Sources of Regional (non) Convergence in Mexico. Gerardo Esquivel. Miguel Messmacher

Transcription

1 Sources of Regional (non) Convergence in Mexico Gerardo Esquivel El Colegio de México and Miguel Messmacher Banco de México July 2002 Abstract In this paper we study the main sources of regional convergence in Mexico between 1960 and We show that the behavior of labor productivy was an important source of convergence in output per capa between 1960 and This trend, however, did not materialize in the 1960s and the 1980s as a result of the divergent behavior in labor market variables (employment and participation rates). In contrast wh this trend, labor productivy was the main factor behind the regional divergence pattern in output per capa that was observed during the 1990s. An analysis of the determinants of productivy suggests that the change in the regional pattern of productivy is associated to a structural change that took place during the 1990s, and that is associated to an increasing relevance of variables such as education and infrastructure in the determination of output per capa and productivy. The authors would like to thank Luz Marina Arias, Leonardo Armas, Francisco Javier Arias and Pedro José Martínez for their excellent research assistance. The opinions expressed in this document are those of the authors and do not represent the point of view of the instutions to which they are affiliated.

2 I. Introduction Differences in product per capa across Mexican states and regions have tradionally been high and fairly persistent. Product per capa in the richest state in 1950 was 6.7 times that of the lowest, and 6.1 times in In addion, there are other indicators suggesting that regional inequalies in Mexico also represent a substantial inequaly in terms of the level of welfare of the Mexican population as a whole. For example, a large proportion of the population inhabs states wh low product levels. As of 2000, 62% of the population inhabed states wh less than one third of the GDP per capa of the richest state (18 out of 32 states), and 81% of the population inhabed states wh less than half of the highest GDP per capa (27 out of 32 states). 2 The large degree of regional inequaly in Mexico could be less of a concern if were falling over time. Indeed, several studies found that there had been significant regional convergence in product per capa in different sub-periods between 1940 and However, they have also shown that regional convergence becomes insignificant in the eighties and that divergence is observed in the nineties. 3 This phenomenon is confirmed in this study using slightly different data from those employed in previous studies. Thus, the big question is: why did convergence stop in the eighties and why was there divergence in the nineties? Several hypotheses have been put forward to explain this phenomenon. Esquivel (1999) suggests that the lack of regional convergence in output per capa might be due to a lack of absolute convergence in post-primary education and to a slow process of migration across states. Messmacher (2000) argues that there has been relative regional convergence 1 There is though a very large fall wh respect to the difference observed in 1940, when the ratio was 11.8 times. 2 If we exclude the DF, which has a much higher level of GDP per capa in 2000 than any other state, and consider Nuevo León as the state of comparison, 67.7% of the population would inhab in a state wh less than half s level of GDP per capa. 3 See, for example, Juan-Ramon and Rivera-Batiz (1996), Esquivel (1999), Messmacher (2000), and Sánchez- Reaza and Rodiguez-Pose (2002). 2

3 in education so that, if the returns are higher for lower levels of education, then convergence in output per capa should have been observed. 4 This is the pattern that has generally been found in the microeconomic lerature estimating returns to education, 5 so there needs to be an addional factor explaining why convergence in education did not lead to convergence in output per capa. One possibily is an increase in the return to education, and particularly to higher levels of education. Recent technological change seems to have been complementary wh human capal, effectively leading to an increase in the return to human capal. 6 This is an issue we discuss in more detail further on, as there are other explanations for the increase in this return. Since the 1980 s, there have been far-reaching structural reforms in the Mexican economy. These have been oriented towards reducing the role of the state as a direct participant in economic activy and have included the deregulation of economic activy, privatizations of most government enterprises as well as a considerable opening up of the economy that started wh the entry of Mexico into GATT in 1986 and that was consolidated wh the start of NAFTA in Thus, the development model followed by Mexican authories has changed dramatically from an inward-oriented one wh strong government participation to an export oriented one wh the private sector driving economic activy. The effect of the structural reforms on regional inequaly is, in principle, ambiguous. Theoretical arguments can be made for and against increases in inequaly in general, and regional one in particular. For example, deregulation of economic activy may benef large and well-established firms, but could also promote the opening of new firms. Privatizations leading to cost-cutting and employment adjustments in previously 4 Another factor considered in cross-country studies is the growth rate of the population. Wh low migration at the regional level, this could also be a factor that might explain divergence. However, Messmacher (2000) finds that there was significant regional convergence in this variable between 1970 and For surveys of the lerature on returns to different levels to education see Schultz (1988), Psacharopoulos (1993), and Card (2000). 6 Papers that review the evidence on the effect of recent technological change on the return to highly skilled workers include Katz and Murphy (1992), Autor and Katz (2000) and Acemoglu (2002). 3

4 government-operated firms could lead to a reduction in the income of workers employed previously in these firms and in the regions where they operate. It should be noted, however, that is likely that these workers have higher skills than the average population, which then makes unclear what might be the effect of these measures on inequaly. Financial sector liberalization would ideally have benefed mostly small and non-tradable producing firms. Finally, the effect of opening to trade depends on which are the factors, regions and sectors that benef, so they depend on the specific circumstances of an economy and s trading partners. 7 Esquivel and Messmacher (2002) present an extensive review of the possible effects of greater economic integration on regional inequaly. We will return to this point later on, but by now is worth mentioning that trade could be one of the factors, besides technological change, that may explain the increase in the returns to human capal. Even though ex-ante was considered by many that NAFTA would disproportionately benef unskilled workers in Mexico given that this is a more abundant factor in Mexico than in the US or Canada, from a global point of view is probably the case that Mexico is relatively abundant in workers wh intermediate skills. And from the viewpoint of Mexico, these workers are probably located in the upper levels of the income distribution. 8 Messmacher (2000) argued that the structural reforms and, particularly, opening to trade and NAFTA may have favored certain types of activies, regions and factors that led to the divergence in the nineties. This analysis was done by looking at the performance of different sectors of economic activy and their proportion in a state s GDP. It was found that after NAFTA the sectors that outperformed the rest during the most recent period were, noticeably, manufacturing production and transport and communications. These are located mostly in the northern states and in a manufacturing belt around Mexico Cy, going from 7 For reviews of the reform process in Mexico see Aspe (1993) and Lustig (1998), in a general Latin American framework see Edwards (1995) and in Mexico s financial sector see Ortiz (1994). 8 It seems that a similar phenomenon occurred in other Latin American countries wh natural resource endowments being a relatively abundant factor from a worldwide point of view. For an extensive review on the subject see de Ferranti et.al (2001). 4

5 Jalisco to Veracruz. This characteristic may explain why these states outperformed the rest of the country in terms of economic growth. However, even if this is the case, there are still open questions about what were the factors that determined that some states specialized in these types of activies and why the rest have failed to catch up. Esquivel and Messmacher (2002) and Esquivel, Lederman, Messmacher and Villoro (2002) undertake growth regressions by decade and include a number of policy variables to analyze what were the determinants of growth in GDP per capa during the nineties and if there might have been some type of structural change around They present empirical evidence suggesting that locations lacking human capal, adequate infrastructure and wh a higher proportion of employment in the public sector tended to grow at lower rates than other states. In addion, they also find that distance to foreign markets and inial concentration of manufacturing or population might have been important, indicating the relevance of transport costs and economies of scale through agglomeration and other factors. An important finding is that, whereas there is evidence of absolute convergence in the period , there is still convergence in the period but only after controlling for the aforementioned variables (this is also known as condional convergence). Thus, these studies suggest that poorer and low-growth states need to undertake an important set of reforms such that policy variables might compensate the potential effect of distance and the inial advantage provided by economies of scale. This paper complements Esquivel and Messmacher (2002) and Esquivel et. al. (2002) in several important ways. The first thing that is done is a more detailed analysis of the evolution of the components of GDP per capa dividing them into productivy (i.e. product per employed worker) and a set of labor market and demographic variables. The policy implications of lack of convergence from productivy, labor market or demographic variables are substantially different, and the analysis of these can help us distinguish between the alternative explanations. In second place, we look at the evolution of productivy in greater detail. For this, we do two types of analysis. The first is a growthaccounting exercise separating productivy differences into those due to human capal and 5

6 an augmented residual. The second is an analysis of differences in productivy across states due to differences in the composion of productive activy and differences in productivy whin each sector. Finally, we undertake a panel analysis of growth in GDP per capa and in productivy that allows us to identify if there was structural change in the 1990 s. II. Sources of Differences in State GDP per capa In this section, we explore the evolution of state GDP per capa during the period decomposing into demographic, labor market and productivy components. The employed decomposion of GDP per capa is the following: Y N Y = L L EAP EAP WAP WAP N (1) where Y is real GDP, N is population, L is the economically active population that is employed (so Y/L is GDP per worker or labor productivy), EAP is the economically active population (so L/EAP is the employment rate), WAP is the working age population (so EAP/WAP is the participation rate), WAP/N is the proportion of working age population to total population, and i and t are sub-indexes by state and year. Equation (1) can be reexpressed in logarhms as: y n = y l + l eap + eap wap + wap n (2) From this expression we can decompose differences over time and across states in order to account for the contribution of demographic variables (wap/n), labor market structure and participation (l/eap and eap/wap), and differences in productivy (y/l) to the GDP per capa regional variance in Mexico. 6

