Dynamics of Employment and Productivity in Developing Countries

Transcription

1 Dynamics of Employment and Productivity in Developing Countries M. Jahangir Alam Department of Economics University of Calgary Canada Abstract: I establish a systematic pattern of employment and productivity distributions over establishment age with development indicators. For example, establishments grow less in developing countries than in developed countries. These patterns can help to understand the cross-country income differences. Using the Enterprise Surveys dataset, I show that the average size of an establishment grows as an establishment ages; however, this growth varies across countries. As well, the relative size of older establishments is much smaller in developing countries than in developed countries. Keywords: Employment, Productivity, Manufacturing, Pattern, Cross-country JEL classification: O140, O470 Preliminary Draft: Please do not cite or circulate without the author s permission. I am grateful to Joanne Roberts and Trevor Tombe for their valuable comments. Matthew Webb for helping on bootstrapping techniques. I am thankful to

2 1 Introduction Employment and productivity distributions over the life cycle 1 of manufacturing establishments 2 potentially play a larger role in cross-country income differences. Hsieh and Klenow (2012) show that moving from a U.S. life cycle of plants to the Indian or Mexican life cycle could plausibly account for a 25 per cent drop in aggregate total factor productivity (TFP). In the U.S., establishments tend to start small and substantially grow as they age (Dunne et al., 1989; Atkeson and Kehoe, 2005). Specifically, the employment level, which is the number of employees in an establishment, is lower in early stages of the life cycle and eventually grows with age. The level of labor productivity follows a similar pattern to employment. Employment and productivity level over the life cycle of manufacturing establishments differs by country. To establish the fact, Hsieh and Klenow (2012) show that the employment level over the life cycle of plants is much smaller in Mexico than in the U.S. and is much lower in India than in Mexico. Surprisingly, there are no studies estimate employment and productivity distributions by establishment age broadly over several countries. To identify the reason for a specific pattern, the first step is to establish a pattern of employment and productivity distributions for several countries, especially for developing countries. Thus, I ask the research questions: what is the pattern of employment and productivity distributions by establishment age in developing countries? Do the employment and productivity distributions by establishment age for India and Mexico hold in a larger set of countries more broadly? To estimate employment and productivity distributions by establishment age, I use the rotating panel dataset from the Enterprise Surveys 3,whichprovidescomparabledataacross countries and industries. To estimate the employment distribution, I regress the employment level on age cohort dummies for each industry and country in a given year. Using industry value added weights, I calculate weighted coefficients by aggregating industry coefficients. To estimate the relative employment level, I rescale the weighted-coefficients by considering 1 According to Atkeson and Kehoe (2005), manufacturing establishments have a clear life cycle: they are born small, grow substantially with age, and eventually die. 2 Establishments are defined by a specific physical location. A firm may be composed of one or more establishments. The establishment identifier remains the same even when the establishment changes ownership. 3 Enterprise Surveys ( The World Bank, January

3 the younger cohort as a numeraire. Due to the small sample size in the Enterprise Surveys, I use a bootstrapping method to calculate standard errors and confidence intervals. In that manner, I estimate the productivity distribution by establishment age. To validate a general consistency of my estimates, I compare the employment and productivity point estimates of the Enterprise Surveys with the Hsieh and Klenow (2012) estimates for Mexico. The Hsieh and Klenow (2012) estimates fall within the bootstrap confidence intervals established in the results here, indicating a general consistency. Since the estimates for Mexico are consistent, Iusesimilarmethodtoestimatethedistributionsfortheothersurveyedcountriesincluded in the Enterprise Surveys. My results show that the employment distribution by establishment age, with few exceptions, demonstrate a systematic pattern in the surveyed countries of the Enterprise Surveys. In roughly two-third countries, establishments with age 25 or more have the employment level far below with same cohort in U.S. establishments. Even though the remaining one-third countries have similar employment levels compared to U.S. establishments, the employment coefficients have higher bootstrap standard errors and wider bootstrap confidence intervals. Thus, the Hsieh and Klenow (2012) results for India and Mexico hold more broadly in a larger set of countries. The results further suggest that the relative size of older establishments is much smaller in developing countries than in developed countries. This implies that the relative size of establishment can be used as a proxy to measure of cross-country income differences. To support my argument, I plot point estimates of age cohort 25 or more against per capita income. The results show that the age cohort coefficient is positively correlated with per capita income. Thus, employment and productivity distribution by establishment age can be used as a measure of cross-country income differences. The existing literature on growth and productivity has developed along two themes: the vintage effect and the survival effect. The vintage effect says that recent cohorts start out with higher productivity than earlier entrants did; the survival effect says that productivity increases as establishments age (Jensen et al., 2001). The literature on the survival effect is further expanded to include selection mechanisms and survivor growth. Selection mechanisms, which reflect establishment-specific decisions and productivity shocks, are im- 2

4 portant determinants of the observed relationship between age and productivity (Jovanovic, 1982; Baily et al., 1992; Ericson and Pakes, 1995; Luttmer, 2007). With survival growth, as establishments age, managers accumulate experience, gain from learning by doing, undertake new investments, or achieve economies of scale, all of which can improve productivity of establishments (Bahk and Gort, 1993; Bartelsman and Doms, 2000; Hsieh and Klenow, 2012). My study contributes by establishing new a fact to explain cross-country income differences to survival growth literature, which emphasizes the importance of the accumulation of plant-specific organization capital. Organization capital increases with age. In the U.S. manufacturing sector, around 40 percent of intangible capital payments are attributed to organization capital (Atkeson and Kehoe, 2005). Employment and productivity distributions over the life cycle of establishments are driven by the accumulation of organization capital (Atkeson and Kehoe, 2005). The empirical literature on the evolution of establishments, through employment and productivity distributions estimation, focuses mainly on U.S. establishments. My study is closely related to Hsieh and Klenow (2012), who compare employment and productivity distributions over the life cycle of plants in India and Mexico to that in the U.S. They find that forty-year old plants in U.S. are almost eight times larger, in terms of employment, when compared with plants less than five years old. In India, old plants are no larger than young plants. In Mexico, 25 year old plants are more than twice the size of new plants. Akeycontributionhereistoestablishafactthatemploymentandproductivitydistributions by establishment age demonstrate a systematic pattern. To establish this fact, I estimate employment and productivity distributions more broadly in a larger set of countries. In addition, I estimate the bootstrap estimate of standard errors and the bootstrap confidence intervals of employment and productivity distribution estimates. The paper is organized as follows. Section 2 provides a detail description of the Enterprise Surveys data and steps taken to clean the dataset. It also includes the empirical method to estimate employment and productivity distributions by establishment age. Section 3 contains estimation of employment and productivity distributions by establishment age and results. The paper concludes with Section 4, which also includes the potential areas of future research. 3

5 2 Method 2.1 Data To estimate employment and productivity distributions by establishment age, I use the rotating panel dataset from the World Bank Enterprise Surveys. The Enterprise Surveys, conducted by private contractors on behalf of the World Bank, contains an expansive collection of data on 130,000 establishments in 135 countries. The primary sampling unit of the Enterprise Surveys is the establishment. A major strength of the dataset is that the Enterprise Surveys uses standardized instruments and a uniform sampling methodology. This minimizes measurement error and provides comparable data across countries and industries. Due to this feature of the dataset in the Enterprise Surveys, this dataset is appropriate to examine whether the Hsieh and Klenow (2012) results for India and Mexico hold more broadly in a larger set of countries. The Enterprise Surveys is a representative sample of an economy s private manufacturing industry. Establishments with 100% government or state ownership are not included. The sample frame of the Enterprise Surveys is prepared by compiling a list of eligible establishments from the country s statistical office and on occasion from other governments agencies, such as tax or business licensing authorities. Formal (registered) manufacturing establishments with 5 or more employees are interviewed. The Enterprise Surveys dataset contains detailed information about these establishments access to finance, corruption, infrastructure, crime, competition, and performance measures. To ensure the proper sample representation, the Enterprise Surveys uses three levels of stratification: two-digit industry, establishment size, and geographical region. Two-digit industries are food, textiles, garments, chemicals, plastic and rubber, non-metallic mineral products, basic metals, fabricate metal products, machinery and equipment, electronics, and others. In the Enterprise Surveys, establishment size are 5-19 (small), (medium), and 100 or more employees (large-sized establishments). In general, large-sized establishments are oversampled since larger establishments tend to have greater impact on employment and productivity growth. In each country, geographical regions are selected based on which cities or regions collectively contain the majority of economic or business activity. 4

6 In the Enterprise Surveys, the overall sample size within an economy depends on the sample size for each level of stratification. 4 Moreover, the stratification depends on the size of the economy, measured by the Gross National Income (GNI). In general, 150 surveys are conducted in small economics, 360 interviews are conducted in medium-sized economies, and for large economics, surveys take place. To estimate employment and productivity distributions, I use sixteen countries dataset from the rotating panel 2006 and 2010 and thirteen countries dataset from the rotating panel 2005 and 2009 (Tables 1 and 2). I drop five countries from dataset in 2005 and 2009 because of insufficient observations. In addition, three types of sampling weights are calculated in the Enterprise Surveys since To compare estimates across countries, it is recommended by the Enterprise Surveys to use the median weights. To ensure proper representation of the Enterprise Surveys sample for each country, I use sampling weights for calculating descriptive statistics and estimating employment and productivity distributions. A potential problem of the Enterprise Surveys is that the resulting datasets represent only establishments that are willing to participate in the survey, refers to self-selection bias. Due to attrition and participants self-selecting out of the survey, additional establishments were surveyed to reach the original target sample size per stratum. In addition, the non-response observations are another problem in the Enterprise Surveys. I follow several steps to clean the dataset (see Appendix A). The number of observations after cleaning, both panel and cross-section, are shown in Tables 1 and 2. Since few countries are surveyed in low-income and higher income categories, I make two groups based on the income grouping from the Enterprise Surveys: the low-income and lower middle-income categorize as low-income countries; the upper middle-income and higher income categorize as middle-income countries. Table 3 shows means and standard deviations of both employment and age of establishments in 2009/2010. The employment level of manufacturing establishments has higher 4 1 n = N + N 1 1 N P (1 P ) k z 1 α Where n is the sample size, N is the population size, P is the population proportion, k is the desired level of precision, and z 1 α is the standard normal value for a desired level of confidence, 1 α. 2 5

7 variation (standard deviation) compared to age variation of establishments. It implies that the employment level significantly varies across establishments. Based on the Kernel density estimation, the age distribution of middle-income countries is positively skewed compared to low-income countries (Figure 1.1). Due to small sample size in the Enterprise Surveys, I group the age variable into three age cohorts. The age cohorts are less than 10 year, 10 to 24 year, and 25 year or more. Without using the value added weights 5,theemploymentdistributionbyestablishmentage is shown in Table 4 for 2009/2010. It shows that establishments with age 25 or more have higher standard deviation than age less than 10 and age This result implies that manufacturing establishments start with a similar number of employees, but employment level grows differently as they age. Based on the Kernel density estimation, the employment distributions of different age cohorts are not symmetric around at the mean. In addition to this, the employment distribution of older age cohorts are positively skewed than younger age cohorts (Figure 1.2). Thus, the employment level of manufacturing establishments increases as they age. Figure 1.3 shows the employment distribution by establishment age. Without using the per capita income weights 6,theresultarepresentedfor2009/2010forlow-incomeand middle-income countries. To show employment level variation for a given age cohort, I add confidence intervals. It shows that the employment level grows over establishment ages in both groups of countries. Moreover, middle-income countries have higher employment level at each age cohort. It also shows that the employment level of establishments that are age 25 or older have three times the employees in middle-income countries than they had at age less than 10. In low-income countries, establishments that are age 25 or older have two and half times the employees than they had at age less than 10. Furthermore, low-income countries have higher standard errors of the employment level compared to middle-income countries (Figure 1.3). 5 I calculate value added weights for a given age cohort by aggregating the value added for each two-digit industry and then divided by the total value added of industries. 6 I calculate per capita income weights for aggregating measures across countries by using per capital income data from the Penn World Table. 6

8 2.2 Method of Estimation To estimate employment and productivity distributions by age for a given year, I regress a measure of distributions (E eij )onagecohortdummies,onlythreedummiesuseinthisstudy for each highly aggregated two-digit industry, for each industry (i) and country (j): E eij = c I c I c I β 1ij Age 1ij + β 2ij Age 2ij + β 3ij Age 3ij + eij (1) j=1 i=1 j=1 i=1 j=1 i=1 where Age1, Age2, Age3 areagecohortdummiesandcisthetotalnumberofcountriesina given year. I use three measures of E eij.thesemeasuresareemploymentlevel,employment growth, and productivity growth. Using the value added weights (α i ) of manufacturing industries (I), I aggregate age cohort coefficients. For a given age cohort and country, the sum of value added weights is one. I multiply each two-digit industry coefficient for given age cohort and country by the calculated value added weights and then add to calculate the aggregate age cohort coefficient. β aj = I α i β aij,where i=1 I α i =1 a =1 to 3, j =1 to c (2) i=1 Irescaletheweighted-coefficientsforeachagecohortbyconsideringtheyoungestcohort as a numeraire. This implies that the age cohort coefficients of employment and productivity distributions by establishment age are reported relative to the youngest cohort: δ 1j = β 1j β 1j =1,δ 2j = β 2j β 1j, and δ 3j = β 3j β 1j (3) To estimate age cohort coefficients, I use Generalized Method of Moments (GMM) since it requires less restrictive assumptions. Since bootstrapping provides more accurate inferences when the sample size is small and also it does not require any distributional assumptions, I calculate bootstrap standard errors for age cohort coefficients of employment and productivity distributions. Moreover, bootstrapping avoids some unnecessary technical complexities to 7

9 calculate standard errors for aggregate coefficients based on the individual standard error of two-digit industry. In addition to the bootstrap standard errors, I calculate the asymmetric bootstrap confidence interval of estimates. I calculate the asymmetric confidence intervals since employment distributions for all age cohorts are not symmetric at the mean (Figure 1.2) Wild Bootstrap To calculate the bootstrap estimate of standard errors and the percentile bootstrap confidence intervals of estimates, I use the Wild bootstrap (Wu, 1986; Angrist and Pischke, 2008). I use this method since it does not require the assumption of the error terms are independently and identically distributed. To avoid extreme value in the bootstrapping, I use this method since large-sized establishments are oversampled in the Enterprise Surveys. The idea is, like the residual bootstrap, to leave the regressors at their sample value, but to resample the response variable based on the residuals values. This method assumes that the true residual distribution is symmetric and can offer advantage over simple residual sampling for smaller sample sizes. The most straightforward approach of bootstrapping is to collect the response-variable value and regressors for each observation. This method is called pairwise bootstrap. Flachaire (2005) shows with some experiments that the pairwise bootstrap is sensitive to high leverage observations. The Wild bootstrap does not suffer from this limitation. The following are the steps followed for the Wild Bootstrap: 1. Estimate the equation (1.1) by GMM estimation and retain the fitted values, Êeij and the residuals, ˆ eij, (e =1, 2,...,n). 2. For each replicate, the data-generating process (DGP) is- E eij = c I c I c I ˆβ 1ij Age 1ij + ˆβ 2ij Age 2ij + ˆβ 3ij Age 3ij + f(ˆ eij )v (4) j=1 i=1 j=1 i=1 j=1 i=1 where Age1, Age2, Age3 aretheoriginalagecohortdummies.f(ˆ eij )isatransformation of the estimated residuals and v is a random variable with mean 0 and variance 8

10 1. Possible options of the f(ˆ eij )andthev are explained in Appendix B1. 3. Calculate the weighted-coefficients using the value added weights of industries. 4. Rescale the weighted-coefficients by considering the younger cohort as a numeraire, and then save the rescaled weighted-coefficients, which are employment and productivity distribution estimates by establishment age. 5. Draw another replication 7 and follow the previous steps, from two to four Bootstrap Standard Error The bootstrap estimate of standard error (Efron and Tibshirani, 1993) is the standard deviation of the bootstrap replications (B): SE( ˆδ a )= B b=1 ( ˆδ a δ a) B 1 (5) where ˆδ a is the bootstrap coefficient for each bootstrap replication and δ a is the mean of all replications coefficient ˆδ a Asymmetric Percentile Bootstrap Confidence Intervals Icalculatetheasymmetricpercentilebootstrapconfidenceintervals(EfronandTibshirani, 1993; Davison and Hinkley, 1997) of estimates. The percentile bootstrap works well in cases where the bootstrap distribution is symmetrical and centered on the observed statistics. In addition to this, the percentile bootstrap works where the sample statistic is medianunbiased and has maximum or minimum concentration in the data. Suppose, the vector of employment and productivity distribution estimates by establishment age is δ a = η(f )and that T n = η(the empirical distribution of E 1,E 2,...,E n ) 7 A method of choosing the number of bootstrap replications B developed by Andrews and Buchinsky (2000). The method of Andrews and Buchinsky (2000) applies to a wide variety of bootstrap statistics and application, though the method use in this study on selecting B to estimate standard errors and percentile confidence intervals. In addition, to identifying number of replications, Davidson and Flachaire (2008) have alternative method, which applies only to bootstrap p-values. 9

11 Based on B bootstrapped values T n(1),t n(2),...,t n(b),defineorderedvalues: T n(1) T n(2)... T n(b) To locate the lower and upper percentile bootstrap confidence intervals of the B bootstrapped valuesk L = α 2 (B +1) and k U =(B +1) k L where α(b +1) is the largest integer less than or equal to α (B +1). Thus, the asymmetric 2 2 percentile bootstrap confidence intervals for the weighted-coefficients- T n(kl ),T n(k U ) (6) 3 Estimation and Results To estimate employment and productivity distributions by establishment age, I use both cross-section and two-year panel observations. To measure the employment distribution using the cross-section data, I estimate age cohort coefficients by regressing employment level on age cohort dummies for each industry and country. Using two-year panel observations, I estimate the employment growth distribution by age. This employment growth distribution is measured by comparing the average size of establishments for a given age cohort in 2005/2006 with the same cohort in 2009/2010. Specifically, I estimate age cohort coefficients by regressing employment growth on age cohort dummies for each country in 2005/2006. Similarly, I estimate the productivity growth distribution by age. Due to small sample size in the Enterprise Surveys and avoid unnecessary technical difficulties to calculate standard errors, I calculate the bootstrap estimate of standard errors of age cohort coefficients using the Wild Bootstrap technique. In addition, I estimate the asymmetric percentile bootstrap confidence intervals of age cohort coefficients. To estimate the bootstrap standard errors and the percentile confidence intervals of age cohort coefficients, I 10

12 use 999 bootstrap replications. Furthermore, I use HC 3 and Rademacher weights to calculate the bootstrapping value for each replication of the Wild bootstrapping (see Appendix B1). In order to validate the results of employment and productivity distributions by establishment age using the Enterprise Surveys, I use the Hsieh and Klenow (2012) estimates to compare. By calculating mean of each age cohort where the youngest cohort is normalized to one, I derive three age cohort estimates from their nine estimates. These three estimates are compared to the results of employment and productivity distributions using the Enterprise Surveys dataset. 3.1 Employment Distribution This section presents the employment distribution by establishment age using the crosssection data. First, I calculate the employment distribution by age in Mexico and compare these results with the Hsieh and Klenow (2012) estimates to validate the results. This allows me to develop a method to estimate the employment distribution by establishment age for other countries. Second, I apply this method for estimating the employment distribution by age to other surveyed countries of the Enterprise Surveys in 2005/2006 and in 2009/2010. Lastly, I summarize the results into low- and middle-income countries, by aggregating age cohort coefficients using the calculated per capital income weights. I consider the Hsieh and Klenow (2012) estimates in the U.S. as a base case scenario, with which to discuss age cohort coefficients of low- and middle-income countries. In addition, I plot age cohort coefficients against per capita income to justify that age cohort coefficients can be used as a proxy to measure of cross-country income differences Mexico To estimate the employment distribution by establishment age, I use the Enterprise Surveys dataset from 2006 and In order to deal with insufficient observations for each industry, Icombinealltwo-digitindustries,exceptforfood,garments,chemicals,non-metallicmineral products, and machinery and equipment, into an other category. This allows me to include 11

13 an independent regressor for each two-digit industry in GMM estimation. Since the sample in the Enterprise Surveys is stratified, I estimate GMM regressions using sampling weights to properly represent the population. To estimate the employment distribution by age, I aggregate age cohort coefficients using the calculated value added weights. Ireportagecohortcoefficientsofestablishmentsrelativetotheyoungestcohortcoefficient. The age cohort coefficient, the standard error, and the percentile interval of youngest cohort are respectively one, zero, and one. In 2006, as reported in Table 5, the coefficient of age cohort 2 (age 10-24) for Mexico is 1.40, which implies that establishments with the age cohort 2 have 1.4 times more employees as establishments with the age cohort 1 (less than 10 year). Similarly, in 2006, establishments with the age cohort 3 (age 25 or more) have 1.73 times more employees as establishments with the age cohort 1. As reported in Table 6, age cohort coefficients in 2010 are higher than those for Based on these age cohort coefficients in 2006 and 2010, it is clear that establishments with the age cohort 3 have more employees than establishments with the age cohort 2. Thus, the employment distribution by age suggests that the employment level increases as establishment ages. The results of employment distribution by age for Mexico are consistent with empirical prediction. The employment level of establishments increases as they age. This empirical prediction may arise from two main channels. First, age differences refer to as age channel, across establishments generate differences in their employment level. This employment level difference is due to establishment accumulating organization capital over their life cycle. Second, establishment adjustment to positive productivity shocks refer to as adjustment channel, across establishments generate differences in their employment level. Empirically, through both channels, the employment level of establishments will increase as they age. Specifically, the coefficient of age cohort 2 should be greater than one, as the coefficient of age cohort 1 is normalized to one. The coefficient age of cohort 3 should be greater than the coefficient of age cohort 2. This implies that establishments in the age cohort 2 should have more employees than those in the age cohort 1. Establishments in the age cohort 3 should have more employees than those in age cohort 1 and 2. Thus, it implies that employment level of establishments will increase as they age. The employment level of Mexican manufacturing establishments for a given age cohort 12

14 varies across establishments. In 2006, both standard errors of age cohort 2 coefficient and age cohort 3 coefficient are similar; however, in 2010, the standard error of age cohort 3 coefficient is higher than the standard error of age cohort 2 coefficient. The results are expected that standard error of older cohorts coefficients would be higher than younger cohorts coefficients. This implies that the percentile confidence intervals of age cohort coefficients for older cohorts would be wider than those for younger cohorts. In 2006, as reported in Table 5, the percentile bootstrap confidence intervals for age cohort 2 coefficient (0.93 to 2.02) is shorter than age cohort 3 coefficient (1.28 to 2.39). Similarly, in 2010, age cohort 2 coefficient has a smaller interval than age cohort 3 coefficient (Table 6). The higher standard errors and wider percentile confidence intervals of age cohort coefficients demonstrates that the employment level of manufacturing establishments increasingly varies as establishment ages. The cause of higher standard errors and wider percentile confidence intervals as cohorts age could be explained by misallocation of resources, and differences in entry and adoption of technology. Misallocation of resources could arise from age and adjustment channels that lead to increased variation of employment level across establishments. In addition, differences in entry and adoption of technology could lead to increased variation in employment level. Specifically, the employment level of establishments for a given age cohort varies by establishment because the productivity of establishments increases at different rates. As establishments age, more productive establishments could employ more employees than less productive. It is expected that older establishments have more variation in their employment level than younger establishments. Younger establishments have lower variation could be due to selection problem, only observe surviving establishments. Establishments choose to exit could be due to misallocation of resources and negative productivity shocks. Percentile confidence intervals should not include zero; therefore, all age cohort coefficients in Mexico are significant, because they do not include zero. In addition, if the lower limit of percentile interval is greater than one then the number of employees of a given age cohort would be significantly higher than to the youngest cohort. Based on the percentile confidence intervals, the results show that the number of employees of the older age cohort (age cohort 3) is significantly higher relative to the youngest cohort, with the exception of the age cohort 2 in Thus, the percentile intervals further establish the fact that employ- 13

15 ment level of manufacturing establishments increasingly varies, also with higher variation of employment levels by establishment, as establishment ages. To compare the estimates of the Enterprise Surveys with the Hsieh and Klenow (2012) estimates, I plot both age cohort coefficients in Figures 1.3 and 1.4. In addition, I plot the asymmetric percentile bootstrap confidence intervals of age cohort coefficients for predicting the distribution of age cohort coefficients. For 2006, the age cohort coefficients of establishments are consistent with the Hsieh and Klenow (2012) estimates in Mexico. For 2010, although the age cohort coefficients are not consistent with the Hsieh and Klenow (2012) estimates, my results are still comparable because the Hsieh and Klenow (2012) estimates fall within the percentile intervals of age cohort coefficients. The inconsistencies in the point estimates of age cohort coefficients could be due to sample selection bias. In addition, these inconsistencies could be caused by insufficient observations in each two-digit industry in the Enterprise Surveys. Because the Enterprise Surveys oversamples the large establishments, which are typically older, there are fewer observations in the youngest cohort. Since the results of age cohort coefficients in Mexico are consistent with the Hsieh and Klenow (2012) estimates, the approach used to estimate the employment distribution by establishment age for Mexico can be used as a method to determine the employment distribution for each surveyed countries of the Enterprise Surveys. In addition, it possible to generalize a systematic pattern for age cohort coefficients of establishments across countries Enterprise Surveys Countries Iapplysimilarapproachforestimatingtheemploymentdistributionbyestablishmentage in Mexico to the other surveyed countries in the Enterprise Surveys dataset, with minor exceptions. These exceptions stem from inconsistencies in the number of observations in each manufacturing industry in the Enterprise Surveys. The number of observations is varying between each year and countries in the Enterprise Surveys. As a result, the number of independent regressors in GMM estimation for each two-digit industry differs across countries and years. Specifically, the Enterprise Surveys have more establishments in some two-digit industries in 2005/2006 than in 2009/2010 and vice versa. If an industry does not have sufficient observations for including an independent regressor in GMM estimation, I include 14

16 that industry in the category other. In addition, I estimate GMM regression coefficients without sampling weights for 2005 dataset because the Enterprise Surveys does not provide sampling weights in that year. The estimated age cohort coefficients of establishments are reported in Tables 5 and 6. With few exceptions, the coefficient of age cohort 2 and cohort 3 in 2005/2006 and in 2009/2010 are greater than one. It implies establishments with age and 25 or more have more employees than establishments with age less than 10. