Inter-American Development Bank Banco Interamericano de Desarrollo Latin American Research Network Red de Centros de Investigacin Research network

Working paper #R-395

Geography and Income Convergence among Brazilian States

By

Carlos R. Azzoni* Naercio Menezes-Filho* Tatiane A. de Menezes* Raul Silveira-Neto*

*University of So Paulo

May 2000

Cataloging-in-Publication data provided by the Inter-American Development Bank Felipe Herrera Library

Geography and income convergence among Brazilian states / by Carlos R. Azzoni [et al.]. p. cm. (Research Network Working papers ; R-395) Includes bibliographical references. 1. Income--Brazil--Mathematical models. 2. Income--Brazil--Effect of Human services on-Mathematical models. 3. Income--Brazil--Effect of Human geography on--Mathematical models. 4. Income distribution--Brazil--Mathematical models. I. Azzoni, Carlos Roberto. II. Inter-American Development Bank. Research Dept. III. Series. 339.3 G385--dc21

82000 Inter-American Development Bank 1300 New York Avenue, N.W. Washington, D.C. 20577 The views and interpretations in this document are those of the authors and should not be attributed to the Inter-American Development Bank, or to any individual acting on its behalf. The Research Department (RES) publishes the Latin American Economic Policies Newsletter, as well as working papers and books, on diverse economic issues. To obtain a complete list of RES publications and read or download them, please visit our web site at: www.iadb.org/res/32.htm.

Abstract The objective of the study is to identify the role of geographical variables in explaining differences in per capita income among Brazilian states. It also aims at ascertaining the degree to which such variables affect convergence or divergence trends in per capita income among these states. In order to investigate these issues it uses micro-data, instead of the more traditional aggregate data, averaged up from household to birth cohort level. Both the level and the change in average household income per capita across Brazilian states are correlated to geographical and household variables. The aim is to capture not only the influence of household human capital and wealth variables on the convergence of per capita income (along the lines of the neoclassical model), but also that of spatial or geographical characteristics, such as public infrastructure, health and education services. Therefore, this paper simultaneously considers data on geographical variables and repeated cross-sections of household surveys. The use of cohort level data means that we can construct cohort/state/year means for all variables of interest and control for state, life cycle and composition effects for the first time in this literature. The results indicate that the geographical variables seem to be important determinants of income levels and growth. Altogether, the results indicate that human capital and infrastructure variables are important areas for government intervention, as these are some of the main factors behind the differences in steady-state rate of income growth in Brazil.

Acknowledgments This paper is part of the Inter-American Development Bank research program on the effects of geography on development. Support from NEMESIS - Ncleo de Estudos e Modelos Espaciais Sistmicos, FUJB No. 7077-7, FINEP/PRONEX No. 41.96.0405.00 is also acknowledged. We are very grateful to audience members of the IDB seminar at Cuernavaca, to Fabiana del Felicio at FEA-RP for excellent research assistance, to Denisard Alves for the climate data and to seminar participants at UFMG, IPEA/RJ, FVG-RJ, and at the European Meeting of the Regional Science Association International.

CONTENTS 1. Introduction 2. Regional Inequalities in Brazil 3. Theory 4. Econometric Methodology 4.1. The Construction of Cohorts 5. Data and Data Sources 6. Econometric Results 7. Conclusions and Policy Implications References TABLES Table 1 Description of the main variables Table 2 Correlation coefficients among variables Table 3 Regression results for income levelsDependent Variable ln(y) Table 4 Regression results for income growthDependent Variable ln(y) Table 5 Convergence results Table 6 Decomposition of income differentials among regions Table 7 Interstate Gini coefficients Table 8 Variables averages and standard deviations by state FIGURES Figure 1 Per capita income for Brazilian macro regions Figure 2 Per capita income and age of household head across cohorts Figure 3 Education and household age across cohorts Figure 4 Education across Brazilian macro regions Figure 5 Geographical variables for Brazilian states Figure 6 Urban/rural situation and density across Brazilian states Figure 7 Infrastructure variables for Brazilian states Figure 8 Human development variables for Brazilian states 10 11 12 12 13 13 14 14 23 24 25 26 27 27 28 29 5 5 6 8 9 10 15 19 21

1. Introduction In mainstream economics, the starting point for understanding the existence of poor regions is the neoclassical exogenous growth model developed by Solow and Swan in 1956. According to this model, per capita income differentials among regions are determined by their respective initial endowments of resources, so that it is not so much that there are poor regions, as given areas that have a greater concentration of poor families. Such models are presented in greater detail in Barro and Sala-i-Martin (1995). Recent studies, such as Hall and Jones (1996), Chang (1994), Ravallion and Jalan (1996) and Jalan and Ravaillon (1998b) have highlighted the importance of geographical, institutional and political variables in determining regional income differentials. According to these authors, in certain circumstances the presence of poor families is determined endogenously, and not by the more widespread exogenous variables assumed in growth models. Differing levels of geographical capital, such as climate, local infrastructure, access to public services and knowledge about the local physical reality and adequate technologies, would influence the use of private capital. That is, geographical variables would affect the marginal return of private capital. The imperfect mobility of factors, usually assumed in this sort of model, would create the conditions for the persistency of inequalities. The coexistence of increasing returns to geographical capital with non-increasing returns to private capital is conceivable in this line of reasoning. Poor people tend to live in poorly supplied regions. With the same personal characteristics, they would be better off if living in a richer region. This difference in diagnosis is reflected in the important issue of policy recommendations. According to the first class of models, regional inequality is to be solved by allowing free mobility of factors that should in turn result in long-term convergence of growth rates. The second strand of the literature, on the other hand, can be used to justify regional policies aiming at reducing regional inequality, such as public investments in geographical capital. The objective of this paper is to offer evidence on this controversy in the case of state income inequalities in Brazil over the period 1981-1996. In Section 2 some general information on regional inequalities in Brazil is provided as well as a review of the some existing empirical studies, creating the background for the study to be performed in the paper. Section 3 presents a methodological discussion of the model to be estimated. Section 4 presents the data used in the research and discusses the advantages and limitations of using micro-data. Section 5 presents and discusses the econometric results, which are commented on in the concluding section. 2. Regional Inequalities in Brazil Brazil is well known for its high levels of regional income inequalities. In 1960, Brazil had a GDP per capita of US$1,449. Thirty-five years later, in 1995, this figure had risen to US$3,556, corresponding to an average growth rate of 2.6% per year. Data on per capita GDP for Brazilian states indicate that only three Brazilian states had figures above the national average, namely, So Paulo, Rio de Janeiro and Rio Grande do Sul. The state of So Paulo was notable for a GDP per capita that was almost 2 times the national average. The poorest state was Piau, with a GDP per capita that was 4.5 times less than the Brazilian average, and 8.9 times less than that of the state of So Paulo. It is notable that nine of the ten poorest states in Brazil were in the Northeast, and three of the four states of the Southeast were among the five richest states in Brazil.

