SciELO - Scientific Electronic Library Online

vol.60 issue4Equivalent martingale measures and Lévy processesMedindo a fidelidade das torcidas brasileiras: uma análise econômica no futebol author indexsubject indexarticles search
Home Pagealphabetic serial listing  

Services on Demand




Related links


Revista Brasileira de Economia

Print version ISSN 0034-7140On-line version ISSN 1806-9134

Rev. Bras. Econ. vol.60 no.4 Rio de Janeiro Oct./Dec. 2006 

Economic integration, poverty and regional inequality in Brazil*



Joaquim Bento de Souza Ferreira FilhoI; Mark Jonathan HorridgeII

IEscola Superior de Agricultura "Luiz de Queiroz", Universidade de São Paulo. Departamento de Economia, Sociologia e Administração. Av. Pádua Dias, 11. Piracicaba, SP. CEP - 13.418-900. Fone: (019) 34178700.
IICentre of Policy Studies, Monash University. Melbourne, Australia. Email:




In this paper we address the issue of how trade integration can affect poverty and income distribution in Brazil. The analysis is conducted through a static General Equilibrium and Microsimulation Model of Brazil, calibrated with the PNAD 2001 data. The model comprises 112,055 Brazilian households and 263,938 adults, distinguishing 42 activities, 52 commodities, and 27 regions. The FTAA formation is simulated, and its impacts upon poverty and income distribution in Brazil analyzed. Results point to the fact that even a large shock like the one simulated wouldn't have a strong impact on poverty reduction in Brazil, although the results are concentrated on the poorest households.

Keywords: Trade Integration; Poverty; Income Distribution; Micro-simulation; Applied General Equilibrium Models.
JEL Code: F15; C68; I32; R59.


Este trabalho analisa os efeitos potenciais da formação da ALCA sobre os níveis de pobreza e distribuição de renda no Brasil. A análise é conduzida através de um Modelo Aplicado de Equilíbrio Geral e de Micro-simulação estático, calibrado com os dados da PNAD 2001. O modelo distingue na sua estrutura 112.055 domicílios e 263.938 adultos, 42 atividades produtivas, 52 produtos, e 27 regiões. Os resultados apontam para o fato de que mesmo mudanças tarifárias grandes como as aqui simuladas não trariam um forte impacto sobre a pobreza no Brasil, embora os resultados estejam concentrados nos domicílios mais pobres.




One of the most striking aspects of the Brazilian economy is its high degree of income concentration. Despite the changes the economy has faced in the last twenty years, ranging from the country's re-democratization, trade liberalization, hyperinflation, many currency changes, and finally, to the macroeconomic stabilization in the mid-nineties, the country still shows one of the worst patterns of income distribution in the world. The resilience of this income distribution problem has attracted the attention of many researchers all over the world, and is the central point of a lively debate in Brazil. The problem is, of course, extremely complex, related to a great number of socio-economic variables, which makes it a particularly difficult analytical issue, since the effects of many variables upon poverty are uncertain.

At the same time, new changes in the economic environment now challenge the Brazilian economy. Among them, the participation of the country in new free trade areas may be one of the most important. A complex phenomenon in itself, the economic integration poses new questions relating to the prospects for the poor. This paper is an attempt to address these questions with a systematic and quantitative approach. For this purpose, an applied general equilibrium model of Brazil tailored for income distribution and poverty analysis will be used. The model has also an inter-regional breakdown, which will make it possible to assess the regional inequality associated issue.

The plan of the paper goes as follows: the next section shows some figures about the problem of poverty and income distribution in Brazil, with a brief review of the recent literature on the topic. Then, we present the methodological approach to be pursued here, with a discussion of the relevant literature on the many different approaches. Then the model itself is presented, with a discussion of its main aspects and of the database. Finally, results and conclusions are presented.



It has long been recognized that, although Brazil is a country with a large number of poor people, its population is not among the poorest in the world. Based on an analysis of the 1999 Report on Human Development, Barros et al. (2001a) show that around 64% of the countries in the world have per capita income less than in Brazil, a figure that mounts to 77% if we consider the number of persons in the same condition. The same authors show that, while in Brazil 30% of the total population is poor, on average only 10% are poor in other countries with similar per capita income. Indeed, based on the same report the authors define an international norm that, based on per capita income, would impute only 8% of poor for Brazil. That is, if the inequality of income in Brazil were to correspond to the world average inequality for countries in the same per capita income range just 8% of the Brazilian population would be expected to be poor.

Taking the concept of poverty in its particular dimension of income insufficiency, the same authors show that in 1999 about 14% of the Brazilian population lived in households with income below the line of extreme poverty (indigence line, about 22 million people), and 34% of the population lived in households with income below the poverty line (about 53 million people). Even though the percentage of poor in the population has declined from 40% in 1977 to 34% in 1999, this level is still very high and, it seems, stable. The size of poverty in Brazil, measured either as a percentage of the population or in terms of a poverty gap, stabilizes in the second half of the eighties, although at a lower level than was observed in the previous period.