7 In what follows, we first analyze the difference at a given moment in time of each of the GDP per capa components wh respect to the state wh the highest GDP per capa, then we study the presence of convergence in each of the different variables over time, and finally we decompose the rate of convergence in GDP per capa into that owed to s different components. Data on GDP by state comes from Esquivel (2002), whereas all the labor force and demographic data are from the censuses of 1940, 1950, 1960, 1970, 1980, 1990 and II.1. Differences in GDP per capa and s components across states Using the decomposion in (2), we can attribute, at a given moment in time, differences in GDP per capa across states to demographic, labor market or productivy differences. Specifically, we calculate the following expression: y n y n bt = y l y l bt + l eap l eap bt + eap wap eap wap bt + wap n wap n bt where underscore b stands for a base state against which the comparison is done. In this case, we have chosen the Federal District (DF), as is the enty wh the highest GDP per capa throughout the whole period. Table 1 shows the average difference per decade in each of the variables. The average difference wh the richest enty in GDP per capa fell from 1960 to 1980, but then increased in the last two decades. The increase in the average difference in GDP per capa from 1980 to 1990 was a result of an increase in the average difference in 9 Two adjustments were done to the data. A first problem is that oil production is allocated mainly to three states: Campeche, Chiapas and Tabasco, and the creria for allocation between these three states has changed over time. Given that oil includes rents and not only value added, and is considered to be a national asset (so tax revenues from are allocated to all the states), we substracted the original value of oil production to each state GDP and then reallocated according to population. Of course, this does not solve the problem of the economic rents vs. value added, but allocates the rent in such a way that does not bias inter-state comparisons of production. The second adjustment is due to the recognion that the population of Chiapas and Oaxaca was underestimated in 1980 (which in turn led to an overestimation of GDP per capa in these states). Thus, we make a projection of population in these two states between 1970 and These adjustments are described in more detail in Esquivel (2002). 7

8 productivy (y/l), and smaller ones in the participation rate (emp/wap) and the proportion of working age population (wap/n). The larger increase in the average difference in GDP per capa from 1990 to 2000 is a result of a relatively large increase in the differences in y/l, compensated by reductions in the average differences in all the other components. Table 1 Difference between the logarhm of a variable for a given state and the DF Averages y/n y/l l/emp emp/wap wap/n 1960 Average Std. Deviation Average Std. Deviation Average Std. Deviation Average Std. Deviation Average Std. Deviation Average over the period At any point in time, the variable that explains the largest part of average differences is labor productivy (82% of the difference in output per capa for the whole period), wh labor participation (eap/wap) in a distant second place (10% of the difference). In addion, the evolution of labor productivy seems to have been unfavorable from a convergence point of view during the last two decades, given the relatively large increase in s average difference. On the other hand, the evolution of labor market and demographic variables in the last two decades has not contributed to a reduction in the output per capa differences between the average state and the Federal District. 8

9 II.2. Convergence of GDP per capa and s components In this section, we analyze the evolution of GDP per capa and s components over time, wh particular emphasis on whether there has been beta and sigma convergence across Mexican states. Evidence of beta (sigma) convergence implies that relative (absolute) average differences in GDP per capa are falling over time. In order to analyze whether there has been beta-convergence we do a regression of the type: x t, t 1 = + βxt 1 α (3) where a negative and significant coefficient implies that relative differences have fallen across states, since states wh lower x have a higher growth rate of this variable. On the other hand, sigma-convergence is captured by the evolution of the standard deviation of x over time. Table 2 shows the results from beta convergence regressions of each of the variables. As has been noted in previous work (Esquivel (1999) and Messmacher (2000)), there is evidence of beta-convergence in GDP per capa for the period as a whole, but the pattern has changed considerably over time. The rate of convergence is posive from 1960 to 2000, as a whole, though by decade is only significant during the 1970 s. Finally, the point estimate in the nineties indicates slight divergence, though is not significant. This convergence pattern masks some important differences in terms of the causes of nonconvergence during each decade. The rate of convergence of labor productivy (y/l) is higher than that of output per capa (y/n) for the period of analysis as a whole. The rate of convergence is not found significant between the 1960 s and 1970 s, as in the case of GDP per capa, but increases over time and is significant in the 1970 s and 1980 s. Thus, the evidence suggests that from 1960 to 1990 convergence in productivy was becoming stronger. However, in the 1990 s there is significant evidence of strong divergence. 9

10 Table 2 Results from regressions of the following form: x = α βx t, t 1 + t 1 i) GDP per capa Constant Y/N t-1 Adj R *** ** (1.02) (0.12) (0.62) (0.1) (0.33) *** (0.04) *** (0.77) (0.08) (0.43) (0.05) Y/N * Less than 10%, ** Less than 5%, *** Less than 1% Standard error in parenthesis. Convergence rate ii) GDP per worker Y/L Constant Y/L t-1 Convergence rate Adj R *** *** (0.79) (0.10) 1.34 ** (0.52) (0.07) 1.29 *** *** (0.32) (0.04) 2.79 *** *** (0.65) (0.08) ** 0.13 ** (0.52) (0.06) %, ** Less than 5%, *** Less than 1% in parenthesis. 10

11 Constant iii) Employment rate Adj R (0.003) *** (0.11) *** (0.01) ** (0.27) *** *** (0.1) *** ** *** (0.01) (0.16) *** *** (0.15) * Less than 10%, ** Less than 5%, *** Less than 1% Standard error in parenthesis. L/Eap t-1 L/Eap Convergence rate iv) Participation rate Eap/Wap Constant Eap/Wap t-1 Convergence rate Adj R (0.092) *** (0.19) *** (0.09) *** (0.18) *** (0.123) *** (0.19) *** *** *** (0.14) (0.3) *** *** (0.082) (0.13) * Less than 10%, ** Less than 5%, *** Less than 1% Standard error in parenthesis. 11

12 v) Proportion of working age population Wap/N Constant Wap/N t-1 Convergence rate Adj R *** *** (0.117) (0.18) *** *** (0.11) (0.16) (0.08) (0.11) (0.09) (0.14) *** *** (0.07) (0.13) * Less than 10%, ** Less than 5%, *** Less than 1% Standard error in parenthesis. In the employment and participation rates, we observe high, fairly constant and significant rates of convergence during each decade and for the whole period. However, the rate of convergence for is smaller than that for any decade. This suggests that we may actually have repeated leapfrogging across states in these variables, wh states that are lagging behind picking up and surpassing the other states. The proportion of working age population also converges during the period as a whole, but there are important differences across decades. There is inially a strong convergence during the 1960 s, but there is no evidence that this process continued during the following two decades. It is until the 1990 s that we find significant convergence again. Graph 1 shows the evolution of the standard deviation of each variable over time. We find slight divergence in GDP per capa from 1960 to 1970, but then very strong convergence from 1970 to During the 1980 s we observe slight divergence and a very strong one in the 1990 s. Thus, while there is evidence of slight β-convergence from 1960 to 2000, implying that the difference in relative incomes fell during this period, the 1980 s is the only decade of the period when there is actually a reduction in the absolute differences in product per capa. 12

13 Graph 1 Standard Deviation of GDP per capa Across States and s Components i) GDP per capa ii) GDP per worker employed iii) Employed as a proportion of EAP iv) Participation Rate

14 v) Proportion of working age population Examining the different components of output per capa, can be seen that there is sigma convergence in all the components of GDP per capa during the 1960 s, wh the exception of the employment rate (l/eap). On the other hand, the strong output per capa convergence observed during the 1970 s is mainly due to the convergence in labor productivy. The latter component remained converging strongly during the 1980 s, but s effect on GDP per capa was more than compensated by the unfavorable evolution of the participation rate and the demographic variables. In the 1990 s, the evolution of these three components is the oppose, wh strong divergence in productivy, but a significant convergence in the other variables (though none of the variables returned to the levels observed in 1970). Thus, the above results show that the absence of significant beta and sigma convergence in GDP per capa during the 1980 s and 1990 s is actually due to two different phenomena. During the 1980 s, the evolution of the working age population and labor participation changed in such a way that output per capa diverges, even though productivy per worker is still converging. The decade of the 1990 s shows the oppose phenomenon, which could be due to the structural reforms, to the technological change or to NAFTA. We will return to this issue later. 14

15 II.3. Decomposion of the rate of convergence of GDP per capa The rate of beta convergence in GDP per capa can be decomposed, in an exercise similar in spir to the growth-accounting methodology, into that due to each of s demographic, labor market or productivy components. The methodology used is that of de la Fuente (1998, 2002). The growth rate of GDP per capa is: ( n) = ( y / n) ( y n) = xikt xikt 1 = xikt, t y, t 1 / 1 1 k k k / (4) where x ikt are each of the k (four) components of GDP per capa mentioned previously. From (4), is clear that the growth rate of GDP per capa is just the sum of each of the growth rates, and that the beta convergence equation defined in (3) can be simply reexpressed as: k x α βy (3 ) ikt, t 1 = + / n 1 It is also possible to estimate the effect of the evolution of each component separately, by regressing the growth rate of each component against the inial level of GDP per capa: x ikt, t 1 = k + β k y / n 1 α (3 ) The term β k gives us the partial convergence in GDP per capa due to the evolution of variable k. Table 3 shows the results for each of the components of GDP per capa, as well as the estimated coefficient for the GDP per capa. 10 During the entire period, the main driver of convergence in GDP per capa has been productivy, while the evolution of the proportion of working age population has led to a slight divergence in the margin. The coefficients estimated for the other components are also posive, though not significant. 10 In principle, the sum of all the partial convergence parameters β k will be equal to the total convergence parameter, β. 15

16 Results in Table 3 also show that productivy contributed to convergence from 1960 to 1990, though the coefficient was not significant during the sixties. On the other hand, productivy led to divergence during the nineties. The evolution of the participation rate limed convergence in a significant way during the 1980 s, but contributed to convergence in the 1970 s and 1990 s. In this last decade, somewhat compensated for the strong divergence due to the evolution of productivy. The evolution of working age population always lead to output per capa divergence to different degrees, though s effect is only significant during the 1980 s and, more strongly, when we analyze the whole period. Table 3 Results from regressions of the form: x = α β y n ikt, t 1 k + k / 1 dy/n dy/l dl/emp demp/wap dwap/n ** *** ** (0.118) (0.107) (0.027) (0.021) (0.071) (0.068) (0.006) (0.027) (0.020) *** *** ** ** ** (0.036) (0.039) (0.006) (0.026) (0.009) *** *** (0.083) (0.079) (0.003) (0.029) (0.016) ** *** *** (0.045) (0.048) (0.003) (0.026) (0.019) * Less than 10%, ** Less than 5%, *** Less than 1% Standard error in parenthesis. 16

17 III. Decomposion of Productivy The above results indicate that differences in productivy account for most of the differences in GDP per capa across states. In addion, productivy has diverged in the last decade, contrary to the evolution of demographic and labor market variables. Thus, is of particular importance to analyze in more detail the determinants of labor productivy. The evolution of productivy over time and, more recently, productivy differences across countries or regions, have usually been analyzed using a growth-accounting framework based on the following decomposion: y l = ln( A ) + α k l + β h l + γ t (5) where ln (A) is the log of total factor productivy (TFP); k is physical capal; h is human capal; l is employment; t is an addional component capturing things as infrastructure, land, climate and transport costs; α, β and γ are the output elasticies of physical capal, human capal and infrastructure, respectively. The above specification assumes constant returns to scale in k, h and l, and increasing returns when infrastructure is taken into account. 11 The data for the above decomposion is unfortunately not available as we don t have appropriate measures of A and k at the state level in Mexico, and we definely need at least one of them (the other can be calculated as a residual if we know α, β and γ). Two indirect alternatives will be to analyze differences in productivy across states and over time. The first approach is based on the use of data on human capal (education) and employment, and the second is based on the analysis of productivy by sector of economic activy in each state. 11 For an extensive revision of the growth-accounting methodology see Barro (1999). 17

18 III.1. Analysis of productivy using human capal and employment Equation (4) can be expressed in the following way: y l = A ) + β ln( h l (6) where ln( A ) = ln( A ) + α k l + γ t can be thought of as an unobserved and augmented measure of TFP. Using (6) we proceed along two types of analysis. First, we attribute differences in productivy across states to differences in education and to the augmented residual. We can do this by calculating the following expression: y / l bt bt bt y / l = β ( h / l h / l ) + ln( A ) ln( A ) (7) where b is chosen as a state against which all others are compared. The difference in augmented TFP is not observed, but is calculated as a residual given that we know the cross-state differences in productivy and in human capal per worker. In order to assign a value to β, we assume that the aggregate production function exhibs constant returns to scale and that there is perfect competion in the economy. Under these condions, β is equal to the share of human capal in income. Unfortunately, as noted by Bernanke and Gurkaynak (2000), there is serious measurement error in the calculation of human and physical capal shares in non-industrial countries. In the particular case of Mexico and several other Latin American countries, most of the national statistical offices attribute the income of the self-employed to physical capal income, since they treat all of their income as if they were profs. This means that no account is made of the implic salary of the self-employed and, since a large segment of the self-employed people in Latin America is composed of relatively low-skilled workers in the informal sector, this means that their implic salary is probably a large fraction of their total income. This explains the relatively high physical capal shares found for Latin 18