In 1995 a larger number of states were above the national average. Of these, the first two, So Paulo and Rio de Janeiro, are in the Southeast region, while the next three, Rio Grande do Sul, Paran and Santa Catarina, belong to the South. The other Southeastern states achieved visible improvements with regard to the national average (e.g., the state of Mato Grosso). Of the ten poorest states, eight are in the Northeast, and two in the North, Amazonas and Par, were notable for having declined in relative terms since 1960 to a standing among the ten states with the lowest GDP per capita. So Paulo was still the richest state in 1995, with a GDP per capita that was 1.7 times the national average, with an average growth rate of 2.1% per year between 1960 and 1995, 0.5% below the national growth average. Piau, on the other hand, was still the poorest state in Brazil. It is nevertheless interesting to note that while, in absolute terms, the state s position was considerably worse than the rest of the country, in relative terms it had improved, with a GDP per capita that was 3.7 times smaller than the national average, and 6.1 times smaller than that of the richest state. Another interesting point is that in the state of Piau GDP per capita grew 3.1% per year, one percentage point above that of the state of So Paulo for the same period. These observations raise the question of whether there is an income convergence trend among Brazilian states. Taking the neoclassical growth model as a basis, Ferreira and Diniz (1995), Schwartsman (1996) and Zini (1998), analyzing the period initiated in 1970, could not reject the hypothesis of absolute convergence. They estimated speeds of convergence among Brazilian states that were above the levels predicted by the model. Azzoni (1999), working with a longer series (1939-1996), also found indications of absolute convergence of income, but at a much lower speed. It remains to be investigated whether there are spatial externalities that condition regional growth. That is to say, geographical characteristics such as climate, public and private infrastructure (which could be understood as reflecting geographical capital ) may be affecting the growth rates of states or regions by influencing the productivity of individual or family capital. This would have decisive implications for public sector policies designed to combat poverty and to improve incomes at regional and state levels, since differences in living standards would result not only from the initial conditions of families, as assumed by neoclassical growth models, but also from differences in the geographical capital between regions. Since the previous models used for calculating convergence did not consider the effect of geographical variables, and these would tend to be positively correlated with initial income, their results could, in fact, be reflecting conditional instead of absolute convergence, and they could be underestimating the true velocity to this convergence. We thus propose to estimate the variation in household income per capita for Brazilian states as a function of geographical, state and household variables, in order to capture not only the influence of the individual characteristics of households on the convergence or divergence of per capita income (along the lines of the neoclassical model), but also that of spatial or geographical characteristics.

3. Theory This section aims at presenting the neoclassical growth model with exogenous savings, geographical variables and fixed effects, as in Islam (1992), to be taken to the data. Consider the production function, with labor augmenting technical progress:

Y (t ) = K (t )( A(t ) L(t ) ) G
1

where Y = product, K = capital, L = labor and G = geographical capital (fixed in the economy). The public input is thus complementary to private inputs, and we have:
L(t )= L(0 ) e nt A(t )= A(0 ) e gt

where n and g are the (exogenously determined) population and technology rates of growth. Capital accumulation per effective worker in the steady state will be given by:
dk ) = s y (n + g + )k dt

where s = savings rate, = depreciation rate and:

= y

Y (t ) = K (t ) ,k A(t ) L(t ) A(t ) L(t )

This equation implies steady state levels of capital and product (per effective worker), which are described by: sG 1 k* = ( ) n+ g+ y *= ( sG 1 ) 1 G n+ g+

(t )G , we can approximate its time (t ) = k With the product per effective worker given by y variation around the steady state to get (in logs):

(t )) d ln( y (t ))] = [ln( y *) ln( y dt

where = (n+g+ )(1-). The growth of per-capita product during the period = t2 - t1 around the steady-state will be: (t )) = ln( y (t )) ln( A(0)) gt where we have used the fact that ln( y
ln y (t 2 ) ln y (t1) = 1 e

1 e ln G + 1 e ln y (t1)+ g t 2 e t1 + (1 e ) ln A(0) 1

)1 ln (s ) (1 e )1 ln(n +

g+ )+

One can include human capital in the production function:

Y (t ) = K (t ) [ A(t )L(t )]
1

H (t ) G

and around the steady state we will then have:

ln y (t 2 ) ln y (t1) = 1 e t ln y (t1)+ 1 e

)1

ln s 1 e

+ )1 ln (n +

g+ )+

(1 e )1

ln h * + 1 e

( ) (

)1

G + 1 e ln A(0) + g t 2 e

where h* = human capital investment as a fraction of income in the steady state. 4. Econometric Methodology This section aims at representing the last equation in a form that can be estimated with the data at hand. As the main aim of this research is to investigate the roles of geographical and human capital variables on growth, we propose: y it = y it 1 + S i + 1 H it + 2 Git + t + it Where:
y it 1 = ln y (t1) y it = ln y (t 2 ) ln y (t1) = 1 e

)
1 )1 [ ln s ( + ) ln( n + ) + ln A(0) / g+ 1 ] 1

S i = 1 e

t = g t 2 e t1

where G and H are human capital and geographical variables.

4.1 The Construction of Cohorts The data we use come from repeated cross-sections of very rich household surveys, carried out by the Brazilian Census Bureau (see below). The use of micro-data to examine issues of convergence has not, as far as we know, been done in the literature before. It is well known in the consumption and labor supply literature (see Browning et al., 1985 and Attanasio and Browning, 1994) that with repeated cross-sections it is possible to construct demographic cohorts based on date of birth, and calculate cohort-year means for all variables of interest, including income, education, labor force participation and living conditions. We propose to extend this methodology to include the state of residence as another grouping variable and derive state-cohort-year means for the variables of interest. The value attributed to each cohort in state s in year t is

n cst where: n cst is the proportion of household heads born in an interval of determined years (e.g., 1940 to 1945), living in state s in period t. Ten cohorts were constructed for each state in each year. The same procedure was applied for all other variables included in the analysis, so that we have, for example, the average number of years of education for people included in each cohort; the same holds for all other variables.
The advantages of such a procedure are many-fold. Firstly and most importantly, the use of micro-data allows us to control for changes in the composition of the population in each state that cannot be controlled for with the use of aggregate data. Secondly, we can control for life-cycle and generation effects, which means that we are, in effect, considering the effect of geographical variables on income growth and convergence within a generation or for a population with the same age. Thirdly, one can identify state fixed effects without having to rely only on the time component of the series, since we have obtained various (10 in our case) observations for a given state in a given year. Finally, one can rely on differences across generations within a state-year group to identify the effects of human capital on growth, for example, that are not readily identified using aggregate data (Islam, 1994). The main disadvantage of using cohort level data is that if there are measurement errors at the household level they are likely to be carried out to the cohort means unless the cell sizes are large. In effect, we will be estimating an equation of the form: Where the subscript c means that the variables are now cohort-state-year means, X = household y ict = y ict 1 + S i + 1 H ict + 2 Gict + 3 X cit + 4 LFcit + t + it controls and LF = life cycle variables.

y cst =

ln y
i

ncst

5. Data and Data Sources The implementation of the model described above requires panel data, not easily available in general. In the case of Brazil we have repeated cross-sections of a yearly household survey (PNAD Pesquisa Nacional por Amostra de Domiclios), conducted by IBGE, the Brazilian Census Bureau, that can be used as a pseudo-panel, by constructing a model that looks like an individual-level model but is for cohorts (see Ravallion, 1998). Due to data limitations, only 19 out of 27 states were considered in the study; the excluded states belong to the almost uninhabited northern region (including the Amazon region). The total number of households considered in different years ranged from 49,514 to 90,776, with an average of 58,328.1 The average number of households per cohort is 269. Due to the small number of observations in some cohorts, only 2,166 were considered in the analysis. The main variables in the analysis are as shown in Table 1. The dependent variable used throughout (y) is per capita monthly labor income from the main job. We have also made some experiments with two other variables, hourly labor income and total income, with no apparent change in the results.2 In Figure 1 the evolution of per capita income over the period 1981-1996 for four macro-regions of Brazil is displayed. The first observation relates to the level of income: the average for all states in Brazil over the 16 year period is of US$45 per month, or US$540 per year, indicating the low average level of income in the country. Secondly, a slight declining trend is observed for all states and regions, showing that the average income fell in Brazil during the period considered.

South
Log of p. c. income (R$/month) 4.5 4.3 4.0 3.8 3.5 3.3 3.0 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 Northeast

Center-West

Figure 1 - Per capita income for Brazilian macro regions

Table 8 presents averages and standard deviations for all variables for each state Instead of the 3,040 possible cases (10 cohorts x 19 states x 16 years). The small cohorts are located among the youngest in the first years of the period and among the oldest in the more recent years. Only cohorts with at least 30 households were included in the sample.
1 2

10

It is interesting to point out the peak for all regions in the year 1986, related to the Cruzado macroeconomic stabilization plan, which included heavy price controls and brought inflation down. The plan was preceded by a sharp depression in 1983, and even before 1986 some growth in income could be observed. Income levels fell after 1986 and began to recover only in 1994, with the implementation of the Real macroeconomic stabilization plan. It was applied in July of 1994 and succeeded in reducing inflation gradually during the following month, producing the peak observed in 1995. Finally worth mentioning is the difference between the Northeast region and the other three regions in terms of income levels, with that region presenting lower levels of income and the other three presenting similar higher levels. The average annual figures for the South, Southeast and Center-West are close to US$800, and for the Northeast they are around US$360. Over time, there is no indication that this difference is falling. In Figure 2 the per capita income per household by cohort (or age) is presented. Each cohort is represented by a line, the set of lines indicating the general trend. The traditional life cycle behavior of income is present, although not as clear as if individual data were used. Income grows as individuals grow older but, up to a certain age, so does the number of children, preventing growth in average income. For people over 40, the average income falls as expected. For each cohort (each line) the traditional life cycle behavior of income is also present.

5.0 Log of income (R$/month) 4.5 4.0 3.5 3.0 Each line represents a Cohort 2.5 2.0 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62 64 66 Household head age 68

Figure 2 - Per capita income and age of household head across cohorts

The same cohort disposition is utilized in Figure 3 for the household head s education. In general, the number of years of education is very low, with an overall average of 3.8 years for the country as a whole. The positive slopes of the cohort curves indicate that the number of years of study grows in each cohort as the household head grows older, but only up to a certain age range (mid-40s); after that the curves are flat, indicating that household heads do not invest in education any more. The overall declining disposition of the curves show that the younger generations present higher levels of education, as compared to older generations, as a result of the improvements made in the Brazilian education system over time.

11

6.5 Years of education 5.5 4.5 3.5 2.5 1.5 22 25 28 31 34 37 40 43 46 49 52 55 58 61 64 67

South Center West Northeast
1995

Figure 3 - Education and household age across cohorts

This can be seen also in Figure 4, in which the total number of years of education for each macro region is displayed. An increasing trend is observed for every region, but differences in levels are important: the differences between the South and the Southeast are small, with the Center-West falling behind and the Northeast with a much lower level. The average number of years of education for the regions are: Southeast, 5.4; South, 4.8; Center-West, 4.0, and Northeast, 2.9

6.0

Southeast
5.5

Years of education

5.0 4.5 4.0 3.5 3.0 2.5 2.0 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994

1996

Figure 4 - Education across Brazilian macro regions

The pure geographical variables are rainfall, (the negative of) latitude3 and temperature in June (the month with the highest variation across states in Brazil). These are displayed in Figure 5, in which the states are positioned in the horizontal axis in decreasing order of per capita income levels. It is clear from the figure that there is a positive correlation between income level and the variables rainfall and latitude and a negative correlation between income and temperature.

Throughout the paper the negative of latitude will be utilized in order to facilitate the interpretation of the coefficients. Thus, Southern (richer) states will show a higher value and Northern (poorer) states a lower value for this variable.

Rainfall (cm3/year)
5 4

Log

3 2 1
SP MS RJ RS MT SC PR GO MG ES SE BA PE

Temperature (Degrees C)

Latitude (negative of degrees)

AL RN CE PB MA PI

Figure 5 - Geographical variables for Brazilian states

Two control variables are displayed in Figure 6: urban(1)/rural(0) situation and density (persons per room in the household). The former is positively correlated with income, meaning that poor states have a higher percentage of people living in rural areas. Density, on the other hand, is negatively related to income, for poorer states tend to have more persons per room in the household.

1.4 1.2 1 0.8 0.6 0.4 0.2 0 SP RJ MT PR MG SE PE RN PB PI

urban density

Figure 6 Urban/Rural situation and density

The infrastructure variables include access to public sewage, electricity and garbage systems. The average levels for states are shown in Figure 7, in which states are also displayed in decreasing order of income levels (household connected to the service = 1; not connected = 0). As can be seen, there is a positive correlation between the percentage of households connected and income for all three variables.

13

Electricity
Connection to public service = 1 no connection = 0

1 0.8 0.6 0.4 0.2 0 SP MS RJ RS MT SC PR GO MG ES SE BA PE AL RN CE PB MA PI Garbage Sewage

Figure 7 - Infrastructure variables for Brazilian states

Finally, Figure 8 displays two proxies for the efficiency of public services: infant mortality and life expectancy. It is clear that life expectancy increases with the state s income level and that infant mortality decreases with income. It is also clear that there is a big difference between the Northeast region states and the others, for the former present much higher levels of infant mortality.

120

70

100

Infant mortality

65 80

60

Life expectancy Infant mortality

60

40 55 20

0
SP MS RJ RS MT SC PR GO MG ES SE BA PE AL RN CE PB MA PI

50

Figure 8 - Human development variables across Brazilian states

Other human capital variables are education of the household head s partner and a measure of children s delay in education (number of years of schooling effectively attended by the child divided by the expected number of years of attendance, given the child s age). The life cycle variables include age and participation of household head, partner and children in the labor force. Other controls variables include the gender of household head and measures of household wealth, such as whether the household possesses a stove and refrigerator ( fridge in tables). Time and cohort dummies were also included. The simple correlations among the main variables used in the analysis are presented in Table 2.

Years

6. Econometric Results We have estimated the main equation of interest in two different forms for the dependent variable: the logarithm of per capita income and the first differences of the logarithm of per capita income for each state.4 Most right-hand side variables vary across cohorts, years and states; the exceptions are infant mortality and life expectancy, which vary across years and states but are invariant across cohorts, and geographical variables, which vary only across states. The standard errors have been corrected by the fact that data may come from a population with a grouped structure, according to geographic location (state). The presence of intra-group error correlation and the fact that the regressors include variables with repeated values within states (e.g., climate variables) mean that the OLS estimated standard errors may be invalid for statistical inference. Therefore, we adopted the approach described by Moulton (1986) and adjusted the OLS standard errors accordingly; the standard errors are also heteroskedasticity robust. Dummy variables for each year were included in all regressions. Table 3 presents the results of the estimations using the levels of income as the dependent variable. Column (1) presents results for the regression including only time dummies and geographical variables. Only (the negative of) latitude is statistically significant, its sign indicating that the farther South, the higher the state s income. In Column (2) human capital controls are included. As expected, the coefficients of the household head s education and labor force participation enter positively while, conditional on this result, the coefficient of the partner s participation in the labor force enters negatively. Education of the partner is not significant but whether or not the children participate in the labor force is significant and the coefficient is positive. It is interesting to point out that in this column latitude remains positive and significant, but temperature also becomes significantly positive. These alternating signs for the two variables pose a problem, for states farther South (higher values for latitude) are also cooler (smaller values for temperature). We ran some experiments with the variable distance from the state s capital city to the sea, to control for the fact that coastal cities present higher temperature, regardless of latitude, and that most capital cities in the Northeast are sea cities. On the other hand, the rich and fast growing states in the Center-West present high temperatures, are in the middle range of latitude and very distant from the sea. An interaction dummy for temperature and distance to the sea was tested, which appeared not significantly, but caused the coefficient on temperature to become non-significant too. From these experiments, it is clear that latitude is the most stable of the geographical variables, with the others being more sensitive to the inclusion of additional variables, as will become clear below. With respect to the wealth variables included in Column (3), only refrigerator possession makes a significant and positive difference on the level of household income. All the other variables present similar results as in Column (2). Variables indicating the rural/urban situation and infrastructure capital for households are included in Column (4). In this case, refrigerator possession is no longer significant,5 but density becomes significant, keeping its negative sign as before. This indicates that states with a higher coverage of public garbage collection and with fewer persons per room tend to present higher levels of income. Gender of the household head becomes significant, indicating that household headed by males tend to present higher per capita income (the sign is the
4 5

Given that the income levels are expressed in logarithms, the first differences indicate the rates of growth. Multicolinearity among the regressors turns the t test inefficient but the coefficients are non-biased.

15

same as in the previous columns, but the coefficients were not statistically significant in those). As for the geographical variables, rainfall appears as significant, the sign indicating higher incomes for states with higher rainfall levels; the other two geographical variables remain significant and with the same signs as before. The proxies for the efficiency of public services (children s education, life expectancy and infant mortality) included in Column (5) are not significant; refrigerator possession reappears as significant. Column (6) drops some of the variables to avoid multicolinearity and some minor changes in the results occur. The household head s education becomes insignificant, as does temperature. The life cycle variables now appear as significant and with the expected sign; metropolitan status is significant and positive. Only a small drop in the R2 is observed from Column (5) to Column (6). Finally, Column (7) drops the geographical variables and includes the state dummies,6 with some changes in the coefficients resulting: education of the household head is again significant; the life cycle variables remain significant and with the right signs. The biggest change is the inclusion of infant mortality as significant and with a negative sign, indicating that infant mortality is negatively associated with the state s income level. Therefore, the substitution of dummy variables for geographical variables does not provoke large changes in the results and in some cases they become more reasonable, as in the signs for the life cycle variables. Of the results as a whole, latitude seems to be the most stable of the geographical variables, appearing in all regressions with the expected sign and significantly in five out of six specifications; the other two appear only in some specifications and temperature always with contradicting sign. Participation of the household members in the labor force is the most robust of the human capital variables, with positive signs for the head and the children and negative signs for the partner. This indicates that partners tend to participate more in the labor force when the household per capita income is lower; as income grows, partners tend to participate less. Education of the household head was not significant in only one case. As far as the other variables, the number of persons per room (density) appears always with a negative sign and is significant in four out of five specifications. Household capital, indicated by refrigerator possession, appear as significant and positive only in two out of five specifications, although always with a positive sign. The existence of garbage collection service is also important, for this variable was significant in all cases in which it was included. Table 4 reports the results of the first-differences specifications, with the inclusion of the initial level of income as a regressor. In the first column, one can see that there is no significant unconditional relationship between changes in income and the pure geographical variables; the same holds for the initial income level. The results become a bit more interesting when controls for human capital and labor force participation are included (Column 2). It can be seen that the initial level of income presents a significant negative sign, indicating the states with lower initial income levels tend to grow faster than states with higher income levels. Latitude and temperature appear as significant, with the same conflict of signs as in the results of Table 3.7 The signs for education and labor force participation are the same as in the regressions with income levels presented before: the more educated the household head and the children, the higher the labor force participation and the lower the partner s participation in the labor force, the higher the income growth.
6

Notice that the geographical variables as well as the state dummy variables are time invariant. As such, they are substitutes for each other and cannot enter simultaneously in the regression. 7 The same experiments for the equations in levels of income presented similar results in this case, indicating that latitude is the most stable of the geographical variables considered in the study.

16

The inclusion of household capital and density information (Column 3) causes latitude to be non-significant but does not change the sign and significance of the initial level of income and of temperature. Only refrigerator possession is (positively) associated with income growth. All the education and labor force participation variables keep their status as in the previous column. The inclusion of the metropolitan/urban/rural situation and public infrastructure made in Column (4) makes latitude and density significant, as compared to the previous column. Of the newly included variables, only connection to public garbage collection is (positively) associated with income growth and density becomes significant. In Column (5) all variables are included, the last being children education and human development variables, which turned out to be insignificant. No important changes with respect to the results of Column (4) are observed. Column (6) drops some of the variables to deal with multicolinearity among the regressors. Comparing Columns (5) and (6), there is a substitution of rainfall for latitude. Education of the household head and refrigerator possession both become insignificant; there is also a small decrease in R2. Equation (7) substitutes state dummies for the geographical variables, with a small gain in R2 and some minor changes in results: now life cycle variables and education of the household head become significant, while refrigerator possession becomes insignificant. As in the regression in income levels, infant mortality appears as negative and significant, indicating that states with higher infant mortality tend to grow less than states with lower infant mortality levels. Finally, in Column (8) income growth was measured in third-differences and all right handside variables at their levels in the initial year of the three-year periods to address causality concerns. As can be seen, the initial income remains as significant as in the other specifications. Rainfall and latitude appear as significant, with the same sign contradiction; life cycle indicators are significant and with the expected signs; education of the household head, however, appears as insignificant, and with a wrong sign. Labor force participation, density, garbage and infant mortality present the same results as before. Thus, even taking into account possible causality problems, the main results seem to hold. Considering these results on income growth, it is clear that the initial level of income is an important variable, being present in all specifications. The same holds for the participation of the household head and children (positive sign) and partner (negative sign) in the labor force. Density is always negative, but significant in only three out of six specifications. Education of the household head is significant in six out of eight specifications; household capital (refrigerator possession) is positive and significant in four out of six specifications. Public infrastructure, represented by garbage collection, was positive and significant in all cases it was included in the regressions. Among the geographical variables, temperature is significant in five out of seven specifications and latitude in four, with the contradictory signs commented on before. As mentioned before, in all specifications but the first, the coefficient for the lagged income is negative and significant, indicating that states with smaller levels of initial income tend to have faster income growth. In Table 5 the coefficients on the lagged dependent variable for the different specifications of Table 4 are presented again, together with the implied and the estimated velocity to half convergence (column numbers are the same as in Table 4). In the first column there are no other variables apart from the lagged dependent variable. The estimated coefficient is very small and not significantly different from zero, which basically implies persistence in income differences among states. That is, there is no evidence of absolute convergence among states in Brazil in the

17

period analyzed. Column (1) includes the time dummies and pure geographical variables, with no important changes in the results. Column (2) adds age, age squared and human capital variables, with important changes in the results, for now the coefficient on the initial year income is negative and significant. The implied and the resulting number of years to achieve half convergence are also shown in the table. Column (3) adds household capital variables, increasing marginally the speed of convergence. In Column (4) the public infrastructure variables are added, and the speed of convergence increases reasonably, implying 0.8 years for half convergence. The introduction of human development variables additionally increases the speed of convergence, as shown in Column (5). Reducing the number of variables to decrease multicolinearity returns the speed of convergence back to the same level of Column (3). The substitution of dummy variables for geographical variables, as done in Column (7), causes the speed of convergence to increase and the number of years to half convergence to be 1.04. All in all, the results indicate a very high speed of convergence of income among Brazilian states, after provision is made for differences in geography, human capital, labor force participation, infra-structure and human development conditions. It is interesting to point out that in this case we are dealing with conditional convergence; that is, each state is converging to its own steady state situation. Thus, instead of indicating a positive situation, in which inequalities among states in Brazil would be diminishing at a high speed, the results show that the states are close to their steady state income levels, meaning that the resulting equilibrium inequality situation is close to the high inequality level presently observed in the country. This result is perfectly compatible with the abovementioned absence of absolute (or unconditional) convergence of income in Brazil. The different state provisions of education, infrastructure, etc. prevent the equalization of income levels across states. Table 6 shows the results of the decomposition of the observed differences in per capita income across regions. It is interesting to highlight the fact that the Northeast/Southeast observed differences in per capita income are pretty much the same as the Northeast/Center-West differences; the South/Northeast differences are slightly higher but in the same value range as the other two. This is expected, considering the information on regional incomes presented in Figure 1. The most important variables in explaining these differences are the geographical variables rainfall and latitude, infrastructure capital, represented by the existence of a public system for garbage collection, and infant mortality. Other variables present a much smaller importance in the decomposition. Temperature appears as important in explaining differences between the South and the Northeast only. Table 7 shows the Gini inequality index across states in Brazil (weighted by the population of each state). The number in the first column (Average Income) is representative of inequality using the observed income data. The inclusion of geographical variables causes a drop of 15% in the Gini coefficient. The addition of human capital variables reduces the coefficient by another 6.4% (as compared to its already lower value). The further inclusion of public infrastructure and household capital does not produce any noticeable change in the Gini coefficient.

18

7. Conclusions and Policy Implications The main objective of this paper is to shed some light on the effects of geographical variables on the per capita income patterns of growth of Brazilian states. To achieve this aim, we proposed, for the first time in this strand of the literature, a methodology to examine the issue of convergence using micro-data. We constructed cohort/state/year averages of all variables of interest, and regressed income levels and income growth on a variety of human capital, life cycle and geographical variables. The main results indicate that geographical variables are important in explaining the differences in income levels and in growth across Brazilian states. This is shown by their joint impact on the level and growth regressions. The results also show that there is no sign of absolute convergence in Brazil in the period analyzed. On the other hand, conditional convergence will take place rapidly in Brazil, once human capital, infrastructure and geographical variables are controlled for. The estimated high speeds of conditional convergence indicate that Brazilian states are close to their steady state levels of income, indicating that the equilibrium regional income inequality in the country is not much different from the present one. Considering the prospects of government intervention, it is clear that investments in public infra-structure do play a role, especially in reducing the differences between the Southeast and the Northeast (after controlling for differences in the provision of garbage collection services, income inequality among those regions drops to 36%, from the previous 82%). Other important factors in the regressions are household density (number of persons per room) and household capital (refrigerator possession). Human capital variables appear as significant in general, especially participation in the labor force. However, the decomposition of income differentials indicates a minor quantitative role for this variable. However, no control is made for variations in the quality of education provided among states, and recent indications are that poorer states also present poorer education quality. Unfortunately, there is no information on this variable for the period considered. Thus, investments in this area, although with a positive impact, do not seem to have a major influence on the levels of income inequality in the country. Investments in the quality of education might have a better effect, although no evidence on that aspect can be derived from this study. As for geographical variables, they enter significantly into most regressions, meaning that, even after controlling for differences in education, labor force participation, social and private capital, etc., there still are differences among states that are related to geographical aspects. The decomposition of income differentials indicates that these variables are quantitatively very important. The variable rainfall is nature-related and probably includes its effects on agricultural productivity. Latitude, on the other hand, could be understood as encompassing institutional, cultural, religious and other sorts of regional characteristics that influence the productivity of factors in the regional production functions. The same conclusion can be drawn from the results of the specifications that substituted state dummies for geographical variables, with the coefficients of the dummies being significant, leading to the same interpretation on the geographical aspects considered. Another indication in this direction is the relatively low R2 in the growth equations, indicating that other factors but the ones included in the regressions may account for the variance in the state s rates or growth. This argument is not valid for the level regressions, since the R2 there are over .9. Unfortunately, it is impossible to draw any conclusive lesson from the results on the state dummies, although they are

19

interesting enough to suggest that institutional differences may have an important role in shaping regional inequality in Brazil. In general, the results of this study indicate that investments in public infrastructure and in education may help reducing regional inequality in Brazil. Another important aspect is participation in the labor force, for this variable appeared almost always as an important factor in the definition of income levels and income growth. Thus, provision of job opportunities seems to be a relevant factor. Even after considering the positive expected effects of public investments, there is some indication that income inequality is self-fulfilling, for wealthier states and states with dynamic labor markets tend to have higher income levels and to growth faster. Moreover, the importance of geographical variables indicates that probably a good deal of government intervention should be directed towards institution building, government efficiency improvement, etc. Even after controlling for variables related to human, household and social capital, there still is a great deal to be explained in terms of differences in income growth between Brazilian states.

20

References
Attanasio, O. and Browning, M. 1995. Consumption over the Life Cycle and the Business Cycle. American Economic Review. 85 (5): 1118-1136. Azzoni, C. 1999. Economic Growth and Regional Income Inequalities in Brazil. Annals of Regional Science. Forthcoming. Barro, R. and Sala-i-Martin, X. 1995. Economic Growth. New York, United States: McGraw Hill. ---- 1997. Technological Diffusion, Convergence, and Growth. Journal of Economic Growth. 2 (1): 1-26 Barro, R., Mankiw, G. and Sala-I-Martin, X. 1995. Capital Mobility in Neoclassical Models of Growth.NBER Working Paper 4206. Cambridge, United States: National Bureau of Economic Research. Blundell, R., Browning, M and Meghir, C. 1994. Consumer Demand and the Life-Cycle Allocation of Household Expenditures. Review of Economic Studies. 61 (1): 57-80. Browning, M., Deaton, A. and Irish, M. 1985. A Profitable Approach to Labor Supply and Commodity Demands Over the Life Cycle. Econometrica. 53 (3): 503-544. Chang, R. 1994. Income Inequality and Economic Growth: Evidence and Recent Theories. Economic Review, Federal Reserve Bank of Atlanta. 79: 1-9. Deaton, A. 1985. Panel Data from Time Series of Cross-Sections. Journal of Econometrics. 30: 109-126. Ferreira, A. H. and Diniz, C C. 1995 Convergencia entre las rentas per capita estaduales en Brasil. EURE-Revista Latioamericana de Estudios Urbano Regionales. 21 (62): 17-31. Hall,R. and Jones, C. 1996. The Productivity of Nations. NBER Working Paper 5812. Cambridge, United States: National Bureau of Economic Research Helliwell, John, 1996. Do Borders Matter for Social Capital? Economic Growth and Civic Culture in U.S. States and Canadian Provinces. NBER Working Paper 5863. Cambridge, United States: National Bureau of Economic Research. Islam, N. 1995. Growth Empirics: A Panel Data Approach. Quarterly Journal of Economics. 110 (4): 1127-1170. Jalan, J. and Ravallion, M. 1998a. Are There Dynamic Gains from a Poor-Area Development Program? Journal of Public Economics. 67 (1): 65 - 85. ---- 1998b. Geographic Poverty Traps? Discussion Paper No. 86. Boston, United States: Boston University, Institute for Economic Development.

21

Jones, C. 1995a. Times Series Tests of Endogenous Growth Models. Quarterly Journal of Economics. 110 (2): 495-525. ----. 1995b. R & D Based Models of Economic Growth. Journal of Political Economy. 101 (4): 759784. ---- 1997. On the Evolution of the World Income Distribution. NBER Working Paper 5812. Cambridge, United States: National Bureau of Economic Research. Moffitt, R. 1993. Identification and Estimation of Dynamic Models with a Time Series of Repeated Cross-Sections. Journal of Econometrics. 59: 99-123. Moulton, B. 1986. Random Group Effects and the Precision of Regression Estimates. Journal of Econometrics. 32: 385-397. Ravallion, M. 1998a. Poor Areas. In: A. Ullah and D. Giles, editors. Handbook of Applied Economic Statistics. New York, United States: Marcel Dekker, Inc. ----. 1998b. Reaching Poor Areas in a Federal System. Policy Research Group Working Paper 1901. Washington, DC, United States: World Bank. Ravallion, M. and Jalan, J. 1996. Growth Divergence Due to Spatial Externalities. Economic Letters. 53 (2): 227-232. Ravallion, M.and Wodon, Q. 1998. Poor Areas or Just Poor People? Policy Research Working Paper 1798. Washington, DC, United States: World Bank. Romer, P. 1986. Increasing Returns and Long Run Growth. Journal of Political Economy. 94: 1002-37. Solow, R.W. 1956. A Contribution to the Theory of Economic Growth. Quarterly Journal of Economics. 70: 65-94. Schwartsman, A. 1996. Convergence Across Brazilian States. Discussion Paper, n 02/96. So Paulo, Brazil: Universidade de So Paulo, IPE. Swan, T.W. 1956. Economic Growth and Capital Accumulation. Economic Record. 32: 33461. Zini, A. A., Jr. 1998. Regional Income Convergence in Brazil and its Socio-Economic Determinants. Economia Aplicada. 2 (2): 383-411.

22

Table 1: Description of the main variables Code

Dependent Variable: Monthly per capita household income Family: Gender of the household head (male = 1; female = 0) Education: years of education of household head Education of partner Participation in the labor market (hh head previous week) Participation of partner Participation of children (average) Household: Density occupants per room Availability of stove (yes = 1; no = 0) Availability of refrigerator (yes = 1; no = 0) Electricity (supplied with electricity = 1; not supplied = 0) Water (supplied by the public system = 1; not supplied = 0) Sewage (served by the public system = 1; not served = 0) Garbage (served by the public system = 1; not served = 0) Geographical: Urban/rural (urban = 1; rural = 0) Metropolitan region (metropolitan = 1; non-metropolitan=0) Life Expectancy (IDH) Infant Mortality (IDH) Average temperature in the Winter - Celsius degrees Latitude degrees Rainfall mm/year Metrop. Urban Life exp. Infant mort Temp Latitude Rain fall HH dens. Stove Fridge Electricity Water Sewage Garbage Gender_head Educ_head Educ_part Part_head Part_part Part_child y

Brazil Avg SD
71103 0,835 3,858 3,151 0,847 0,304 0,246 0,913 0,915 0,572 0,769 0,621 0,484 0,497 0,164 0,675 61,199 65,473 19,892 -13,315 103,039 156911 0,079 1,654 1,429 0,161 0,115 0,171 0,206 0,110 0,208 0,166 0,185 0,170 0,191 0,227 0,140 4,001 35,407 4,334 7,094 24,553

NE and CO Avg SD
56522 0,832 3,199 2,747 0,857 0,298 0,238 0,982 0,876 0,457 0,694 0,522 0,387 0,418 0,082 0,618 59,694 82,187 21,778 -8,839 92,485 121330 0,077 1,367 1,305 0,145 0,111 0,161 0,203 0,121 0,147 0,152 0,140 0,108 0,163 0,149 0,121 3,640 34,048 3,738 3,049 22,393

South and SE Avg SD

96097 0,840 4,987 3,845 0,830 0,315 0,260 0,796 0,983 0,771 0,898 0,792 0,650 0,633 0,303 0,773 63,778 36,821 16,661 -20,987 121,133 201583 0,081 1,486 1,366 0,183 0,119 0,187 0,152 0,016 0,136 0,092 0,116 0,122 0,157 0,266 0,114 3,196 10,758 3,241 5,240 16,079

23

Table 2 Correlation coefficient among variables

Ln(y) Ln(y) ln(y) Educ_head Gender_head Part_head Educ_part Part_children Stove Fridge HH Density Metrop Urban Water Electricity Sewage Garbage Rainfall Latitude Temp life expec. infant mort.
1 0.378* 0.710* 0.541* 0.208* 0.664* 1 0.134* 0.178* 0.110* 0.153* 1 0.497* 0.418* 0.946* 1 0.321* 0.656* 1 0.517* 1

ln(y) Educ_ Gende Part_ Educ_ Part_ head r head head part ch

Stove Fridge

HH Metrop Urban Water Electrici Sewage dens ty

Garbage Rainfall

Latit

-0.266* -0.095* 0.433* 0.558* 0.002 0.296* 0.508* 0.554* 0.455* 0.536* 0.505* 0.419* 0.607* 0.033 0.036 0.077* 0.008 0.063* 0.053* 0.039* 0.029 0.03 0.002 0.033

-0.594* -0.677* -0.233* -0.710* 1 0.476* 0.658* -0.012 0.043* 0.385* 0.048* 0.107* 1 0.666* 1

-0.051* 0.252* 0.495* 0.159* -0.011 0.255*

-0.187* 0.488* 0.414* 0.652* 0.617* 0.662* 0.605* 0.609* 0.314* 0.548*

-0.255* -0.492* -0.561* 1 -0.063* 0.316* 0.419* -0.072* 0.711* 0.773* 0.031 -0.019 0.054* 0.011 0.086* 0.063* 0.698* 0.838* 0.743* 0.881* 0.690* 0.821* 0.672* 0.818* 0.105* 0.490* 0.581* 0.788* -0.174* 1 -0.485* 0.605* -0.506* 0.448* -0.612* 0.376* -0.508* 0.535* -0.527* 0.447* -0.191* 0.073* -0.495* 0.377* -0.228* 1 0.785* 1 0.805* 0.818* 1 0.839* 0.853* 0.792* 0.864* 0.800* 0.834* 0.165* 0.327* 0.231* 0.602* 0.749* 0.630* 0.498* -0.592* -0.535* 0.463* 0.485* 0.681* 1 0.830* 0.364* 0.704* -0.510* 0.505* 1 0.270* 0.645* -0.541* 0.539* -0.539* 1 0.645* -0.454* 0.429* -0.605* 1 -0.858* 0.534* -0.756*

-0.078* -0.013

-0.039* 0.039* 0.494* -0.019 0.004 0.467*

-0.074* 0.264* 0.531* -0.059* -0.03 0.429*

-0.071* 0.093* 0.460* 0.144* 0.067* 0.050* 0.221* -0.002 0.384*

-0.546* -0.033 0.251* 0

-0.450* -0.099* -0.040* -0.333* -0.062* -0.547* -0.664* 0.451* 0.473* -0.157* 0.449* 0.362* 0.119* 0.419* 0.738*

-0.620* 0.158* -0.214*

-0.455* -0.01

-0.530* 0.018

-0.189* -0.391* -0.103* -0.519* -0.775* 0.563*

0.536* -0.624* -0.6175* -0.642*

24

Table 3 - Regression results for income levels dependent variable is ln (y)

(1) Constant Rainfall Latitude Temperature Age Age^2 Education head Participation head Gender head Participation partner Education partner Participation children Stove Fridge Density. Metropolitan Urban Sewage Electricity Garbage Education children Life Expectancy Infant Mortality Time Dummies State Dummies R2 x 0,48 X 0,88 X 0,89 X 0,91 -0,064 (0,692) 0,424 (0,222) 0,563 (0,113) 0,206 (0,155) (2) -2,942 (0,494) 0,236 (0,195) 0,327 (0,108) 0,351 (0,159) 0,009 (0,012) -0,0008 (0,0001 0,241 (0,051) 2,390 (0,227) 0,621 (0,400)( -0,855 (0,304) -0,024 (0,057) -1,442 (0,230) (3) -2,734 (0,421) 0,219 (0,198) 0,236 (0,107) 0,470 (0,199) 0,008 (0,019) -0,0001 (0,0002) 0,168 (0,035) 2,726 (0,257) 0,636 (0,418) -0,885 (0,289) -0,047 (0,044) 1,225 (0,230)-0,204 (0,251) 0,821 (0,280)-0,365 (0,188) (4) -2,369 (0,459) 0,301 (0,136) 0,284 (0,056) 0,464 0,129) 0,003 (0,013) -0,0001 (0,0001) 0,093 (0,029) 2,656 (0,206) 0,689 (0,299) -0,451 (0,131) -0,019 (0,035) 1,326 (0,161) -0,408 (0,243) 0,529 (0,278) -0,504 (0,135) 0,180 (0,118) 0,422 (0,247) -0,392 (0,256) -0,521 (0,303) 0,896 (0,231) (5) -0,436 (5,552) 0,303 (0,151) 0,281 (0,063) 0,455 (0,170) -0,003 (0,013) -0,00004 (0,0001) 0,089 (0,028) 2.637 (0,193) 0,617 (0,284) -0,439 (0,166) -0,017 (0,037) 1,320 (0,164) -0,374 (0,238) 0,528 (0,224) -0,457 (0,128) 0,191 (0,120) 0,401 (0,302) -0,381 (0,255) -0,523 (0,348) 0,894 (0,190) 0,138 (0,071) -0,401 (1,297) -0,030 (0,130) X 0,91 (6) -2,894 (0,720) 0,443 (0,153) 0,163 (0,081) 0,316 (0,170) 0,041 (0,009) -0,0004 (0,0001) -0,002 (0,035) 2,775 (0,156) 0,290 (0,261) (7) 3,697 (0,770)

0,063 (0,011) -0,001 (0,0001) 0,060 (0,024) 2,713 (0,127) 0,209 (0,200)

0,184 (0,249) -0,585 (0,138) 0,275 (0,131)

-0,061 (0,230) -0,781 (0,137) 0,469 (0,351)

0,949 (0,214)

0,320 (0,134)

-0,125 (0,123) X 0,89

-1,101 (0,239) X x 0,92

25

Table 4 Regression results for income growth Dependent variable is ln (y)

(5) (6) (7) (8) -2,176 -1,719 3,481 -1,274 (3,891) (0,468) (1,002) (0,470) Ln yt-1 -0,557 -0,486 -0,635 -0,550 (0,058) (0,061) (0,060) (0,074) Rainfall 0,136 0,186 0,304 (0,091) (0,073) (0,087) Latitude 0,165 0,079 0,143 0,047) (0,042) (0,050) Temperature 0,294 0,179 0,109 (0,091) (0,071) (0,084) Age -0,021 0,016 0,037 0,059 (0,009) (0,006) (0,009) (0,008) Age^2 0,0001 -0,0001 -0,0003 -0,0008 (0,0001) (0,0001) (0,0001) (0,0001) Education head 0,066 0,017 0,056 -0,014 (0,021) (0,018) (0,021) (0,015) Participation head 1,828 1,756 2,011 1,159 (0,147) (0,176) (0,159) (0,295) Gender head 0,354 0,233 0,231 -0,077 (0,211) (0,189) (0,201) (0,263) Participation partner -0,215 (0,101) Education partner -0,003 (0,027) Participation children 0,892 (0,111) Stove -0,173 (0,170) Fridge 0,384 0,100 -0,018 -0,279 (0,161) (0,143) (0,181) (0,161) Density. -0,269 -0,274 -0,473 -0,182 (0,087) (0,083) (0,109) (0,118) Metropolitan 0,051 0,098 0,402 0,129 (0,063) (0,052) (0,267) (0,065) Urban 0,303 (0,224) Sewage -0,182 (0,137) Electricity -0,478 (0,261) Garbage 0,532 0,514 0,229 0,650 (0,135) (0,122) (0,110) (0,166) Education children 0,032 (0,038) Life expectancy. 0,069 (0,894) Infant mortality 0,038 -0,051 -0,931 -0,133 (0,079) (0,062) (0,252) (0,070) Time Dummies X X X X X X X X State Dummies X R2 0,29 0,52 0,52 0,56 0,56 0,52 0,58 0,66 Regression (8) uses as the dependent variable the income rate of growth over three year periods (third differences); regressors are taken at their levels in the initial years of the periods. Constant (1) -0,114 (0,090) -0,026 (0,029) 0,006 (0,027) 0,032 (0,024) 0,028 (0,016) (2) -1,694 (0,293) -0,453 (0,065) 0,067 (0,091) 0,150 (0,064) 0,209 (0,081) -0,006 (0,007) 0,0001 (0,00008) 0,124 (0,024) 1,518 (0,166) 0,371 (0,266)( -0,395 (0,151) 0,001 (0,029) 0,821 (0,124) (3) -1,545 (0,298) -0,476 (0,068) 0,065 (0,091) 0,111 (0,064) 0,274 (0,113) 0,007 (0,011) 0,0001 (0,0001) 0,093 (0,020) 1,716 (0,196) 0,381 (0,290) -0,428 (0,149) -0,010 (0,026) 0,762 (0,135) -0,106 (0,167) 0,406 (0,138) -0,180 (0,109) (4) -1,504 (0,312) -0,560 (0,057) 0,122 (0,082) 0,157 (0,042) 0,314 (,069) -0,009 (0,008) 0,00009 (0,0001) 0,067 (0,021) 1,822 (0,160) 0,423 (0,223) -0,203 (0,077) -0,007 (0,026) 0,891 (0,107) -0,237 (0,170) 0,341 (0,173) -0,285 (0,086) 0,063 (0,062) 0,245 (0,174) -0,196 (0,151) -0,395 (0,209) 0,570 (0,153) -

26

Table 5 Convergence results Columns from table 4 Constant 0.0153 (0.0334) Ln (yt-1) -0.0086 (0.0086) 0.0086 Implied Years to half convergence 79.9 Time Dummies Geographical Variables Age and age^2 Human Capital Household capital Public Infra-structure IDH variables State dummies R2 (1) (2) (3) (4) (5) 0.0114 -1.694 -1.545 -1.504 -2.176 (0.090) (0.293) (0.298) (0.312) (3.891) -0.026 -0.453 -0.476 -0.560 -0.557 (0.029) (0.065) (0.068) (0.057) (0.058) 0.0263 0.6033 0.6463 0.8210 0.8142 26.3 1.2 1.1 0.81 0.81 x x x x x x x x x x x x x x x x x x x x x x x x 0.56 (7) -1.990 (0.494) -0,635 (0,060) 1.0000 0.68 x x x x x x x 0,58 (6) 3.607 (1.054) -0,486 (0,061) 0.6655 1.04 x x x x x x x 0,52

0.0005

0.29

0.52

0.55

0.56

Table 6 - Decomposition of income differentials among regions Southeast South Center-West - Northeast - Northeast - Northeast Income 0,816 0,833 0,816 Rainfall 0,173 0,245 0,246 Latitude 0,178 0,222 0,141 Temperature -0,053 -0,142 -0,022 Head education 0,055 0,047 0,027 Head participation -0,094 -0,021 0,049 Gender 0,001 0,009 0,009 Fridge 0,056 0,060 0,036 Density 0,081 0,119 0,069 Metropolitan 0,075 0,023 -0,030 Garbage 0,358 0,180 0,128 Infant mortality 0,114 0,132 0,113 Decomposition used: [X1BARSE-X1BARNE].. Coefficients estimated in table 3, Column (6)

27

Table 7 - Interstate Gini Coefficients Columns in Table 3 Gini coefficient Average Income 0,6044 (1) 0,5139 (2) 0,4808 (4) 0,4830 (5) 0,4831

Variable for Columns (1)-(4): Yi = Yi (Xi XBar). (i referring to state) (1) Income, controlling for the effects of geography. (2) Income, controlling for the effects of geography and human capital. (4) Income, controlling for the effects of geography, human capital, household capital, urban and metropolitan status, and infra-structure (5) Income, controlling for the effects of geography, human capital, household capital, urban and metropolitan status, infra-structure, and human development variables

28

Table 8 Variables averages and standard deviations by state Variable So Paulo Rio de Janeiro Minas Gerais Paran Santa Catarina Avg SD Avg SD Avg SD Avg SD Avg SD Y 12097 246894 117826 236551 70110 14247 84753 17165 94613 205678 0 0 6 Gender 0,836 0,076 0,797 0,087 0,828 0,084 0,858 0,066 0,875 0,076 head Educ_head 5,551 1,255 6,292 1,050 4,239 1,246 4,363 1,466 4,860 1,445 Educ_part 4,080 1,308 4,344 1,199 3,408 1,312 3,402 1,378 4,003 1,360 Part_head 0,813 0,204 0,791 0,218 0,830 0,165 0,856 0,156 0,836 0,182 Part_part 0,261 0,099 0,274 0,090 0,262 0,106 0,322 0,106 0,382 0,120 Part_child 0,259 0,190 0,210 0,160 0,255 0,185 0,293 0,193 0,295 0,204 Density 0,834 0,153 0,785 0,127 0,796 0,136 0,861 0,168 0,744 0,153 Stove 0,988 0,011 0,984 0,010 0,983 0,011 0,978 0,017 0,976 0,026 Fridge 0,881 0,068 0,878 0,065 0,601 0,101 0,690 0,114 0,838 0,078 Electricity 0,963 0,075 0,973 0,019 0,806 0,101 0,853 0,096 0,928 0,039 Water 0,934 0,033 0,869 0,062 0,742 0,072 0,751 0,095 0,734 0,144 Sewage 0,851 0,041 0,750 0,044 0,681 0,046 0,572 0,032 0,506 0,061 Garbage 0,890 0,032 0,722 0,065 0,527 0,082 0,619 0,104 0,543 0,132 Metropolit 0,514 0,030 0,804 0,028 0,220 0,029 0,246 0,024 an Urban 0,923 0,013 0,941 0,016 0,739 0,036 0,717 0,056 0,654 0,065 Life exp. 63,78 3,31 62,27 2,72 62,74 3,38 63,71 3,00 65,38 2,83 Infant 36,43 9,10 38,11 11,94 41,42 10,50 41,98 10,58 34,46 9,33 mort Temperatu 16,82 18,32 18,26 14,12 13,46 re Latitude -21,27 -21,16 -18,17 -22,85 - -26,31 Rain fall 127,98 113,36 119,45 132,17 - 143,9 0

Rio Grande do Sul Avg SD 10881 22725 3 3 0,836 0,078 5,213 4,132 0,842 0,391 0,244 0,741 0,986 0,816 0,899 0,793 0,549 0,654 0,340 1,322 1,333 0,173 0,119 0,168 0,146 0,010 0,076 0,053 0,086 0,052 0,085 0,029

Esprito santo Mara Avg SD Avg 75598 15214 34395 6 0,849 0,076 0,835 4,393 3,542 0,840 0,314 0,267 0,810 0,985 0,690 0,865 0,723 0,638 0,478 1,385 1,356 0,168 0,112 0,193 0,138 0,017 0,118 0,089 0,092 0,048 0,084 2,340 2,108 0,898 0,380 0,252 1,193 0,634 0,302 0,507 0,294 0,246 0,134

0,741 0,049 0,695 0,062 0,373 65,46 2,72 63,10 2,91 59,84 26,07 6,27 39,28 7,95 97,08 12,89 -26,96 122,44 22,76 -10,19 88,63 25,99 -5,03 116,92

29

Variable Y

Gender head Educ_head Educ_part Part_head Part_part Part_child Density Stove Fridge Electricity Water Sewage Garbage Metropolit an Urban Life exp. Infant mort Temperatu re Latitude Rain fall

Table 8 Variables averages and standard deviations by state (continued) Rio Grande do Pariba Pernambuco Alagoas Sergipe Bahia Mato Grosso Norte do Sul Avg SD Avg SD Avg SD Avg SD Avg SD Avg SD Avg SD 50877 104970 43305 88464 55808 10899 50889 111278 53628 10988 57743 11403 90099 18299 2 1 5 0,833 0,085 0,814 0,083 0,808 0,077 0,824 0,090 0,805 0,091 0,823 0,072 0,861 0,061 3,276 3,044 0,838 0,290 0,218 0,977 0,867 0,457 0,784 0,578 0,418 0,560 0,665 58,45 107,19 1,196 1,328 0,160 0,105 0,158 0,173 0,063 0,110 0,108 0,066 0,068 0,090 0,047 4,07 27,67 3,236 3,050 0,831 0,286 0,217 0,946 0,940 0,403 0,734 0,600 0,428 0,492 1,279 1,398 0,146 0,113 0,146 0,179 0,027 0,101 0,106 0,067 0,053 0,079 3,479 2,879 0,824 0,288 0,230 0,938 0,938 0,468 0,785 0,595 0,452 0,461 0,410 1,031 1,170 0,159 0,087 0,154 0,159 0,024 0,091 0,079 0,077 0,056 0,075 0,038 2,780 2,239 0,821 0,271 0,236 0,994 0,845 0,442 0,735 0,500 0,352 0,455 0,588 56,71 122,99 23,63 -8,67 91,10 1,132 1,135 0,178 0,100 0,156 0,180 0,070 0,112 0,111 0,078 0,062 0,110 0,070 2,76 17,54 3,084 2,666 0,843 0,313 0,236 0,962 0,923 0,497 0,751 0,593 0,407 0,473 0,613 59,07 79,95 22,76 -10,19 88,63 1,199 1,217 0,160 0,109 0,165 0,186 0,035 0,130 0,119 0,091 0,066 0,098 0,101 3,41 14,49 3,071 2,467 0,869 0,304 0,229 0,952 0,892 0,417 0,669 0,510 0,372 0,357 0,224 0,579 60,52 70,91 18,69 -10,84 76,33 1,088 1,050 0,129 0,090 0,157 0,159 0,037 0,074 0,085 0,057 0,053 0,058 0,040 0,054 2,92 11,61 4,284 3,420 0,874 0,273 0,247 0,842 0,967 0,669 0,820 0,691 0,505 0,600 0,766 63,06 37,02 13,04 -14,28 84,08

1,467 1,366 0,129 0,114 0,168 0,139 0,019 0,109 0,103 0,117 0,021 0,105

0,651 0,050 0,720 0,041 57,53 3,55 58,84 2,76 113,48 25,94 100,7 22,36 7 22,84 21,97 -6,83 67,76 -8,05 63,46 -

0,054 3,26 9,51

30