Barros and Mendonça (1997) have analyzed the relations between economic growth and reductions in the level of inequality upon poverty in Brazil. Among their main conclusions, these authors point out that an improvement in the distribution of income would be more effective for poverty reduction than economic growth alone, if growth maintained the current pattern of inequality. According to these authors, due to the very high level of income inequality in Brazil it is possible to dramatically reduce poverty in the country even without economic growth, just by turning the level of inequality in Brazil close to what can be observed in a typical Latin American country.

The poverty in Brazil has also an important inter-regional dimension. According to calculations due to Rocha (1998), in a study for the 1981/95 period, the South-East region of the country, while counting for 43.84% of total population in 1995 had only 33% of the poor. These figures were 15.37% for the South region (8.15% of poor), and 6.81% for the Center-West region (5.23% of poor). For the poorer regions, on the contrary, the share of population in each region is lower than the share of poor: 4.56% (9.32% of poor) for the North region, and 29.42% (44.31% of poor) for the North-East region, the poorest region in the country.

In terms of evolution of regional inequality, Rocha concludes that no regular trend could be observed in the period. Moreover, the author also concludes that the yearly observed variations in concentration are mainly related to what happens in the state of São Paulo (South-East region) and in the North-East region. This reinforces the position of these two regions in the extremes of the regional income distribution in Brazil. The author also points out that once the effects of income increase that followed the end of the hyper-inflation period in 1995 run out, the favorable evolution in the poverty indexes and its spatial incidence will depend mainly on the macroeconomic determinants related to investment. Also, the author concludes that even keeping unchanged the actual level of poverty, the reduction in the regional inequality will require the reallocation of industrial activity to the peripheral regions.

And, finally, the same author also concludes that the opening of the economy to the external market (mainly in relation to the formation of Mercosur) would help reduce regional inequality in Brazil. This would happen through reduced consumer prices in the poorest regions, which are fortunately lacking in the industries most threatened by new trade flows.

Green et al. (2001) analyzed the behavior of wages and the allocation of labor throughout the 1980-99 trade liberalization period in Brazil. Among the main findings the authors point out that wage inequality remained fairly constant for the 1980s and 1990s, with a small peak in the mid 80s. The main conclusion of the study is that the egalitarian consequences of trade liberalization were not important in Brazil for the period under analysis. As caveats, the authors note the low trade exposure of the Brazilian economy (around 13% in 1997), as well as the low share of workers that have completed college studies in total (1 in 12 workers at that time).



Computable general equilibrium (CGE) models have long been used for poverty analysis. In the traditional analysis, however, the Representative Household formulation has been used to represent consumer behavior in the model. This formulation, although adequate for many purposes, limits the investigation of poverty and income distribution analysis. More recent approaches were developed to deal with these constraints.

Savard (2003) provides a thorough discussion on the topic. According to that author, the models dealing with poverty and income distribution analysis can be classified into three main categories: models with a single representative household (RH), models with multiple-households (MH), and the micro-simulation approach that links a CGE model to an econometric household micro-simulation model.

The Representative Household model is the traditional method, and has been widely used in the literature. The main drawback of this model for income distribution and poverty analysis is that there are no intra-group income distribution changes, as the households are all aggregated into a representative one. This, of course, limits the scope for economic behavior in the model.

The second approach, the multiple-household model (MH), consists of multiplying the number of households. Increasing computation capacity allows us to have a large number of households in the model. To take an extreme case, the total number of households in a household survey could be used. This approach then allows the model to take into account the full detail in household data, and avoids pre-judgment about aggregating households into categories. The main disadvantages of this type of approach are that data reconciliation can be difficult, and that the size of the model can become a constraint.

The third approach, which we call MS, draws on micro-simulation techniques. Here, a CGE model generates aggregate changes that are later communicated to a micro-simulation model based on a large unit record database. Savard (2003) points out that the drawbacks to the approach are coherence between models, since the causality usually runs from the CGE model to the micro-simulation model, with no feedback between them.

The approach pursued in this paper takes advantage of the same general idea raised by to overcome the difficulties posed by the three first options abovementioned: the use of a CGE model linked to a micro-simulation model, but with a bi-directional linkage between them that would guarantee a convergence of solution for both models. Savard (2003) links the models by running them in a repeated sequence of CGE-MS model runs, first computing the CGE simulation, then the MS model simulation, in a looping way, until convergence occurs. The main advantages of this approach are that: there is no obligation to scale microeconomic data to match the aggregated macro data; we can accommodate more households in the MS model; and the MS model may incorporate discrete-choice or integer behavior that might be difficult to incorporate in the CGE model.

