Structural Change in Brazil Reconsidered:

(1)

Structural Change in Brazil Reconsidered:

Accounting for regional differences in labor productivities

.

Tim Aalvink

2036932

timaalvink@gmail.com

University of Groningen

Faculty of Economics and Business

(2)

Abstract

This paper studies structural change and labor productivity growth in Brazil in the period 1991 – 2010. Conducting a shift-share analysis it seeks to set out the contribution of structural change in Brazil’s labor productivity change. First, the analysis is conducted on a ten sector level of the economic structure of Brazil. After that the analysis is replicated while accounting for differences in labor productivities across the 27 federative units of Brazil. The results suggest that structural change has contributed a small share to increase labor productivity. When accounting for differences across federative units the results do not significantly change.

(3)

In the 18th and 19th century many of today’s advanced countries started to industrialize and experience faster economic growth. Along economic growth industrial revolution affected everyday life and improved the living standards of the population to above subsistence level. For development economists the likely existence of the interrelatedness between industrialization and economic development provided the basis for the construction of models of structural change in the mid twentieth century. The models tried to explain the mechanism behind changes in the economic structure and specify the importance of structural change for economic growth. Building on these models research has put effort in uncovering the contribution of structural change in changes in national productivity using the shift - share analysis introduced by Fabricant (1942). It provides an estimation of the amount that structural change contributed to productivity growth, after decomposing productivity growth into structural change and within sector productivity gains.

However, the contribution of structural change in national productivity growth can be disputed depending on the level of detail the shift-share analysis is applied to. A country’s economic structure can be defined in different amounts of sectors depending on the level of disaggregation a scholar uses. This paper builds on the discussion between McMillan and Rodrik (2011) and de Vries et al. (2012) on the role of structural change after applying the shift-share analysis on different detailed level of an economy. McMillan and Rodrik (2011) applied the shift-share analysis on a nine sector disaggregation and find that structural change was negative for countries in Africa and Latin America. In turn, de Vries et al. (2012) point out that applying the shift-share analysis on three sectors or 35 sectors can change the contribution of structural change in productivity growth. Additionally, they argue that, when taking into account informal activities in sectors, movement of workers from these activities also contributes. They find that disaggregating between formal and informal activities overturned the contribution of structural change in Brazil.

This paper will add to this discussion by introducing another possible way of disaggregating sectors. Instead of considering a country in terms of sectors, it also consists of regions. In a large country such as Brazil, different regions can have very different economic structures. For instance, some regions will rely more on mining since they have large natural resource endowment. Each one of these regions in turn has its own economic structure with their own labor productivity levels. And these labor productivity levels will most likely differ from other regions. In this case, making a distinction between sectors in different regions might also result in a different contribution of structural change in the shift-share analysis.

(6)

will split up the ten sectors into a total of 250 different sectors for Brazil. I suspect that accounting for different labor productivities across regions will increase the contribution of structural change. Therefore, I compare both analyses, before and after disaggregating, to see if structural change increases.

For the period 1991 – 2010 I find that, when only considering Brazil consisting of ten sectors, structural change contributed to 0.22% of the changes in labor productivity. At the same time labor productivity within a sector contributed to 5.36% of changes in labor productivity of Brazil. After disaggregating the nation into 25 separate regions the contribution of structural change becomes 0.25% and that of within sector labor productivity becomes 5.33%. Thus it seems that, in the case of Brazil in 1990-2010, Accounting for differences across regions does not significantly change the contribution of structural change.

The small difference of the structural change terms in the first and second analysis suggests that in this case the interpretation of the role of structural change stays the same. Whether disaggregating a country into regions also doesn’t change the results of a shift-share analysis for other countries or periods remains unknown. However, Brazil is one of the largest countries in the world and splitting its economic structure op in smaller regional structures, would seem to make more sense than for a small country. In any case, taking into account that regions have different labor productivity levels in the same sectors provides a more detailed picture of the economy. And instead of relying on aggregated national data, it might be important to take this into account.

Section 2 will discuss Lewis two sector model and provide a snapshot of findings in the literature on structural change. After that it discusses the findings by McMillan and Rodrik (2011), and de Vries et al. (2012) and introduces the hypothesis. In section 3 the methodology is set out and discusses some limitations, it furthermore provides a numerical example on how accounting for differences across regions can change the result. . Section 4 discussed the collection of data and the construction of the variables. Section 5 the results from the regressions and shift share analysis will be discussed. Finally, in the section 6 the paper will conclude and set out a discussion for future research.

2. Literature review

2.1 Definition of structural change

(7)

other economic processes which revolve around structural change such as the rise in urbanization, the demographic transition and the income distribution. This highlights the complexity and it is, therefore, important to state how structural change is defined in this paper. At the risk of oversimplifying I define structural change as “the change in the relative importance of sectors in an economy in terms of employment”. This definition is useful, since it emphasizes the importance of sectors, measured by their relative weight. The essence of this thesis lies around models of structural change and it is therefore necessary to set out the particular kind of model that is used in the argumentation. The next sub-section will set out the Lewis model with unlimited supply of labor.

2.2 Lewis two sector model with unlimited supply of labor

Lewis (1954) sets out a dual economy model consisting of agriculture and industry in explaining the elements of income growth and industrial employment. He assumes that there is a fundamental asymmetry between the two sectors’ use of factors. In his model agriculture uses only the production factors labor and land and industry uses the production factors labor and capital. Another important assumption is that there is a surplus of labor in agriculture. Lewis argues that this surplus is realistic for developing countries in Asia where, compared to land and capital, the amount of inhabitants is relatively large and most people are employed in agriculture. This means that too many people are working in agriculture and the same amount of production can be obtained with fewer laborers plotting the land. If this is the case the marginal productivity in that sector is zero or may even be negative. This overpopulation of agriculture allows for the economic development of other sectors without a labor constraint. In Lewis’ dual economy the other sector is the industry. The industrial sector uses capital in production, but in the early stages of development the industry is still small and capital is scarce. As mentioned above the second factor used in industry is labor. A manager decides how many laborers he employs in order to maximize his profits. He therefore employs as many laborers until their marginal productivity of labor equals their wage. The marginal productivity of labor depends on the ratio between capital and labor in the industry. If more capital is accrued by the industrial sector, more labor can be hired in the industry sector. However, people will only move to the industrial sector when the wage in the industrial sector is higher than, or equal to, their nominal wage in agriculture plus the cost of migration. This is because a farmer would only want to work in the industrial sector when his wage is covering the cost of changing his own life. To simplify his model, wages are fixed at a certain level above the wages of agriculture.