19 American countries in studies such as Elias (1990). Thus, we do the analysis using a value for β of 0.4 as a central case, and also analyze the sensivy of the results to slightly higher or lower values for this parameter. 12 A second alternative to assigning a value to β is to estimate by running a regression across states and over time in eher levels or in differences wh respect to a given state: y l = α + β h + ε (8) l y / l y / l = β ( h / l h / l ) + ν (9) bt bt The problem wh the regressions is that human capal is likely to be posively correlated wh physical capal, infrastructure and TFP, so the coefficient on this variable would be biased upwards. However, if education is the factor wh the lowest rate of accumulation, and determines the levels of A, k and t (particularly the first two, being capal and technology complements to human capal), then this bias would actually be capturing economically significant effects from human capal accumulation. This interpretation is probably more relevant in an international context where labor mobily is more limed. For to be valid in an intra-country analysis, is necessary that migration is limed due to high costs or instutional and social factors, or that capal and technology are more mobile than labor. Thus, an alternative interpretation is that agents have more incentives to accumulate human capal in regions where there are high levels of technology and physical capal per worker. Although, may also be the case that individuals wh high human capal have more incentives to migrate to these regions. 12 Values of shares estimated for industrial countries are typically 0.3 for physical capal and 0.7 for labor, out of which 0.4 corresponds to human capal and 0.3 to simple labor. See Barro and Sala-i-Martin (1995) and Bernanke and Gurkaynak (2002). On the other hand, Gollin (2002) shows how simple adjustments to the standard calculation of factor shares in most countries produce estimates of the labor share similar to those observed in developed countries. These estimates tend to cluster in a range that varies from 0.60 to 0.85 depending on the adjustments made. These results are compatible wh our assumption of β=0.4 19

20 In order to proxy for human capal per worker in the empirical analysis that follows, we use average years of schooling for the population 15 years and older. This information is from the Mexican Ministry of Education (SEP). The results from a growth-accounting type decomposion using (7) are shown in Table 4, using the Distro Federal (Mexico Cy) as the state of comparison. The table shows that differences in the augmented measure of TFP account for a substantial part of the differences in productivy at any point in time. Moreover, is found that the importance of augmented TFP has continuously increased over time as a source of the differences in productivy across states (s contribution increased monotonically from 62% of the productivy differences in 1960 to 86% in the year 2000). The average difference in the productive component of human capal by state has fallen continuously during the period in both absolute and relative terms. 13 Table 4 Decomposion of the difference in productivy between the DF and other states into differences in human capal and in an augmented TFP (β=0.4) ln Y/L i - lny/l df ln A i - lna df β(ln H/L i - lnh/l df ) 1960 Average Standard Dev Average Standard Dev Average Standard Dev Average Standard Dev Average Standard Dev Average for the entire period Using higher values of β we find that human capal accounts for a larger fraction of the differences in productivy across states. However, the trends described above are found to be the same. 20

21 Table 5 shows the results from a regression of productivy on the average level of schooling as indicated in equation (8). We do several exercises, first including all the observations for all states, then wh year dummys, next wh fixed state effects, and finally wh interactive year dummys. The estimated parameter associated wh average schooling is always strongly significant. It is que large, ranging from 0.32 to Of special interest are the results wh the interactive time dummys. It is found that there are gradual increases in the coefficient associated wh education from 1970 to 1990 and finally a very large increase during the 1990 s. This suggests that the importance of average schooling as an explanatory variable for the cross-state variation in productivy has become significantly more important over time, wh a very large increase during the last decade. In all cases, is found that average schooling contributes to explain a substantial proportion of the inter-state variation in productivy, as attested by the adjusted R 2 statistics. Table 5 Results from regressions of the form: y l = α + β h + ε l Period (1) (2) (3) (4) (5) Year dummies Interactive Year Fixed and fixed year dummies effects effects dummies Constant ** *** *** *** (0.079) (0.076) (0.134) (0.136) (0.096) Average Schooling *** *** *** *** *** (0.052) (0.076) (0.045) (0.119) (0.106) A.S * D ** (0.178) A.S * D *** (0.208) A.S * D ** (0.268) A.S * D *** (0.320) Adj. R-squared Observations * Less than 10%, ** Less than 5%, *** Less than 1% Standard error in parenthesis. 14 The latter value is obtained by adding up the general coefficient on average schooling in the last column and the coefficient for the year 2000 dummy. 21

22 Finding coefficients associated wh human capal that are larger than what is presumably s share of income implies that we are eher identifying some type of externaly, or that human capal is posively correlated wh the level of TFP and physical capal in a state. As mentioned previously, this would be expected if human capal is complementary to TFP and physical capal. In such a framework, the evolution of human capal would actually condion the evolution of TFP and physical capal if were the variable wh the slowest rate of accumulation. The close relationship between physical capal, technology and human capal can be confirmed by observing the correlation between the augmented TFP differences across states that were calculated employing (7), and the differences in human capal. The rate of correlation between these variables is always strongly posive and reaches s lowest level in 1960 (0.48), precisely when the coefficient on education is the lowest in the specification wh interactive year dummies. Then increases sharply in 1970 and remains high ever since (the average correlation for the period is 0.71, ranging between 0.68 and 0.77). Thus, the above results suggest that differences in the stock of human capal directly explain part of the differences in productivy across states, but still the largest proportion of such differences is accounted for by unobserved variables, probably physical capal and TFP. Differences in the last variables seem to have been increasing over time. Lastly, our regression results suggest that the relationship between human capal and productivy seems to have grown over time, though presumably not by a direct channel that explains differences in productivy but rather due to close links between human capal, physical capal and TFP. III.2. Differences in Productivy due to Differences in Productive Structure Data on labor productivy (Y/L) by broad sector of activy are available at the state level in Mexico. Thus, is possible to decompose differences in state labor productivy into differences in labor productivy by sector and differences in sectorial composion by 22

23 state. 15 In this section we will also calculate the convergence rate in productivy across states by sector of activy and the convergence rate in sectorial composion by state. To begin wh, let us note that GDP per worker in a state can be decomposed as: Y L = n j= 1 Y L ijt ijt L L where i stands for the state and j for sector of activy. ijt (10) We can now decompose the difference in GDP per worker between any state and a benchmark state into three components: one associated to differences in the composion of the productive structure (Diff share ), one associated to differences in productivy (Diff prod ), and a third effect that is a compose of both differences and that can be obtained as a residual (Diff comp ). That is, Y L Y Y L bt L bt = Diff sharei + Diff prodi + Diff where the sub-index b represents the benchmark state, and where the differences in productivy are expressed as percentages of labor productivy in the benchmark state. As before, we will choose as our benchmark the state wh the highest Y/L, so the comparisons will be done against the Federal District. compi Thus, the following three indicators will be calculated: a) the difference in GDP per capa due solely to sectorial composion assuming that the productivy in each sector is equal to s national average in every state: Diff sharei = n j= 1 L L ijt L L bjt bt Y L ijt ijt Y L bt (11) 15 This exercise is similar in spir to a standard shift-and-share analysis. See Esteban (2000) for a recent application of this technique to the study of regional convergence in Europe. 23

24 b) the difference in GDP per capa due solely to productivy differences assuming Diff that the share of each sector in employment is equal to s national average in every state: prodi = n j= 1 Y L ijt ijt Y L bjt bjt L L ijt Y L bt (12) c) a compose effect calculated by allowing for both sectorial shares and productivy to vary by state minus the results from (11) and (12): Diff compi Y L Y L bt = Diff share Diff i prodi Y L (13) bt It is also possible to relate the above results of shares and productivy wh education by state. The size of the differences and how they have evolved over time can have important implications in terms of the causes of the differences. For example, if differences in productivy are small or have been falling, while sectorial composion differences are large and have not fallen, would suggest a problem of factor accumulation in consistence wh a Hecksher-Ollin trade model. If is the other way around, then there are probably some type of fixed factors that preclude the employment of the same technology in a given sector. Productivy by sector was calculated using: i) the data on state GDP by sector of activy, created by INEGI, for the years 1970, 1980 and 2000 as well as a value for 1990 that was calculated using the state GDP data of 1988 and 1993, 16 and ii) the data on employment by sector of economic activy from the Census of 1970, 1980, 1990 and The sectors are: agriculture, livestock, forestry and fishing; manufacturing; construction; electricy, gas and water; commerce, restaurants and hotels; transports, communications and storing; other services (financial services and social, communal and personal services). The mining sector was excluded from all the calculations given that s GDP values are 16 As described in Esquivel (2002). 24

25 likely to include a very substantial economic rent and not just value added. Therefore, the standard measure of labor productivy probably does not reflect the average product of labor in this sector. A problem wh the use of Census data on workforce by sector of economic activy is that there is a certain proportion of the economically active population whose sector of activy is classified as Insufficiently specified. This proportion is less than 10% on average for all states in 1970, 1990 and 2000, but in 1980 is slightly larger than 30% on average for all states. In order to control for this, the proportion of workforce in each sector of activy is calculated as the economically active population minus the unemployed minus those classified as working in Insufficiently specified activies. Nevertheless, is likely that the results for 1980 are subject to large measurement error problems. If the Insufficiently specified are randomly distributed across the different sectors of economic activy wh the probabily depending on the number of workers in each sector, then measures of the share of the workforce in each sector should be relatively adequate. However, using a smaller workforce to calculate productivy can lead to a large overestimation of productivy in Thus, we also check our results recalculating productivy by assigning the Insufficiently specified to the workforce in each sector depending on the proportion of the workforce in each sector. The results changed slightly in terms of magnude but not in terms of trend or from a qualative point of view. Table 6 shows average productivy and the share of the employed labor force in each sector of economic activy. The table shows that there has been a very dramatic reduction in the proportion of the labor force working in agricultural activies, as fell during the entire period from 50.3% to 18.5% of total workforce. This has been matched by a very large increase in the proportion of services, including commerce, restaurants and hotels. These have increased from a proportion of 27.8% percent of employment in 1970 to 49.3% in The third change in importance is the increase in the proportion of the workforce employed in manufacturing, from 14% in 1970 to 18.5% in

26 Table 6 Productivy and Proportion of the Workforce by Sector of Economic Activy Year Averages and Standard Deviations Total Agriculture, Livestock, Forestry and Fishing Manufacturing Construction Electricy, Gas and Water Commerce, Restaurants and Hotels Transport, Warehouse, Communications Aggregate Services (1)+(2) Financial Services, Insurance, Real State (1) Communal, Social and Personal Services (2) 1970 GDPi/Li Average St. Dev Li/L Average St. Dev GDPi/Li Average St. Dev Li/L Average St. Dev GDPi/Li Average St. Dev Li/L Average St. Dev GDPi/Li Average St. Dev Li/L Average St. Dev Financial services, public utilies (electricy, gas and water), and transport and communications have, in that order, the highest levels of labor productivy. The three are highly capal intensive activies and they account for very small proportions of the workforce (6.1% in 2000). The following sectors in terms of productivy are commerce, restaurants and hotels; manufacturing; communal social and personal services; construction; and agricultural activies are last. One addional point is that, in the four sectors wh relatively large changes in the proportion of the workforce (agriculture, services, manufacturing and construction), productivy has moved in different direction. While productivy has increased in the agricultural sector, has fallen in services (excluding financial activies and including commerce, restaurants and hotels), manufacturing and construction. In spe of this, productivy in these three sectors has remained above that of agriculture, but is striking to note that, as of 2000, the difference between construction and agriculture is almost negligible while that wh communal, social and personal services has also become much smaller. In addion, there seems to be an 26

27 important degree of convergence in productivy in those sectors that employ more than 5% of the workforce. This would be consistent wh a neoclassical model in which there is a reallocation of labor from low to high wage and productivy sectors. The standard deviation of productivy in each sector is considerably higher than that of the share of the workforce in that sector, wh the exception of agriculture, which already suggests that differences in productivy whin each sector of activy probably account for a large part of the differences in aggregate productivy across states. As to their evolution over time, the standard deviation of the share of the workforce in agriculture and communal, social and personal services has fallen over time, while that of the manufacturing industry and of commerce, restaurants and hotels has increased. In terms of the standard deviation of productivy, there seems to be an important increase from 1980 to 1990, but most sectors don t show any define trends, wh the exception of transport and communications and financial services which show a posive trend. The results from the decomposion of differences in aggregate state productivy in differences due to sectorial productivy and specialization corresponding to equations (11), (12) and (13) are shown in Table 7. We find that in 1970 and 1980, the average difference in the allocation of the workforce to different activies accounts for a large part of the average difference in productivy wh Mexico Cy. The average differences of productivy in a given sector are very small, while compose effects are of intermediate importance. In other words, in 1970 and 1980 differences in average productivy wh Mexico Cy are mostly accounted by the fact that a larger proportion of the workforce in other states is employed in activies wh relatively low productivy levels. The fact that a given sector may be more or less productive in Mexico Cy than in other states explains very ltle of the differences in aggregate productivy during this period. However, this suation changed dramatically in 1990 and The importance of differences in the allocation of the workforce to sectors wh different productivy remains at a similar level in 1990, but then falls significantly in In contrast, there was a very large increase in the differences in productivy for the same sector across different states in 1990, and then even increased again in

28 Table 7 Decomposion of Average Differences in Productivy wh respect to the DF Differences owed to allocation of the workforce, productivy differences for a given sector and compose effects Total Productivy Workforce Share in each Sector Productivy in each Sector Compose Effect Convergence regressions We now performed β convergence regressions for both the share of the workforce and labor productivy in each sector. However, in order to know whether divergence or convergence in a given economic variable will lead to convergence in overall productivy, we need to know the level of productivy of a given sector as well as the correlation between overall productivy wh both the degree of specialization in a given activy and the level of productivy in that activy. Table 8 shows the correlations across sectors of productivy for each year in the sample. Productivy in utilies (electricy, gas and water) has a negative, though low in absolute value, correlation wh overall productivy in a state. This is not surprising as the location of these sectors of activy are strongly affected by the location of the natural resources on which the activy depends, which might be unrelated wh the overall productivy of the state. 17 Productivy in all the other sectors is always posively and highly correlated wh overall productivy in a state, wh the exception of services, which start wh a strong negative correlation in 1970, but then the correlation increases and becomes high and posive in 1990 and These results imply that convergence in productivy of a given sector, wh the exception of utilies in all the period and services in 1970, should lead to convergence in overall productivy of a state. Moreover, these results are indicative that places wh high overall productivy tend to have high productivy in all sectors of economic activy. 17 One might have expected the same about agriculture, but turns out that productivy in this sector is always posively and highly correlated wh overall productivy. 28