The CGE model used here is a static inter-regional model of Brazil based on the well-known ORANIG model of Australia (Horridge, 2000). The model's structure is quite standard: consumption is modeled through the Linear Expenditure System over composite commodities (domestic and imported); production for exports or domestic markets are regulated by CET functions for each firm, production is a nested LEONTIEF/CES structure for primary factors and composite inputs, labor is a CES function of 10 different types of labor. This non-linear model is written in linearized form, solved with the GEMPACK software, and distinguishes between 42 sectors and 52 commodities;1 10 labor occupational categories; and 27 regions inside the country, using a top-down technology. The CGE model was calibrated with data from the Brazilian economy for 1996, obtained from two main sources: the 1996 Brazilian Input-Output Matrix (IBGE., and the Brazilian Agricultural Census (IBGE, 1996a).

On the income generation side of the model, workers are divided into 10 different categories (occupations), according to their wages. These wage classes are then assigned to each regional industry in the model. Together with the revenues from other endowments (capital and land rents) these wages will be used to generate household incomes. We extend the CGE model to cover 270 different expenditure patterns, composed of 10 different income classes in 27 regions.

There are two main sources of information for the household micro-simulation model: the Pesquisa Nacional por Amostragem de Domicílios -PNAD (National Household Survey - IBGE 2001), and the Pesquisa de Orçamentos Familiares- POF (Household Expenditure Survey, IBGE 1996b). The PNAD contains information about households and persons, and shows a total of 331,263 records. The main information extracted from PNAD were wage by industry and region, as well as other personal characteristics such as years of schooling, sex, age, position in the family, and other socio-economic characteristics.

The POF, on the other hand, is an expenditure survey that covers 11 metropolitan regions in Brazil. It was undertaken during 1996, and covered 16,014 households, with the purpose of updating the consumption bundle structure. The main information we drew from this survey was the expenditure patterns of 10 different income classes, for the 11 regions. We assigned one such pattern to each individual PNAD household, according to each income class. As for the regional dimension, the 11 POF regions were mapped to the larger set of 27 CGE regions. Here it must be stressed that the POF survey just brings information about urban areas (the metropolitan areas of the main state capitals).

3.1. Model running procedures and highlights

As mentioned before, our model consists of two main parts: a Computable General Equilibrium model (CGE) and a Household Model (MS). Our approach for the analysis consists in running the two models sequentially, whilst obtaining consistency between them. The logical sequence of this procedure, as well as more details, is described in this section.

The process starts with a run of the CGE model. The trade shocks are applied, and the results calculated to 52 commodities, 42 industries, 27 regions, and 10 labor occupations. The results from the CGE model, then, are used to update the MS model. This update consists basically in updating wages and changes in labor demand, for the 263,938 workers in the sample. These changes have a regional (27 regions) as well as sector (42 industries) dimension.

The model then relocates jobs according to changes in labor demand.2. This is done changing the PNAD weights of each worker (see Appendix for details) to mimic the change in employment. This procedure was called the " quantum weights method".3 In this approach, then, there is a true job relocation process going on. If, as occasionally occurs, some region has insufficient unemployed workers in some occupational category, the already employed workers will increase the number of hours worked to meet the increasing labor demand. Having updated the database, the expenditure results from the MS model are fed back into the CGE model, until the convergence of the results.4 Once the final results are obtained, the change in poverty indexes are calculated and reported.

After that, a new updated income matrix is generated, for the total number of records in the original database (PNAD). This post-simulation matrix has the same number of records as the original one (263,938), and keeps unchanged the original link between workers and households.

One final point about the procedure used in this paper should be stressed. Although the changes in the labor market are simulated for each adult in the labor force, the changes in expenditures and in poverty are tracked back to the household dimension. This is possible since PNAD has a key that links persons to household, that's to say, we know to which household each person belongs. Each household contains one or more adults, either working in a particular sector and occupation, or unemployed. In our model then it is possible to recompose changes in the household income from the changes in individual wages. This is a very important aspect of the model, since it is likely that changes in employment records are cushioned, in general, by this procedure. If, for example, one person in some household loses his job but another in the same household gets a new job, household income may change little, or not at all. Moreover, since households are the expenditure units in the model, we would expect household spending to be cushioned by this income pooling effect.



In this section we extend the above description of poverty and income inequality in Brazil. The reference year for our analysis is 2001. Some general aggregated information about poverty and income inequality in Brazil can be seen in Table 1.



The rows of Table 1 correspond to household income classes, grouped according to POF definitions5, such that POF[1] is the lowest income class, and POF[10] the highest. A fair picture of income inequality in Brazil emerges from the table. We see that the first 5 income classes, while accounting for 52.6% of total population in Brazil, get only 17% of total income. The highest income class, on the other hand, accounts for 11% of population, and about 45% of total income. The Gini index associated with the income distribution in Brazil in 2001, calculated using an equivalent household6 basis, is 0.58, placing Brazil's income distribution among the world's worst.

The unemployment rate is also relatively higher among the poorer classes. This is a very important point to be noted, due to its relevance for modeling. The opportunity to get a new job is probably the most important element driving people out of poverty: hence the importance for poverty modeling of allowing the model to capture the existence of a switching regime (from unemployment to employment), and not just changes in wages. As can be seen in Table 1 above, the unemployment rate reaches 36.5% among the lowest income group (persons above 15 years), and just 7.7% among the richest.

For the purpose of further describing the state of income insufficiency in Brazil we set a poverty line defined as one third of the average household income.7 According to that criterion 30.8% of the Brazilian households in 2001 would be poor.8 This would comprise 96.2%, 76.6% and 53.5% respectively of households in the first three income groups,9 or 34.5 million out of 112 million households in 2001.

The Table 2 below shows how each POF group contributes to the three Foster et al. (1984) (FGT, for short) overall measures of poverty:

FGT0: the proportion of poor households (i.e., below the poverty line);

FGT1: the average poverty gap (proportion by which household income falls below the line);

FGT2: measures the extent of inequality among the poor.



As stated before, this general poverty and inequality picture also has an important regional dimension in Brazil. This is a consequence of the spatial concentration of economic activity, which is located mainly in the South-East region. This is particularly true of industrial activity; agriculture is more dispersed among regions. Table 3 shows more information about the regional dimension of poverty and income inequality in Brazil. The map, Figure 1, shows where regions are located, and shades them according to proportions of households in poverty.





As can be seen in Table 3, the states in the North (N) region account for 8% of total population, compared to 23.5% for the North-East (NE), 45% in the South-East (SE), 16% for South (S), and 7.2% for the Center-West (CW). In the SE region the state of São Paulo alone accounts for 22.9% of total Brazilian population.

The next column in Table 3 shows the share of households below the poverty line in each region, as a proportion of total regional households. As can be seen, the states in the NE region (states numbered from 8 to 16 in the table) plus the states of Tocantins and Para in the N region present the highest figures for this indicator, showing that these states are relatively poorer. If, however, regional population is taken into account, the third column show that the populous regions of Ceará, Pernambuco, Bahia, Minas Gerais and São Paulo give higher contributions to the Foster-Greer-Thorbecke poverty gap index.10 These figures are the contribution of each state to the total poverty gap index in Brazil expressed as a proportion of the poverty line (see column total). We can see that the average poverty gap in Brazil in 2001 is a 14.5% insufficiency of income to reach the poverty line.

The last column in the table above shows the regional insufficiency gap. The picture is similar to what was seen for the number of households below the poverty line, with the states in the NE regions plus the states of Tocantins and Para showing the highest poverty gaps. Two states in the South region (Santa Catarina and Rio Grande do Sul) show the lowest poverty gaps in Brazil, followed closely by São Paulo. Interesting enough, Amapa state (in the North region) shows a poverty gap in line with the richer states of the S-SE. This result, however, should be viewed with caution, since that state has a very small share of total population, which could cause the result to be a sampling bias. Besides that, the PNAD survey doesn't cover the rural areas of the Northern states, where poverty is usually concentrated.

More information about the labor structure of the economy can be seen in Table 4 and Table 5. In these tables sectoral wage bills are split into the model's 10 occupational groups. The occupational groups are defined in terms of a unit wage ranking. More skilled workers, then, would be those in the highest income classes, and vice-versa. As can be seen in Table 3, Agriculture is the activity that uses more unskilled labor (40.5% of that sector's labor bill), while Petroleum and Gas Extraction and Petroleum Refinery are the most intensive skilled labor (10th labor class) using activities, with Financial Institutions coming next. If labor inputs were measured in hours (rather than in values) the concentration of low-skill labor in Agriculture would be even more pronounced.





Agriculture is also the sector that hires the highest share of unskilled labor in Brazil, around 41% of total workers in income class 1 (Table 5). The Trade sector is the second largest employer of this type of labor. As for the higher income classes, we see that the Financial Institutions and Public Administration sectors hire the largest numbers of well-paid workers.

And, finally, Table 6 shows the distribution of occupation wages (OCC) classes among the household income classes (POF classes).



In the table above the rows show household income classes, while the columns show the wages by occupation. It is evident from this table that the wage earnings of the higher wage occupations (OCC10, for example) are concentrated in the higher income households, and vice-versa. Most of the wages earned by workers in the first wage class (OCC1) accrue to the three poorest households, POF[1]-[3]. All the workers in the highest wage class, on the other hand, are located in households from the 8th income class and above.



We will simulate the effects of a trade liberalization shock in the context of the Free Trade Area of Americas (FTAA) formation. As there is no probable detailed scenario arising so far from the negotiating process, we will use here a hypothetical 100% cut in all tariffs in trade between Brazil and its trade partners in the block. The shocks to be applied draw on previous work of the authors , and are generated by tariff changes and prices adjustments results from a previous run of the GTAP11 model with a linked (embedded) detailed Brazilian model, using a methodology described in Horridge and Ferreira Filho (2003).

The shocks to be transmitted to the PAID-BR (Poverty Analysis and Income Distribution Brazilian Model) are the Brazilian export quantities, the CIF import prices and the import tariff shocks to Brazil arising from the tariff liberalization shocks in the global model.