(8)

of labor in the industry sector. An increase of capital in the industry pushes out the marginal productivity of labor, as is shown by the concave lines (N1-Q1;..; N4-Q4), and more labor is hired until it equals the wage again.

Figure 1

Figure 1. Source; replicated after Lewis, A. 1954. Economic development with unlimited supplies of labour. It depicts the wages in agriculture (s) and industry (w), the concave lines depict the marginal productivity of labor. It shows that when the marginal productivity of labor increases, due to capital accruement (an outward shift of the concave lines) in the industry, the manager will hire more labor until the marginal productivity in labor equals wage again.

From the figure it becomes clear that in the Lewis model the accumulation of capital does not result in an increase of wages. Only when the surplus of labor has disappeared, will the wages increase. However, from the start, it will result in the rise of national income as the share of workers in the industry sector gets larger and profits increase for the capital owners. The model shows how movement of labor across sectors can increase productivity and income of a country. In this model output in agriculture stays the same and that of industry rises.

2.3 Empirical findings on structural change

The theory sets out a situation in which structural change improves national income. Structural change has been subject to empirical analyses, at first with regard to now developed countries and later for developing countries. I will make a distinction between advanced and developing countries, in order to address differences in findings. For early developed countries there exists a certain pattern in which the structure of an economy changes with different levels of income. Typically, industrial sectors grow more rapidly than agriculture, increasing the share of the industry (Chenery, 1982).

(9)

of GDP would rise alongside economic development at the cost of the share of agriculture. Further along the development path the rise of GDP will increase the share of services in GDP, this time at the cost of the share of agriculture and industry. Evidence of this pattern of economic structure is provided by several empirical articles, for instance the paper of Chenery and Taylor (1968).

Chenery and Taylor (1968) conduct two regressions on their country sample, first they conduct a multiple regression on the total sample of countries and then they divide the dataset in three different types of countries. In their multiple regression analysis on all countries they find that industry and services shares, in terms of GNP per capita, increase when countries experience growth in GNP per capita while at the same time agricultural shares decrease. They also expected that different types of countries experience different speed in structural change along the development path. Therefore, they divide their sample in three types of countries: 1. Large countries, 2. Small primary resource dependent countries and 3. Small industry oriented countries. When subdividing their dataset into these three types of countries they find stronger results than their pooled regression. For all countries the share of agriculture in GNP per capita still decreased when GNP per capita increased. However, for ‘small countries that mainly depend on their primary resource base’ this decline is slower than for large and small industry oriented countries. They furthermore find that industry shares peak around 37 percent for large countries, after which the decline again with increasing GNP per capita. The large countries and small industry oriented countries experienced more similar patterns of structural change along a rise of GNP per capita, the only difference being that small countries’ effect of capital accumulation becomes smaller. This the authors attribute to the economies of scale that a large country can experience and a small one cannot

(10)

Figure 2

Figure 2. Source: Bah, (2011) Structural Transformation paths across countries. The picture depicts a scatter plot of nine developed countries (Australia, Canada, France, Germany, Italy, Japan, Sweden, Uk, and US) with the polynomial fitted curve and the two standard deviation upper and lower bounds. It shows the relation of each sector share of output at different levels of the natural log of GDP per capita.

Bah (2011) replicates his analysis for 38 developing and emerging countries. He argues that there is not the same pattern of structural change for all developing countries. Instead there are great differences in the path followed. Along different levels of GDP per capita he finds that the model, that fit the developed countries, does not fit those of developing countries. He finds that for the sectors agriculture, industry, and services the R-squared of the respective polynomial function is 0.35, 0.27, and 0.08 respectively. This, he argues, shows that developing countries do not experience the same trend in structural change along the development path as developed countries.

Figure 3

Figure 3. Source: Bah, (2011)

Structural Transformation paths across countries. The figure

(11)

To illustrate this Bah (2011) picked scatter plots of five countries along the services curve for developed countries, shown in figure 3. In the figure only the “stars” of Korea show a very strong correlation with the fitted curve of the developed countries in his sample. Senegal and Ghana experience hardly any pattern, where increases in services shares has not been accompanied by growth in GDP per capita. Pakistan’s “plusses” show a similar trend as developed countries, except that the share of services is much higher, where it is concentrated on the upper bound of a 2 standard deviation bandwidth. Brazil experienced a similar trend as developed countries up until a log of GDP per capita of 8.5 after which services shares increased but GDP per capita did not.

However, there might be some similarities accruable to the region of the country. That is, countries from similar region or continent have some similar paths that they follow. Grouping the countries according to their continent allowed Bah to compare the best fitted curve for different continents and to see a degree of homogeneity within a continent. His results are threefold. First, for Africa there is not one fitted curve that represent a general experience of African countries. Second, only for Asia the R-squared is high enough to suggest that countries follow a similar path. In Asia the agricultural share starts much lower than in developed countries, although this converges with rising incomes. Industry share today is still increasing and much larger than the shares in developed countries have ever been. The pattern of the service sector in Asia is quite close to that of the developed countries. Finally, for Latin America a low R-squared is observed for industry and services, this low R-squared implies that there is a lot of variation between country experiences. However, he suggests that, regardless of the low R-squared, the Latin American case depicts an interesting finding. Industry shares peeked at much lower levels of GDP and at much lower levels of the industry share. This phenomenon is coined premature deindustrialization by Dasgupta & Singh (2006). Moreover, services’ shares took off at an earlier stage of the development process and obtained larger shares in Latin America than in advanced countries at the same level of GDP. In the next section I will discuss premature deindustrialization.

2.4 Premature deindustrialization

The focus of the thesis will be on structural change in Brazil. Findings on premature deindustrialization were particular strong for Latin America, therefore I will discuss this phenomenon. The decline was particularly strong for employment compared to value added, which brings us to the main discussion of this paper. The shift - share analysis will uncover these contributions to productivity gains, this will be discussed in the next section.

(12)

could not have been a result of a global decline in the share of manufacturing since this did not decline since the 1970’s. So deindustrialization in many developing countries is not due to a global decline in the industry employment share (Felipe et al., 2015). The fact that global industry employment shares are not on the decline, suggest that some countries experienced decreases in employment shares but other experienced a rise in employment shares globally. A region that comes to mind, and what Bah (2011) also find, is (South) East Asia where employment shares are still increasing.

