fortaleza -0.28192 Salvador -0.17372 Recife -0.33275 Recife -0.29249 Obs: adjusted coefficient = [exp(estimated coefficient) - 1]

Table 4 - Variable coefficients for model 2, nominal values 1992 1995 1997 Coefficient S. Dev. Coefficient S. Dev. Coefficient S. Dev. Constant 8.959 (0.0035) 0.839 (0.003) 1.001 (0.00) Belém -0.274 (0.0247) -0.190 (0.023) -0.274 (0.02) Fortaleza -0.352 (0.0156) -0.369 (0.014) -0.331 (0.01) Recife -0.355 (0.0148) -0.363 (0.014) -0.405 (0.01) Salvador -0.113 (0.0152) -0.301 (0.013) -0.234 (0.01) Belo Horizonte -0.044 (0.0123) -0.078 (0.011) -0.058 (0.01) Rio de Janeiro -0.049 (0.0065) -0.113 (0.006) -0.131 (0.01) São Paulo 0.173 (0.0046) 0.232 (0.004) 0.227 (0.00) Curitiba -0.055 (0.0159) 0.067 (0.014) 0.042 (0.01) Porto Alegre -0.069 (0.0124) -0.090 (0.011) -0.069 (0.01) Distrito Federal 0.095 (0.0181) 0.179 (0.017) 0.244 (0.02) Education - less than 1 year -0.601 (0.0138) -0.686 (0.015) -0.610 (0.01) 1 to 3 years -0.476 (0.0104) -0.527 (0.010) -0.500 (0.01) 4 years -0.343 (0.0089) -0.415 (0.008) -0.397 (0.01) 5 to 7 years -0.256 (0.0078) -0.311 (0.007) -0.302 (0.01) 8 years -0.073 (0.0093) -0.131 (0.008) -0.130 (0.01) 9 to 11 years 0.213 (0.0065) 0.200 (0.006) 0.174 (0.01) 12 years or more 0.832 (0.0000) 0.904 (0.000) 0.885 (0.00) Age - 18 to 24-0.196 (0.0080) -0.240 (0.008) -0.241 (0.01) 25 to 30-0.029 (0.0050) -0.041 (0.005) -0.036 (0.00) 35 to 44 0.111 (0.0059) 0.107 (0.005) 0.095 (0.01) 45 to 50 0.123 (0.0090) 0.160 (0.008) 0.149 (0.01) 55 to 65 0.030 (0.0141) 0.058 (0.000) 0.113 (0.00) Job situation - formal employee 0.053 (0.0035) -0.023 (0.003) -0.011 (0.00) informal employee -0.233 (0.0098) -0.169 (0.009) -0.181 (0.01) public worker -0.090 (0.0169) -0.086 (0.016) -0.050 (0.02) independent worker -0.071 (0.0079) 0.043 (0.007) 0.013 (0.01) employer 0.443 (0.0164) 0.611 (0.014) 0.579 (0.01) Gender - male 0.081 (0.0030) 0.089 (0.003) 0.082 (0.00) female -0.154 (0.0000) -0.160 (0.000) -0.148 (0.00) Sector - manufacturing 0.093 (0.0068) 0.068 (0.006) 0.036 (0.01) construction -0.023 (0.0120) 0.019 (0.011) -0.016 (0.01) commerce -0.067 (0.0083) -0.077 (0.007) -0.104 (0.01) services -0.085 (0.0068) -0.078 (0.006) -0.056 (0.01) transportation and communication 0.107 (0.0135) 0.053 (0.012) 0.131 (0.01) public administration -0.007 (0.0018) 0.114 (0.017) 0.155 (0.02) social and other activities 0.024 (0.0000) 0.036 (0.000) 0.048 (0.00) Ownership - private sector -0.028 (0.0025) -0.011 (0.002) -0.007 (0.00) public sector 0.158 (0.0000) 0.067 (0.000) 0.042 (0.00) Experience - 2 years or less -0.136 (0.0043) -0.133 (0.004) -0.133 (0.00) more than 2 years 0.107 (0.0000) 0.102 (0.000) 0.103 (0.00) Role in the household - non head -0.094 (0.0051) -0.102 (0.005) -0.091 (0.00) head 0.071 (0.0000) 0.078 (0.000) 0.070 (0.00) Race - non white -0.093 (0.0047) -0.096 (0.004) -0.099 (0.00) white 0.060 (0.0000) 0.059 (0.000) 0.061 (0.00) F 976,871 1374,544 1310,435 R2 0.467 0.529 0.514 Number of observations 37.994 41.692 42.205

References ARBACHE, J. S.(1998) How do economic reforms affect the structure of wages? The case of Brazilian manufacturing, 1984-1996. University of Kent at Canterbury, U.K (Studies in Economics, N. 98/17). AZZONI, C. R. (2001) Economic growth and regional income inequality in Brazil, Annals of Regional Science, Vol 35, No. 1, February 2001 AZZONI, C. R.; DO CARMO, H and MENEZES, T. (2000) Índice de Custo de Vida comparativo para as principais regiões metropolitanas brasileiras: 1981-1999, Estudos Econômicos, Vol. 30 N. 1, Jan/Mar pag 1-22. AZZONI, C., MENEZES-FILHO, N., MENEZES, T. and SILVEIRA-NETO, R. (2000) Geography and regional income inequality in Brazil Interamerican Development Bank, Working Paper BARROS, R. P., et alii (1998). Uma avaliação da evolução recente do diferencial de rendimentos formal e informal. Ministério do Trabalho. (Notas sobre Mercado de Trabalho, n.6). BECKER, G. (1964) Human Capital: A Theoretical Analysis, with Special Reference to Education, National Bureau for Economic Research, Columbia University Press, New York and London BECKER, G. and CHISWICK, B. R. (1966) The economics of education and the distribution of earnings American Economic Review 56 (2): 358-369 CAVALIERI, C. and FERNANDES, R. (1998) Diferenciais de salários por gênero e cor: uma comparação entre as regiões metropolitanas brasileiras. Revista de Economia Política, 18(1):158-175. REIS, J. G. A. and BARROS, R. P. (1991) Wage inequality and the distribution of education A study of the evolution of regional differences in inequality in Metropolitan Brazil, Journal of Development Economics, 36:(1) 117-143 July DUMOND, J. M., HIRSCH, B. T. and MACPHERSON, D. A. (1999) Wage differentials across labor markets and workers: does cost of living matter?, Economic Inquiry 37:(4) 577-598 October FERREIRA, A. (2000) Convergence in Brazil: recent trends and long-run prospects, Applied Economics 3:(4) 479-489, March 2000 GREENE, W. H. and SEAKS, T. G. (1991) The restricted least squares estimator: a pedagogical note. The Review of Economic and Statistics, 73(3): 563-567. HADDAD, E. A. (1999) Regional inequality and structural changes lessons from the Brazilian experience, Ashgate, England. HAISKEN-DENEW, J. P. and SCHMIDT, C. M.(1997) Inter-industry and inter-region differentials mechanics and interpretation. The Review of Economic and Statistics, 79(2): 516-521. HALVORSEN, R. and PALMQUIST, R. (1980) The interpretation of dummy variables in semilogarithmic equations, American Economic Review, 40: 474-475.

KENNEDY, P. (1986) Interpreting dummy variables. Review of Economic and Statistics, 68:174-5. LOVELL, P. A. (2000) Race, gender and regional labor market inequalities in Brazil, Review of Social Economy, 58:(3) 277-293, September MAIER, G. and WEISS, P. (1986) The importance of regional factors in the determination of earnings: the case of Austria. International Regional Science Review, 10(3):211-220. MINCER, J. (1958) Investment in human capital and personal income distribution The Journal of Political Economy LXVI (4): 281-302, August (1974) Schooling, experience, and earnings, New York, National Bureau of Economic Research NORDHAUS, W. and TOBIN, J. (1972) Is growth obsolete? Economic Research: Retrospect and Prospect, Vol. 5, Columbia University Press RAZIN, A. and YUEN, C. W. (1997) Income convergence within an economic union: the role of factor mobility and coordination Journal of Public Economics, 66:(2) 225-245 November REIS, J. G. A. and BARROS, R. P. (1991) Wage inequality and the distribution of education: a study of the evolution of regional difference in inequality in metropolitan Brazil. Journal of Development Economics 36: 117-143. ROBACK, J. (1988). Wages, rents, and amenities: differences among workers and regions. Economic Inquiry, 26(1):23-41. (1982) ). Wages, rents, and the quality of life. Journal of Political Economy, 90 (6):1257-78 ROBERTSON, R. (2000) Wage shocks and North American labor-market integration American Economic Review, Vol. 90 No. 4, p 742-764, September ROSEN, S. (1986). The theory of equalizing differences. in: ASHENFELTER, O. and Layard, R. Handbook of Labour Economics. SAVEDOFF, W. D. (1990). Os diferenciais regionais de salários no Brasil: segmentação versus dinamismo da demanda. Pesquisa e Planejamento Econômico, 20(3): 521-555. SERVO, L. M. (1999) Diferenciais regionais de salários no Brasil, Master Thesis presented to the Department of Economics, Universidade de São Paulo SUITS, D. (1984). Dummy variables: mechanics vs interpretation. Review of Economics and Statistics, 66:177-180. TOPEL, R. H. Regional labor markets and the determinants of wage inequality, American Economic Review, 84:(2) 17-22 May TREJO, S. J. (1997) Why do Mexican Americans eran low wages? Journal of Political Economy, 105:(6) 1235-1268 December WEILER, S. (2000) Industrial structure and unemployment in regional labor markets, Industrial Relations, 39:(2) 336-359 April