5.1. Model's closure

In this section we describe the main aspects of the model's closure used. First, among the many different possible choices for the labor market closure, in this paper we have chosen to hold real wages fixed, with employment adjusting in each industry. With fixed wage relativities, the share of each occupation in each industry is also fixed, meaning that each activity will hire fixed proportions of the 10 model occupations.

In the capital market the capital stock in each sector is held fixed, with rates of return to capital adjusting endogenously. The same applies for land stocks (used just in the Agriculture activity). This closure has a short run flavor in the sense that capital and land stocks cannot adjust in the short run. The ratio investment/consumption is also fixed. The trade balance is fixed, with real consumption, investment and government spending moving together to accommodate it. The trade balance, then, drives the level of these three last macro aggregates. And, finally, the aggregated (national) consumer price index is the model's numeraire.



6.1. The CGE model results

The Brazilian economy is little oriented to external trade. The shares of exports and imports in total GDP were respectively 7% and 8.9% in the 1996 base year. These shares have increased recently, but not by enough to significantly change this picture. Table 6 shows more information about the nature and size of the shocks applied to the model, as well as about the structure of Brazilian external trade. The final column shows simulated changes in output.

As stated before, the shocks applied to the model were generated by a previous run of the GTAP model. The GTAP effects on the Brazilian economy were then transmitted to the PAID-BR model through the following channels: export quantities, foreign currency import prices, and the aggregated (over regions in the global model) trade weighted import tariffs calculated in the GTAP model, version 5 database.

An inspection of Table 7 can give an idea of the importance of these shocks combined with the importance of each commodity in Brazilian external trade. As can be seen, Brazilian exports are spread among many different commodities, with no specialized trend. Imports as a share of each commodity domestic production are concentrated in Wheat, Oil, Machinery, Electric Materials and Electronic Equipment, and Chemical Products. In terms of total imports shares, however, Oil Products (Raw and Refined), Machinery, Electric Materials and Electronic Equipment, and Chemical Products are the most important products.



The changes in the foreign currency import prices in the model are generated by the world price adjustments in the global model. From the export side, we see that there is an export push arising from the trade liberalization in some of the Brazilian main export products: Iron Products, Machinery and Tractors, Other Vehicles and Spare Parts, and Processed Vegetable Products (VegetProcess), to cite the most import products in terms of exported share in the base year. On the other hand exports of Minerals and Vegetable Oils12 contract. From the import side there is a general fall in import tariffs, only partially counteracted by higher world prices.

In what follows, we present some macro results in order to establish a benchmark for the regional and poverty analysis. When interpreting these results one should bear in mind that the model has a " top-down" inter-regional specification, meaning that the national model is solved before the inter-regional one, being exogenous to it.

The first observed result of our simulation is an increase in activity level in the model, as a result of trade liberalization. The increase in exports, consumption, government consumption and investment (which follow household consumption by means of the closure) outweigh the increase in imports, causing GDP to rise by 0.68%. The real exchange rate rises, with corresponding gains in the external terms of trade.

For factor market results, recall that sectoral capital and land are fixed, while employment adjusts to accommodate fixed real wages. As we can see, the average (aggregated) capital rental shows a 1.61% increase. With capital stocks fixed, output increases require employment increases (1.06% overall); so falling capital/labor ratios increase the marginal productivity of capital, and hence capital returns. The price of land also shows a strong increase, reflecting the increase in production of activities using this factor (Agriculture).

Aggregate employment measured using wage bill weights rose by 1.06%, but rose more in terms of hours worked (PNAD head weights): 1.5%. This means that not only did employment rise, but employment patterns also shifted towards the sectors where low-wage workers were employed - boding well for a more equal income distribution.



Table 9 shows results at regional level. With real wages fixed and the CPI acting as a numeraire, each region's wage bills will change in proportion to (wage-weighted) regional employment. The change in aggregate labor demand will be distributed among regions according to their activity level changes. As can be seen in Table 9, some of the more populous states in Brazil (Sao Paulo, Rio de Janeiro and Bahia) show a smaller increase in regional employment. Espirito Santo state, on the other hand, is the one where employment increases the most, a result due to an increase in the production of one commodity (coffee) that is very important for the local economy. But this is a small state compared with the above-mentioned.



6.2. Poverty and income distribution results

We saw in the previous section that model results are differentiated among regions, and among different household income classes. The results of these changes upon the poverty and income inequality measures are presented in Table 10, and discussed below. As it can be seen, the GINI index fell by 0.32% as a result of the simulated scenario.



It should be noted that in the model there is no substitution among workers in different wage classes, which we use as a proxy for skills. The fall in unemployment is a compositional effect arising from the uneven change in economic activity among different regions and sectors. These results show, then, that the integration scenario we simulate would be more beneficial, in terms of reduced unemployment, for the poorest households.

The next table summarizes the results for each household contribution to the FGT indexes (compare with Table 2).



We can see from the table that the three inequality measures are slightly reduced, with again the reductions concentrating in the poorest households: the proportion of poor households, the poverty gap and the extent of inequality all fall in the poorest households. The fall in the number of poor amounts to a 1.99% fall in aggregate poverty if the calculation is performed over households, and 1.77% if over persons.

And, finally, Table 12 show model results relating to the regional breakdown inside Brazil.



It can be seen from Table 12 that the total number of households that would leave poverty in Brazil as a result of the trade liberalization simulation would amount to 307,333, corresponding to 1,074,620 persons. In regional terms, some states in the South and South-Eastern regions would concentrate the bulk of the effects: Minas Gerais, Espirito Santo, São Paulo, and Rio Grande do Sul, while accounting for about 40% of total population, would concentrate about 50% of persons leaving poverty.

It should be also stressed that even in percent terms the states in the South and South-East regions would be more benefited. As seen in Table 9, this is a consequence of the increase in the activity level in these regions, and of the particular composition of labor demand in regional booming industries. The results, then, point to the conclusion that the more developed regions in the country would gain more (although slightly) in terms of poverty alleviation than the poorer states of the North-East and North regions.



A series of points should be highlighted in wrapping up this discussion. As we could see, model results show that even an important tariff shock as that applied here could be not enough to generate dramatic changes in the structure of the Brazilian economy. Even our strong liberalization experiment would have only a moderate effect on aggregate economic activity. The simulated effects on poverty and income distribution, although not negligible, do not seem to be extreme. This highlights two important aspects of this issue, one related to the structure of the Brazilian economy, and other to an aspect of poverty.

In terms of the Brazilian economy, it was shown that it is not very oriented towards external trade. The domestic market is far bigger and more important for the general economy than the external market, an aspect long understood by researchers. This makes it naturally less sensitive to tariff structure changes, as well as to changes in export demand.

But it should also be noted that approaching poverty by the household dimension, instead of by the personal dimension, and tracking the changes in the labor market from individual workers to households is an important modeling issue. In the PNAD 2001 data we used the head of the family income accounts for about 65% of aggregated household income in Brazil. Taking the head of the household as the observation unit may exacerbate the effect of policy changes. To our best knowledge, this is maybe the first methodological approach that tracks employment by sector and region to household income via the incomes of individual family members. If spending (and welfare) is in any sense a household phenomenon, this is the appropriated method for doing so. Even though there may be a somewhat higher computational cost associated with this procedure, it seems worthwhile.

This research can be extended in a series of new directions. Maybe one of the more obvious would be to try to assess in a more direct way the importance of agricultural trade liberalization for poverty in Brazil. As we saw, the agricultural sector is one of the more important sectors in absorbing unskilled workers. Considering that agriculture is still one of the main sticking points in economic integration negotiations, this would be a natural extension for this analysis.

And, finally, it's worth noting that our model does not assess dynamic effects, or effects upon productivity gains, usually thought to be important trade liberalization effects. We have in this paper assessed a more short-run effect, highlighting compositional and regional structure differences.



Barros, R. P., Corseuil, C. H., & Cury, S. (2001a). Salário mínimo e pobreza no brasil: Estimativas que consideram efeitos de equilíbrio geral. Texto para Discussão 779, IPEA, Rio de Janeiro.        [ Links ]

Barros, R. P., Henriques, R., & Mendonça, R. (2001b). A estabilidade inaceitável: Desigualdade e pobreza no brasil. Texto para Discussão 800, IPEA, Rio de Janeiro.        [ Links ]

Barros, R. P. & Mendonça, R. (1997). O impacto do crescimento econômico e de reduções no grau de desigualdade sobre a pobreza. Texto para Discussão 528, IPEA, Rio de Janeiro.        [ Links ]

Ferreira Filho, J. (2003). The free trade area of americas and the regional development in brazil. 6th Annual Conference on Global Economic Analysis, Scheveningen, Holland.        [ Links ]

Foster, J., Greer, J., & Thorbecke, E. (1984). A class of decomposable poverty measures. Econometrica, 52(3):761-66. available at        [ Links ]

Green, F., Dickerson, A., & Saba Arbache, J. (2001). A picture of wage inequality and the allocation of labor through a period of trade liberalization: The case of brazil. World Development, 29(11):1923-1939. available at        [ Links ]

Horridge, J. M. & Ferreira Filho, J. (2003). Linking gtap to national models: Some highlights and a practical approach. 6th Annual Conference on Global Economic Analysis, Scheveningen, Holland.        [ Links ]

Horridge, M. (2000). Oranig: A general equilibrium model of the australian economy. Centre of Policy Studies/IMPACT Centre Working Papers op-93, Monash University, Centre of Policy Studies/IMPACT Centre. available at        [ Links ]

IBGE (1996a). Instituto brasileiro de geografia e estatÍstica. censo agropecuário do brasil. 366p.        [ Links ]

IBGE (1996b). Instituto brasileiro de geografia e estatÍstica. pesquisa de orçamentos familiares.        [ Links ]

IBGE (2001). Instituto brasileiro de geografia e estatÍstica. pesquisa nacional por amostra de domicílios. microdados.        [ Links ]

Rocha, S. (1998). Desigualdade regional e pobreza no brasil: a evolução - 1985/95. Texto para Discussão 567, IPEA, Rio de Janeiro.        [ Links ]

Savard, L. (2003). Poverty and income distribution in a cge-household sequential model. Technical report, International Development Research Centre - IDRC.        [ Links ]



* This research was funded by FAPESP (Fundação de Amparo a Pesquisa do Estado de São Paulo) and CNPq (Conselho Nacional de Desenvolvimento Científico e Tecnológico). The authors thank Alexandre Nunes de Almeida for his invaluable assistance on data preparation stage.
1 One of the activities (Agriculture) produces 11 commodities.
2 The methodology is described in more detail in the Appendix. Here we present only the main ideas.
3 Mark Horridge developed this method for this project.
4 For the simulation in this paper, only 1 loop was needed to converge, since the changes in demands were small.
5 POF[1] ranges from 0 to 2 minimum wages, POF[2] from 2+ to 3, POF[3] from 3+ to 5, POF[4] from 5+ to 6, POF[5] from 6-8, POF[6] from 8-10, POF[7] from 10-15, POF[8] from 15-20, POF[9] from 20-30, and POF[10] above 30 minimum wages. The minimum wage in Brazil in 2001 was around US$76.
6 The equivalent household concept measures the subsistence needs of a household by attributing weights to its members: 1 to the head, 0.75 to the other adults, and 0.5 to the children (eg, to feed 2 persons does not cost double).
7 This poverty line is equivalent to US$ 48.00 in 2001.
8 Barros et al. (2001b), working with a poverty line that takes into account nutritional needs, find that 34% of the Brazilian households were poor in 1999.
9 The proportion of households below the poverty line in the other income groups are 0.284% for the 4th, 0.14% for the 5th, 0.04% for the 6th, 0.008% for the 7th, and 0.001% for the 8th. There are no households below the poverty line for the two highest income classes.
10 The poverty gap and poverty line values are constructed with "adult equivalent" per capita household income.
11 The GTAP version 5.0 database was used for this run.
12 This effect was discussed in more detail in Ferreira Filho (2003).



A. The method of quantum weights for jobs relocation in the micro-simulation

Micro-simulation data is naturally discrete: some families have one child, some have 2 but none has 1.5 children. If the micro data survey contains 5000 workers in some occupation for which demand falls by 3%, then 150 must be fired. But, which 150? Alternatively suppose demand rose by 3%, creating 150 jobs. Which 150 of the 8000 unemployed in the microdata will get these jobs?

Several approaches have been suggested to this problem. For example, Savard (2003) constructs separate queues of employed and unemployed. The most hirable of the unemployed are the first to get jobs, whilst the least productive workers are fired first. Or, hiring and firing could be allocated randomly.

We pursue a different approach altogether, motivated by the following considerations:

  • Our CGE model and microdata identify, in effect, 11070 separate firing problems (10 occupations, 27 regions, 41 PNAD sectors) since workers in each family are tagged with these attributes; and 270 hiring problems (since unemployed have no sector). It might be computationally expensive to construct 11340 separate queues.

  • Perhaps 5000 of the 11070 different percent changes in employment will be negative. For example, employment by occupation 7 in region 3, sector 18 may fall by 5%. Perhaps in the survey data there are only 17 such workers. How do we choose 0.85 ( = 17*0.05) workers to fire?

  • It is typical of CGE simulations that many changes, including many employment changes, are quite small: a subsidy to wheat might cause employment in the plastics sector to fall by 0.006%. This exacerbates the previous problem: we may have to allocate many small changes in employment, which correspond to sub-unit changes in the microdata. Rounding to the nearest worker might bias results: we might include the larger employment rises in wheat whilst overlooking the small falls in other sectors. To avoid this we need a procedure for allocating 0.07 jobs in a particular sector and occupation.

  • In our PNAD microdata, each observation has a weight, ranging from 150 to 850. We have to take these weights into account when computing totals. It will make a difference whether 1 new job is allocated to a household with weight 200 or with weight 600. This complicates the problem of distributing a discrete number of jobs.

Our procedure makes use of the survey weights to account for non-integer changes in employment, so avoiding the problems just listed.

Quantum mechanics teaches that a particle does not have just one location and speed at a certain moment, but is better imagined as a 'probability cloud' showing the likelihood that the particle is in a certain position. Our adoption of the name reflects a feature of our job allocation process described below: instead of trying to decide whether or not a particular worker is fired, we modify our dataset to reflect both possibilities.

Suppose that our survey data file (after it is updated by the CGE model) shows a household, with weight 200, containing only 1 worker and 3 children. We might represent this record as follows in Table 13:



Above, the first row represents household attributes, with an additional row for each adult and his/her attributes. We can see from the JobScore field that employment for workers of this type (Occupation,Sector,Region) has fallen by 5% (originally all JobScores were 1.0). In other words this worker is only working 95% of a normal job. We can restore the JobScore to an integer value by splitting the household into two records, as can be seen in Table 14.



Notice that the weights for the 2 new households sum to the original 200. The first household, with weight 190 ( = 95%*200), is otherwise identical to the original. The adult in the second household (weight 10 = 5%*200) is unemployed, and has no sector or wage. Although the second household has no income, we still label it as POF group 3; the POF group labels refer to initial household income group, and are not updated. Our programs are already equipped to deal with differing household weights (the PNAD requires this) so the only inconvenience of the split is that the number of records is increased.

Now suppose our household had two adults, both working in a sector/occupation/region that was declining (JobScore<1), one by 5% and the other by 10%. To account for Adult 1, 5% of the original record must be split off to create a record where Adult 1 has no job. To account for Adult 2, 10% of the original record must be split off to create a record where Adult 2 has no job. So we get 3 households: one where both adults are employed, one where adult 1 loses the job, and one where adult 2 becomes unemployed. This example can be seen in Table 15.



Notice that, taking the weights into account, the splitting preserves both the total employment and total earnings of the original record. However, the variance of families incomes is increased by the split. We could have created a 4th household where both adults lost their jobs - with weight of 1 ( = 5%*10%*200) but most of the employment changes were too small to justify this step.

In general, we need to create a new household for each working adult with JobScore>1 and for each unemployed adult with an occupation in increasing demand. Since most households have either one or two adults in the labor force, and about half of the occ/sector/region labor demands fall, we need to approximately double the number of households. If we took into account unlucky cases such as the 4th household just mentioned the multiplication of household records could be more severe.

So far we have only examined cases where employment shrank. Let's now see the case where employment expands, say, by 5% in some sector. We would merely truncate the JobScore to convert this 1, as can be seen Table 16.



No new record is created this time. The lost labor time (0.05*200) and lost wages (0.05*200*250) must be preserved (labeled by region and occupation) for later distribution to the unemployed. Once we have processed all adults in a region we know how much labor and wages of each type must be distributed to unemployed. We also know how many unemployed there are of each type (recall, unemployed were assigned to an occupational group). We then pass through the records again, seeking to share out the jobs amongst the unemployed.

Suppose we come upon a record like the one showed in Table 17. This adult represents 150 unemployed of occupation 3 in Sao Paulo. Suppose in total there were 30000 of such adults, so this adult is 0.5% of the total. If there are 20 jobs to distribute, the group represented by this adult should get 0.1 jobs. Therefore we split the record in proportions 149.9/0.1 to get two records, one where the worker is unemployed and the other where the worker is employed, preserving the weight total.



The wage can be worked out since we know how much income we took from over-worked persons of this occupation and region (principle of income conservation). This implies that new workers are assigned an average of the wages paid to this occupation in expanding industries. With wage given, the sector to which the worker is assigned does not affect income or poverty measures, so need not be known. In fact, we do assign sectors to the newly employed, using a random assignment from expanding sectors, with probabilities weighted according to size of sectoral employment increases for the relevant occupation and region.

We used a Pascal program to perform the above procedure. We note two potential problems:

  • the number of new jobs created for a particular region and occupation might exceed the number of unemployed of that type. Potentially the demand for new workers (from the CGE model) might exceed the supply (in the microdata). The problem occurred very rarely in our simulations, mainly for higher-paid occupations in a few regions: recorded unemployment tends to be low amongst these groups. Since our focus was mainly on lower-paid workers, we were not very concerned. In Brazil there is no shortage of less-skilled labor. Our solution to the problem was to first mop up the unavailable unemployed, then to force workers in the bottleneck occupations to work a little harder (i.e., we allowed a few JobScore values to remain above 1).

  • the second problem is subtle and rare (it occurred in 6 out of the 112055 original households). Suppose, for a particular region and occupation, that 2/3 of the unemployed are to get jobs. Suppose we have a household weight 300 with two such unemployed. According to the scheme outlined above we would create 2 new household records. The first, with weight 200 ( = 300 * 2/3) would allocate a job to Adult 1. The second new record, also with weight 200 would show Adult 2 as employed. Since the sum of weights must not change, the weight now assigned to the original household must be -100! Our solution was to assign a zero weight to the original household and weights of 100 to the 2 new households - meaning that a few unemployed were denied the chance to work. Another solution, mentioned previously, would be to create a third new household in which both adults would get jobs.

Our job allocation procedure does not alter numbers employed or wages earned: it only redistributes jobs and income between adults of the same occupation and region. The effect on income distribution within such a group can be large, but the potential for disagreement with the CGE model results is small, as long as the job redistribution within occupations does not move income between the POF income groups which drive consumption. In practice there is a strong correlation between occupational groups (based on individual earnings) and POF income groups (based on household earnings). Hence, job redistribution within occupations affects income distribution within, more than between, POF groups.

Creative Commons License All the contents of this journal, except where otherwise noted, is licensed under a Creative Commons Attribution License