The economic structure of an economy can be measured in terms of input (employment) or output (value added). Rodrik (2016) finds that deindustrialization has been stronger in terms of employment than in terms of value added. Rodrik’s use of value added in his paper represents the output of manufacturing, which is the total output of production minus the use of intermediate goods. Employment is the share of active population working in the particular sector. The connection between value added in an industry and employment is then in terms of labor productivity. A reason why manufacturing experience faster declines in employment than value added is because labor productivity within the manufacturing sector has been growing at the same time. In order to uncover the contributions of structural change and within sector productivity change, researchers often apply a shift-share analysis. The next section discusses the findings of McMillan and Rodrik (2011) and de Vries et al. (2012) and the implication of applying the analysis on a different detailed level of the economic structure. 2.5 Between and within productivity change

McMillan and Rodrik (2011) conduct a decomposition method to split up changes in average national productivity levels into two contributions. They make the distinction between “within” productivity growth and “between” productivity growth. The “within effect” is the aggregate productivity growth (decline) due to increases (decreases) in the labor productivity in a industry, without any movement in labor across sectors. For instance, productivity can grow within industries through capital accumulation or technological change, in this situation firms improve their production technologies becoming more efficient producers. Another possibility is that a reduction of the misallocation across plants can improve productivity. In this case, the more effective firms increase their share of workers in the sector and the least effective firms shrink, while employment in the total sector stays the same. The “between effect” is the structural change component in their paper, measured in terms of employment1

. When employment shares of the more (less) labor productive industries increases, structural change is growth enhancing (reducing).

What they find is that many countries saw growth reducing structural change. That is, many countries experienced a decline of employment shares of relatively high productive sectors and an increase in shares in relatively low productive sectors. This results in a reduction of average labor productivity of a country. For Latin America and Africa, they find that the structural change component has been negative. At the same time the within effect attributed

1

(13)

positively to labor productivity in those countries. On average sectors increased their labor productivity. This effect can be linked to each other if sectors increased their labor productivity without increasing their output with at least the same factor. What happens is that less labor is needed, resulting in sectors shedding their excess labor. If this labor surplus is absorbed into less productive sectors this might decrease the average labor productivity of a country. Thus even though sectors improve their productivity on average, country wide productivity growth may be distorted by growth reducing structural change (McMillan and Rodrik, 2011).

However, whether or not structural change has actually been hampering economic growth in Latin America and Brazil can be disputed. For one thing shift-share analysis itself has been subject to debate, but also the manner in which it is applied. The shift-share analysis which decomposes the labor productivity growth in a between effect or within effect can be applied to a different level of sectors. The level of disaggregation that is applied to the analysis probably depends on the provision of data and the author’s preference. However, this choice of the application might change the conclusion of the analysis.

While McMillan and Rodrik (2011) split up the structure of the economy in nine sectors one could chose to define the economy by three sectors, 35 sectors or any other composition. In their article de Vries et al. (2012) point out that when choosing different disaggregation of sectors the contributions to the within and between effects can change. To show this they conduct the same shift-share analysis for estimating the between and within effect, on a three sector and a 35 sector economic structure in the same period. For instance, for Brazil they find that at the three sector level, the effect of structural change contributed to 0.6% of growth in 1995 – 2008 and, similarly, the within effect contributed to 0.6% of growth in the same period. However, if one would disaggregate the economy in 35 sectors, the results show that structural change contributed for only 0.1% of aggregate labor productivity growth and the within effect contributed for 1%. To clarify, one could consider a country in terms of three sectors, any employment movement of sectors within for example industry, would be assigned to the within contribution. Both manufacturing and construction are sub-sectors of the sector industry. Employment can move from construction to manufacturing, but it would show up as a contribution to labor productivity gains within industry rather than structural change. Applying the analysis to a different level of aggregation would then attribute part of this within effect to structural change.

(14)

productivity, however, separating activities might turn these results. What de Vries et al. (2012) find is that when accounting for informal sectors the structural change effect in 2000-2008 actually is 1.24% instead of 0.17%, and the within effect becomes -0.24% instead of 0.83%. They argue that this results from the fact that firms formalize, which causes a shift of employment to the more productive formal sectors. At the same time the firms that formalize are the most efficient in the informal sector and often less efficient than the existing formal firms. This causes the average labor productivity of both sectors to decrease, resulting in a smaller within effect.

I suggest that disaggregating the economic structure of a country gives a more precise representation of the actual economy. In any case a better disaggregation would then represent a more realistic situation of contributions in growth. Therefore, I will extend this debate and suggest that, when taking into account different economic structures across regions in a country, between and within effect can change when accounting for labor productivity differences across regions.

This paper will focus on Brazil in light of the between and within effect debate discussed above. Brazil is an emerging country with a large geographical size. In such a country it is likely that labor productivities are different across regions. I suggest it is good to take this in the consideration of a shift – share analysis. Its experience with structural change has been peculiar as shown by the literature above. For these reasons, and the availability of data I chose to focus on Brazil in the period 1990 – 2015. The large coverage of data on value added and employment between different federative units in Brazil allows for a disaggregation of structural change across these units. This will provide us with a more in depth analysis of the contributions of structural change to labor productivity increases in Brazil. In the section methodology I will discuss how splitting Brazil up in federative units might change the contribution of the within and between effect. Furthermore, since it is one of the first papers to do so, I will also discuss some of the observations on structural change within and between regions. Below I will shortly state how the economic structure of Brazil evolved over the period of interest. After which I will set out a line of argumentation how splitting up Brazil into federative units might affect the contributions of between effects to labor productivity growth.

2.6 The economic structure of Brazil and the hypotheses

(15)

These policies and the magnitude of economic growth had some effect on the economic structure of Brazil. Table 1 shows the shares of the three main sectors in Brazil for the period 1995 – 2009.

Table 1

Brazil Employment shares Value added shares

Sector 1995 2009 1995 2009

Agriculture 26,0% 17,4% 5,0% 5,6% Industry 19,7% 20,5% 30,7% 26,9% Services 54,3% 62,1% 64,3% 67,5%

Table 1. Source: Aldrighi and Colistete, (2013) Industrial Growth and Structural Change: Brazil in a Long-Run

Perspective. The table shows the employment shares and value added shares of the three main sectors in 1995

and 2009 for Brazil. Services sector has been the largest share in both employment and value added, and has seen increases in its share over this time period. Agricultural shares has been declining for employment but increasing for value added. And industry shares experienced a small increase in shares of employment, but decreases in the share of value added.

In employment terms for the period 1995 – 2009 agriculture declined its share from 26% to 17.4%. Industry remained relatively stable and even slightly increased its employment share from 19.7% towards 20.5%. Services increased its employment share from 54.3% towards 62.1%. On the output side GDP shares increased in agriculture from 5% to 5.6%, for industry the GDP shares dropped from 30.7% to 26.9% of total GDP, and that of services increased from 64.3% towards 67.5% of GDP. Given that employment shares of agriculture declined and GDP shares increased, the labor productivity of that sector probably grew faster than the labor productivity of industry and services. It is interesting to note that they find that Brazil did experience deindustrialization in terms of GDP but not in terms of employment.

Brazil consists of 26 federative units plus 1 federal district in five ‘major regions’. Disaggregating data into these federative units would provide us with an interesting insight on how the economic structures differ. 2 Regional differences in income and growth have been a characteristic of Brazil’s economy since colonial times, while at the same time economic structures differ. Furthermore, the industrialization process has been stronger in regions that are now richer. The north east, for instance, has not only a much smaller share of industry in relation to its population share it also has a GDP per capita much smaller than the south east (Baer, 2008). Thus some regions have different economic structures and this in turn has had an effect on development process for different regions. I will first analyze some experiences for the regions in Brazil, with two hypotheses introduced below.

The literature review discussed that developed countries usually went through a similar stage of structural change, and they end up with most employment and value added in services. However, emerging and developing countries did not experience the same paths and still have

2

(16)

very different economic structures. I will discuss these findings in light of regional development. Do the federative units with higher GDP per capita have similar findings as developed countries? Or are economic structures not different across regions? A common finding is that the least developed countries still have a large share in agriculture, whereas in developed countries most people are employed in services. To test whether these results can be generalized to regions my first two hypotheses will be:

Hypothesis 1: Low GDP per capita regions have a larger share of employment in agriculture than high GDP per capita regions.

Hypothesis 2: High GDP per capita regions have a larger share of employment in services than low GDP per capita regions.

Another common finding is that developed countries have smaller labor productivity differences across sectors (McMillan and Rodrik, 2011). In developing countries sectors often show very different labor productivity levels whereas in developed countries these are converging as employment moves to the more productive sector. Again I will discuss this finding in the light of regional differences in labor productivity. To test whether richer regions also have smaller productivity level differences than the poorer regions, the third hypothesis will be:

Hypothesis 3: High GDP per capita regions have smaller labor productivity differences between sectors than Low GDP per capita regions.

The disaggregation in regions provides us with a possibility to examine how the between and within contributions to productivity growth can change. Because, but not solely, the economic structure is likely to be different from region to region I expect labor productivity levels also to be different for each region. For instance, agriculture might have different labor productivities in Maranhão than in Distrito Federal. When employers in agriculture move from Maranhão to Distrito Federal, and keep working in agriculture, it could change the average labor productivity in agriculture of Brazil. However, when looking at it through a national lens it would seem like labor productivities in agriculture changed without a movement of labor. Therefore, when making a regional distinction the contributions of the between and within effect might change. I present a numerical example how this might be the case in the section methodology.

(17)

Theory proposes several reasons why firms chose to locate near each other, from the earlier work on scale economies to the knowledge spillovers from growth theory. First, scale economies have been the economic rationale for the existence of cities. Without economies of scale all production would be dispersed to save on transportation costs, this is also called the Starretts theorem (Quigley, 1998). Second, the shared inputs in production, implies that the existence of specialized workers (such as accounting, or law) in certain locations reduce the costs for firms (Krugman, 1993; Quigley,1998). Thirdly, works of Acemoglu (1996) suggest that when the stock of human and physical capital in a city increases the returns to human and physical capital increase as well. Fourth, real savings can be made on the fact that the existence of a large pool of workers and firms can control for economic fluctuations. That is employment can be stabilized even when the hiring of workers is imperfectly correlated across firms, because constantly some firms are hiring and others are not (Mills and Hamilton, 1984). Finally, growth theory suggests that firms benefit from knowledge spillovers when being closely located to each other. Lucas (1988) argues that human capital spillovers in cities are born on the creation of new ideas in cities. The existence of large hubs such as the Silicon Valley provides support that technology creation benefits from this spillover (Saxenian, 1994; Glaeser, 1999). Furthermore, some empirical studies have found evidence in support of other of these benefits. Glaeser and Maré (2001) find support for the existence of an urban wage premium that reflects the learning possibilities of workers in urban cities. Rosenthal and Strange (2001) find that knowledge spillovers, input sharing and labor market pooling all positively affect the agglomeration of firms. All these benefits suggest that there exist an economic incentive to be located near each other.

In recent decades decreasing transportation costs and the improved flow of information through computers provided firms with better possibilities to choose their location. Lower transportation cost meant that firms do not have to be as close to their selling market as before, since they could move the product there at lower cost. This is true at least until some critical point, transport costs facilitate agglomeration (Brakman et al., 2001). Therefore, I argue that in Brazil firms choose to move close to each other, benefiting from the advantages mentioned above. This movement does not necessarily show up in a change of employment shares in sectors nationally, but for regions their employment’s share did change. Since agglomeration forces are driven by positive externalities and large labor markets I suggest that firms moved their production base to regions with better labor productivities. Thus structural change, in this case, contributes to the gains in labor productivities for these industries. Therefore, in my shift-share analysis after disaggregating into federative units the contribution of the between effect will be larger than before.

(18)

3. Methodology

In this section I briefly discuss the methodology for testing hypotheses 1, 2, and 3. After that I more extensively discuss the methodology for the main hypothesis of this paper, hypothesis 4. Finally, with a numerical illustration I show how the contributions of between effect and within effect to national labor productivity might change when disaggregating a country into regions.

3.1 Hypothesis 1 and 2: Industry shares and GDP per capita

The first hypothesis suggests that regions with lower levels of GDP per capita have a higher share in agriculture. The second hypothesis suggests that regions with higher levels of GDP per capita have higher shares in services. I conduct a regression analysis with the following log-log fixed effects model for each industry separately.

ln⁡(𝑥𝑖,𝑟,𝑡) = ⁡ 𝛽0,𝑟+⁡𝛽1ln 𝐺𝐷𝑃𝑟,𝑡+⁡𝛽2ln 𝑁𝑟,𝑡+ 𝜀𝑟,𝑡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡ (1) Where xi,r,t is the share of either agriculture, industry, or services in total employment of a region at time t. GDP is the gross domestic product per capita of region r at time t, and control variable N is the population of region r at time t. The expectation is that for agricultural shares β1 is negative and for service shares this is positive. There are several problems with this regression analysis which make it more difficult to discuss the results with some certainty. The regression only uses three observations per region this could cause the within-unit effects of GDP on shares of a sector in employment to be different from the true effect (Clark and Linzer, 2015). In such a situation random effects might be better. However, I will use the fixed effects model over the random effects model because the expectation is that GDP per capita of a region is also correlated with some omitted variable that influence the economic structure in turn. For instance, the federal district has a large level of GDP per capita and at the same time a high share in services partly because the federative unit contains the capital city and political center of the nation. Nonetheless, I will perform a Haussmann test to see whether the random effects is consistent when compared to the fixed effect estimator. Furthermore, another problem with this analysis is that reverse causality could exist when structural change affects GDP per capita. This might very well be the case so in discussing the results I will only argue for a relationship not for causality.

3.2 Hypothesis 3: Labor productivity differences and GDP per capita

(19)

for the average labor productivity of a region at time t. I expect the coefficient of ln(GDP)r,t and ln(lPavg)r,t to both be negative.

ln(𝑠𝑑⁡𝑙𝑝)𝑟,𝑡 = 𝛽0,𝑟⁡+⁡𝛽1ln⁡(GDP)𝑟,𝑡+⁡𝛽2ln(𝑁)𝑟,𝑡+⁡𝜀𝑟,𝑡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(2) ln(𝑠𝑑⁡𝑙𝑝)_𝑟,𝑡 = 𝛽_0,𝑟⁡+⁡𝛽₁ln⁡(LPavg)_𝑟,𝑡+⁡𝛽₂ln(𝑁)_𝑟,𝑡+⁡𝜀_𝑟,𝑡⁡⁡⁡⁡⁡⁡⁡⁡⁡(3) 3.3 Hypothesis 4: The shift – share analysis

Furthermore, the goal of this paper is to show that contributions of between and within effects change when one disaggregates the country into different regions. The hypothesis states that the structural change or between effect will become larger. In order to identify if contributions from within and between effects change I first have to conduct the analysis for the country as a whole. I will use the same shift-share methodology as McMillan and Rodrik (2011) to define the contribution of structural change to changes in labor productivity. They use the following formula to split national changes in productivity levels to (1) changes in labor productivity levels within economic sectors and (2) changes due to the movement of labor across sectors. Labor productivity levels of a country as a whole are average labor productivity and computed by dividing value added of a country by its population employed.

∆𝑌_𝑡= ⁡ ∑ 𝜃_{𝑖,𝑡−𝑘}∆𝑦_𝑖,𝑡 𝑖=𝑛

+ ∑ 𝑦_𝑖,𝑡∆𝜃_𝑖,𝑡 𝑖=𝑛

⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(4)

(20)

This formula in return, will give more weight to the productivity shares, so in order to have the smallest bias it is best to use the average of both formulas. And that is what I will do, the next formula represents this formula with ‘mid-year’ arithmetic means. Here I take the value of both periods and divide them by two, before multiplying it with the change.

∆𝑌𝑡= ⁡ ∑ ( 𝜃𝑖,𝑡−𝑘⁡+⁡𝜃𝑖,𝑡 2 ) ∆𝑦𝑖,𝑡 𝑖=𝑛 + ∑ (𝑦𝑖,𝑡−𝑘+⁡𝑦𝑖,𝑡 2 ) ∆𝜃𝑖,𝑡 𝑖=𝑛 ⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(6)

To continue, the goal of this paper is to see how between and within contributions change when disaggregating to the regional level. I should then start to have a look into the contributions at a regional level. The same shift-share analysis is applied for each region separately. So that the dependent variable is not the change in labor productivity levels for a nation, but instead the change in labor productivity levels in a region which is denoted by a subscript r (∆𝑌𝑟,𝑡). The formula for each region is displayed below by formula (7).

∆𝑌𝑟,𝑡 = ⁡ ∑ ( 𝜃𝑖,𝑡−𝑘⁡,𝑟 +⁡𝜃𝑖,𝑡,𝑟 2 ) ∆𝑦𝑖,𝑡,𝑟 𝑖=𝑛 + ∑ (𝑦𝑖,𝑡−𝑘,𝑟+⁡𝑦𝑖,𝑡,𝑟 2 ) ∆𝜃𝑖,𝑡,𝑟 𝑖=𝑛 ⁡⁡⁡⁡⁡(7)

The formula shows that the change in productivity levels for each region is similarly deconstruct in a contribution of within and between effects. In the formula (7) above the employment shares are a part of total employment of the region and labor productivity is the value added of that sector in that region divided by its respective employment level. The formula nationwide (8), below, is very similar to formula seven for the regional case. However, the employment shares of sector i in region r is now the share of total national employment and defined by λi,t,r. Productivity levels are computed in the same manner as they are regionally, therefore I use the same symbol. Moreover, an additional summation is necessary to combine all the regional contributions into 2 contributions for the nation as a whole. 𝛥𝑌_𝑡= ⁡ ∑ ∑ (λ𝑖,𝑡−𝑘⁡,𝑟+⁡λ𝑖,𝑡,𝑟 2 ) ∆𝑦𝑖,𝑡,𝑟 𝑖=𝑛 𝑟=𝑠 + ∑ ∑ (𝑦𝑖,𝑡−𝑘,𝑟 +⁡𝑦𝑖,𝑡,𝑟 2 ) ∆λ𝑖,𝑡,𝑟 𝑖=𝑛 𝑟=𝑠 ⁡⁡⁡⁡(8)

The employment shares are measured in terms of total employment of Brazil in order to account for changes in the employment levels in regions. If employment shares would be constructed at the regional level (where a region’s total employment is 100%) a problem arises when summing the regions. Since employment shares of a region in Brazil’s total employment change, defining the weight of each region for summation was difficult. Initial, end-period, or mid-period shares of employment did not return the same labor productivity changes as the national analysis from formula (4). Constructing employment shares as a share of Brazil’s total employment solved this issue.

(21)

over time the between effect for the analysis is probably underestimated. One could think of unemployment as a sector with low or no productivity levels. Then a consideration would be to add a sector to the economy with a very low average productivity level and to assign all unemployed to this sector. However, I leave this sector out of the analysis since I consider structural change in the relative weights of sectors.

Formula (8) is the main formula of this paper, and I suggest that the sum of the regional between effects is larger than the nationwide between effect. Thus the second term in formula (8) is larger than the second term in formula (4). That is hypothesis 4 stated that a larger share of the change in national labor productivity levels can be attributed to structural change when disaggregating Brazil into regional economic structures.

An important assumption of this shift – share analysis is that the productivity growth within sectors is not affected by the movement of employment. For movement of employment not to affect the average labor productivity level of a sector the marginal labor productivity should equal the average labor productivity. In other words, as long as the marginal productivity is below average productivity in a sector, the movement of employment away from this sector will raise average labor productivity (Timmer & de Vries, 2009). In the literature review I already quoted Lewis that, in developing countries in Asia, a surplus of labor in agriculture is likely. When this is the case the assumption that marginal labor productivity equals average labor productivity is violated. A suggestion would be to assume the marginal labor productivity of all workers who move away from a sector to be zero, as suggested by Broadberry (1998). However, to allow the marginal productivity of workers that move to equal zero is equally a restrictive assumption. Another possibility is to let marginal productivity of the ‘workers that move’ to range between zero and the value of average labor productivity of the sector they leave (Timmer & de Vries, 2009). However, to calculate marginal productivity as a range somewhere in between these two values I would need the shadow price of labor for each sector in each region, which I do not have. Furthermore, I want to discuss the results of my analysis in light of the findings of McMillan and Rodrik (2011) and therefore use the traditional shift-share analysis setting marginal labor productivity of a sector equal to average labor productivity.

Additionally the shift – share analysis is only a static measure as it relies on the productivity levels of sectors. It only measures the movement of workers to sectors with high (low) productivity levels in considering the effect of structural change. It doesn’t account for productivity growth of sectors. A worker can move to an initially high productive sector but the gains to labor productivity can be depressed by the fact that this sector’s productivity is decreasing at the same time. This negative correlation might exist when workers moving to the sector are employed in the low-tech or informal activities of that sector, their marginal productivity will be low depressing the growth rates (de Vries et al., 2015a)

(22)

for different sector classification. The idea is that the same effect takes place when disaggregating into regions at different considerations of the sector level. Thus the effect is expected to be the same when considering the economy consisting of three sectors.

3.4 Numerical example

Next I set out a numerical example to illustrate how disaggregating a country into regions can change the results of a shift - share analysis for a nation. This is best shown using a one country, two sectors, two regions example.

Table 2

Pala (t) Employment share Employment Labor productivity Value added

Agriculture 0.7 700 0.86 600

Manufacturing 0.3 300 3.00 900

Total 1 1000 1.500 1500

Pala (t+k) Employment share Employment Labor productivity Value added

Agriculture 0.7 700 0.88 614

Manufacturing 0.3 300 3.00 900

Total 1 1000 1.514 1514

Own calculations. Table 2 on changes in labor productivity levels in a fictional country with a dual economy, It shows the employment shares and levels for each industry and the country as a whole in both time periods. Furthermore, it depicts the value added levels, and labor productivity levels. This particular case depicts the change in productivity levels without any apparent movement in employment levels between the sectors.

In table 2 above we have a fictional country “Pala” with two sectors agriculture and manufacturing. Both sectors contribute to the nation’s wealth in terms of value added, and harbor a certain level of employment. Let’s assume that no structural change takes place so that the employment shares of each industry stay the same in time t+k as in time t. However, value added in agriculture increased from 600 to 614, which happened because labor productivity increased. Table 3 shows the contributions of each effect to national labor productivity growth using formula (4). The 0.014 productivity gain refers to the fact that the average worker’s labor productivity increased. In table 2 this can be seen from the increase in total labor productivity, which increased from 1.500 to 1.514. This contribution is all attributable to the within effect because the same share of laborers became more productive in agriculture. Table 3 National Between 0 Within 0.014 Total 0.014

(23)

Table 4

Pala (t)

Emp

share employment Lab

prod val add

Pala (t+k)

Emp

share employment Lab prod val add Agr 0.7 700 0.86 600 Agr 0.7 700 0.88 614 Man 0.3 300 3.00 900 Man 0.3 300 3.00 900 total 1 1000 1.50 1500 Total 1 1000 1.51 1514 West Emp share employment Lab

prod val add West

Emp

share employment Lab prod val add Agr 0.35 350 0.79 275 Agr 0.25 250 0.79 196 Man 0.15 150 3.00 450 Man 0.25 250 3.00 750 total 0.5 500 1.45 725 Total 0.5 500 1.89 946 East Emp share employment Lab

prod val add East

Emp

share employment Lab prod val add

Agr 0.35 350 0.93 325 Agr 0.45 450 0.93 418

Man 0.15 150 3.00 450 Man 0.05 50 3.00 150

total 0.5 500 1.55 775 total 0.5 500 1.14 568

Table 4. this table shows the situation from table 2 with the first four rows. Below that are the economic structures of the two regions at both time periods t and t+k. It shows the values for employment, value added and labor productivity for each sector in each region. It shows that employment moved across sectors in both regions, where the movement occurred in the exact opposite direction such that it doesn’t show for the country as a whole. This particular case shows that there is no within contribution to labor productivity growth, only structural change.

However, when we disaggregate the country into two regions “East” and “West”, we can show that structural change actually took place. Table 4 shows the evolution of employment, labor productivity, and value added for each region and the country as a whole.

(24)

realistic one, but it shows that contributions can change when accounting for separate regions. My expectation is that the between effect will increase when using formula (8).

Table 5

Pala 1 West East Pala 2

between 0 0.221 -0.207 0.014

within 0.014 0.00 0.00 0.00

total 0.014 0.221 -0.207 0.014

Table 5. It shows the contributions of both kinds of effects in the total productivity change. Pala 1 shows the initial situation considering the country as a whole, Pala 2 shows the situation after accounting for differences in the two region. The middle columns show the experience of both regions multiplied with their weight in employment. It shows how results of shift - share analysis can change after disaggregating.

4. Data

This section addresses the construction of a database on value added, employment, and labor productivity for each sector in each region for the period from 1991 - 2010. I will first introduce the main sources used in the collection of data. Then I will shortly discuss combining data on sectors into the ten sector classification used in this paper. After that the construction of the variables are explained, first the employment levels and value added in light of Brazil and thereafter for the federative units.

4.1 The datasources

(25)

respondent, his/her economic status, and the sector he/she is economically active in. Ipums provide 10% samples on census data picking each tenth observation after a random start. A caveat of their sample methodology is that it has the potential to significantly affect the precision of estimates. Fortunately, the industry employment variable is not identified as a variable that is largely subject to sampling error, so I do not expect this to result in any problems. Nonetheless, I consider my correction terms in constructing the variables to see whether these estimations come close to what is provided by IBGE on sectors nationally, which will be discussed later on.

Finally, IBGE also doesn’t provide data on employment levels for 1991 at the national level. Therefore, I turn to the GGDC ten sector database to obtain data on employment levels (Timmer et al., 2015). The GGDC uses data from the IBGE to collect employment levels for Brazil in 1991. However, since the IBGE doesn’t provide enough detail in the services sector the relied on additional sources such as the household surveys and annual surveys from IBGE (de Vries et al., 2015b). The GGDC provides data on employment levels in a ten sector classification for the year 1991.

4.2 Constructing the ten sector database

(26)

4.3 The nation-wide levels on employment and value added

To conduct the shift-share analysis for Brazil I obtain value added and employment levels. The IBGE provides data on value added for the years 2000 and 2010 in constant prices using a reference year of 2000. This data is freely available from the national accounts on their website. I obtained the data on value added for the 51 economic activities and aggregated them to the fifteen sectors presented by Table B in the appendix. In order to obtain data on value added in 1991 I use the table on “the evolution of value added for each sector”. This table provides the volume of value added in each of the fifteen sectors at time 2000 and 1991 in a ratio over 1985 value added volumes. Thus the volume of value added in 1985 is set at 100 and the volumes of value added in the years thereafter are divided by the volume of 1985. For instance, the ratio of value added of agriculture in Brazil in 2000 is 161.3, or, the volume of value added is 1.613 times higher in 2000 than in 1985. Since I want to present the results in constant 2000 years I take the ratio of value added in 2000 and divide the ratio of 1991 by this value. Below formula (9) depicts this process, here 1991VAq2000,i is the value added level of 1991 of a sector in constant 2000 prices. VA2000,i is the value added level of a sector in 2000, RatioVAyear,i is the ratio of value added of 1991 and 2000 respectively, where 1985 is set at 100.

1991𝑉𝐴_{𝑞2000,𝑖} = ⁡ 𝑉𝐴_2000,𝑖⁡×⁡(𝑅𝑎𝑡𝑖𝑜𝑉𝐴1991,𝑖)

(𝑅𝑎𝑡𝑖𝑜𝑉𝐴_2000,𝑖)⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(9) Doing this would set the ratio of volume of value added in 2000 at 100, and those of 1991 in terms of 2000. To give an examples the RatioVA1991 in agriculture is 115, and VA2000 is R$ 56,962 million. The factor with which I multiply the value added of agriculture in 2000 then becomes approximately 0.713, and the value added of agriculture in 1991 using constant 2000 prices is then R$ 40,611.5 million, as shown below.

𝑅$40,611.5⁡𝑚𝑙𝑛 = 𝑅$56.962⁡𝑚𝑙𝑛⁡ × 115 161.3

(27)

agriculture, manufacturing, business and mining. At the same time the 6 sectors left saw increasing in their employment shares over time. The large decrease in employment shares in agriculture provides a first indication that structural change might have been positive over time, although this will become clear from the shift-share analysis. Manufacturing, business and mining experienced a U-shaped evolution of their employment shares, where in 2000 their employment shares had the lowest levels in the three period interval. At the same time “other” services depicts an inverted U-shape pattern. The fact that manufacturing, business and mining all had larger initial labor productivities than “other” services might indicate that the structural change effect is larger for the second decade. However, this would depend on the order of sectors according to their labor productivities in 2000 and 2010, and the experience of the residual sectors.

Figure 4.

Figure 4. It depicts the ten sectors that compose the economy of Brazil in order of initial labor productivity levels. Furthermore, it depicts the employment shares of each sector in each year. It shows how the employment shares have changed for each sector over time, where the largest changes have been in agriculture. Furthermore, manufacturing, business and mining experienced a U-shaped pattern in changes of employment shares.

Figure 5 in turn depicts the scatterplot between changes in employment shares against the initial log of labor productivity ratio. The labor productivity ratio is the labor productivity of the sector over the average labor productivity of Brazil. The agricultural sector (agr) has quite low labor productivity ratios and simultaneously experienced decreasing employment shares over the following decade. Government services (gov) and construction services (con) had above average labor productivity levels, and saw increases in their employment shares. Another interesting observation is that trade services (com) increased its share of employment, while having below average labor productivity levels. Overall, most of the observations (12 out of 20) are either in the top right or bottom left corner, implying that these sectors when having above (below) average labor productivity saw increases (decreases) in the employment shares, which would suggest growth enhancing structural change. At the same time a lot of observations are in the top left or bottom right corner, which would suggest

0 0,05 0,1 0,15 0,2 0,25 0,3

Initial Labour productivity Employment share 1991 Employment share 2000 Employment share 2010

Employment shares for the years 1991, 2000, 2010 in order of their initial labor productivity levels.

(28)

growth reducing structural change. In total the fitted values show an increasing line across the log of labor productivity ratio, which might be a first indication of growth enhancing structural change in Brazil. However, the fitted values are strongly affected by the large decreases in employment in agriculture and when dropping these observations the slope of the fitted values actually turns negative

Figure 5

The scatterplot displays the change in employment shares of industries in the periods 1991 – 2000 and 2000 – 2010 against the log of the ratio of labor productivity over average labor productivity of each sector in 1991 and 2000 respectively. It furthermore depicts the fitted values that shows the increase of employment changes along the labor productivity ratio. The sector abreviations stand for: agr = agriculture, min = mining, man = manufacturing, uti = electricity, gas and water, con = construction, com = wholesale and retail trade, tra = transportation, bus = business, gov = goverment, and oth = other services.

4.4 Regional value added and employment.

The IBGE provides data on value added in the sectors of the federative units in their regional accounts for the years 1991, 2000, and 2010. These regional account tables provide data on the share of each region in a sectors value added, so that when summing the shares over the federative units it amounts to 100% for a sector. The regional accounts are fully comparable with the other regional accounts and compatible with the system of national accounts. The regional accounts are also presented according to the CNAE 2.0 classification (IBGE, 2017). Therefore, they can be used without making additional adjustments. In order to compute the value added of each sector in each region I multiplied the value added shares with the value added level in each sector nationally. The formula below shows this multiplication, where the

VAr,i,t stands for the value added level of sector i in region r at time t, the VAsharer,i,t is the

share of value added of region r in total value added of sector i at time t. and VAi,t, is the value

agr agr min min man man utiuti con con com com _tra tra bus bus gov gov oth oth -. 0 5 0 .0 5 C h a n g e i n e m p lo y m e n t s h a re s f o r 1 9 9 1 -2 0 0 0 a n d 2 0 0 0 -2 0 1 0 -2 -1 0 1 2

log of the ratio of sector lp over average lp at 1991 and 2000

change in employment shares Fitted values

(29)

added in sector i at time t for Brazil from the national accounts mentioned in the section above.

𝑉𝐴𝑟,𝑖,𝑡 = 𝑉𝐴𝑠ℎ𝑎𝑟𝑒𝑟,𝑖,𝑡× 𝑉𝐴𝑖,𝑡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(10)

However, IBGE does not provide data on the distribution of employment across sectors for each region. For data on employment I have to rely on population censuses from the Ipums international database. Ipums population census classifies respondents in nineteen possible sectors. Four of these nineteen sectors were to categorize other answers, these were: unknown, respondents suppressed, not in universe, and other industry n.e.c. (not to be mistaken with other services). When calculating the employment shares for each industry in each region these were excluded. The database provides employment estimates for three time periods, 1991, 2000, 2010. It provided around a total of 8.5 million to 10 million observations per year.

One problem that arose with the Ipums database was that the amount of respondents covered in a region could vary substantially over time. When this is the case estimating the employment shares of a sector in a region using the total amount of economically active in the sample could bias the results. This is best explained using an example. In 2000 the Ipums sample provides 755,751 observations for Rio de Janeiro which amounts to 7.46% of the total amount of observations in that year. In 2010 the number of observations in Rio de Janeiro was only 5.53% of the total amount of observations. This would not have been a problem if Rio de Janeiro also experienced an equal decline in the real population in the same period. However, the IBGE population count estimates suggest that the share of the population residing in Rio de Janeiro decreased from 8.46% to 8.22% which is a much smaller difference. Simply estimating the employment shares by dividing the observations of a sector in Rio de Janeiro by total amount of observation on employment in Brazil would bias the result. Because the Ipums sample has less respondents in Rio de Janeiro in 2010 than in 2000 this would result in negative between effects in all the sectors of Rio de Janeiro as if employment in this region just declined. Therefore, in order to control for this possible effect I calculate the employment shares of a sector in terms of total population of that region, such that I can multiply employment shares with population counts of IBGE for each region to obtain levels.

The population census from Ipums provides data on the employment status of respondent. This allows me to estimate the share of employed from the total number of observations, in order to get an estimate of employment over population of a region. With the use of IBGE population count estimates in regions I can then calculate the level of employed in each region. The formula then becomes,

𝐸𝑀𝑃̃_{𝑖,𝑟,𝑡} = ⁡#𝐸𝑚𝑝𝑖,𝑟,𝑡

(30)

POPr,t, is the population count of region r at time t provided by IBGE. This formula results in an estimate of the level of employment in sector i in region r, at time t (𝐸𝑀𝑃̃_{𝑖,𝑟,𝑡}).

The sum of employment levels over all regions for a certain sector will not add up to the national employment level of a sector, as provided by IBGE. In other words, the number of employed in agriculture nationally is not equal to the number of employed when summing agricultural employment over all regions. I use a correction term, C, for each sector to correct for this difference such that,

𝐸𝑀𝑃_{𝑖,𝑟,𝑡} = 𝐸𝑀𝑃̃_{𝑖,𝑟,𝑡}⁡× 𝐶, 𝑤ℎ𝑒𝑟𝑒⁡𝐶 = ⁡∑𝑟=𝑠𝐸𝑀𝑃̃𝑖,𝑟,𝑡

𝐸𝑀𝑃_𝑖,𝑡 .⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡⁡(12) This corrects the employment of a sector with the same factor for each region. This correction term normally didn’t deviate much from one. Only for the sector mining wat it relatively large for the years 1991, 2000, and 2010, where the correction terms was 0.68, 0.76, and 0.59 respectively. However, I do not expect this to significantly affect my results as the employment levels were relatively small, causing any bias to become larger than for other sectors. Then, using the employment levels resulting from formula (12), I calculate the shares of each sector in each region in terms of total employment in Brazil. Finally, labor productivity levels are calculated for each sector in each region by dividing the value added of that sector by its employment levels.

Furthermore, Ipums data on the residency of a respondent combines four federate states into two before providing observations, coming to a total of 25 regions. They combine Mato Grosso do Sul, and Mato Grosso together as one region, and do the same for Goías and Tocantins. Therefore, in estimating the employment levels of these states I sum the population levels of both regions and multiply it with the share calculated from Ipums. Furthermore, the value added levels are provided for the four regions separately, and they are simply summed up to present two regions.

(31)

Furthermore, several regions (highlighted with red)3 saw decreases in employment shares, most of them situated in Nordeste. However, since employment shares are estimated in total employment of Brazil it might still be that employment levels increased for these regions. Nonetheless, since they experienced a decline in their employment share, the expectation is that structural change has been less beneficial for these regions in comparison to other regions.

Table D in the appendix shows the average labor productivity levels of regions in Brazil in 1000 $real per worker per year. In the table it becomes clear that especially Distrito Federal has high labor productivity levels compared to other regions in Brazil. The relative small region has levels of around three times higher than the average labor productivity level of Brazil. In the same way as above this would imply that structural change effects are given a high weight in the analysis for Brazil as a nation overall. Moreover, São Paulo, Rio de Janeiro and Amazonas had high labor productivity levels. Given that Rio de Janeiro and especially São Paulo have such high employment shares and labor productivity levels, the between and within effects that these regions experience are important for the labor productivity changes of a nation as a whole. I will now start discussing the results from the quantitative analysis with respect to the hypotheses.

5. Results

5.1 Hypothesis 1, 2, and 3

For countries along the development path the share of agriculture declined while at the same time the share of services increased. It is interesting to find out whether these trends are generalizable to regions within a country. The first hypothesis suggested that the poorer regions have a larger share of employment in agriculture than the richer regions. Hypothesis 2 states that the richer regions have a larger share of employment in services compared to the poorer regions. In figure 6 the shares of each industry in total employment of a region are scatter along their respective log of GDP per capita at constant 2000 prices. Combining all years together I get 75 observations for each industry, I added an exponential trend line and its R-squared. The share of agriculture in total employment of a region is in general smaller for regions that are richer. For the poorest regions in the sample, the shares of services and agriculture are more or less similar, but along the log of GDP per capita the gaps between the shares get larger. Their respective trendlines have almost the same intercept, but start to diverge along the log of GDP per capita. The trend line of agricultural shares shows a decreasing slope and explains 59.5% of the variation in the data. The figure also shows a very clear pattern for services where richer regions have a larger share in services than poor regions (R2 = 0.482). Industry also increased when moving across regions with higher GDP per capita, but the changes in shares were not as strong, and the goodness of fit was low (0.2). The analysis is not that conclusive however, since each federative unit enters the plot three times affecting the results when some fixed effects are explanatory for the distribution of

3

Structural Change in Brazil Reconsidered: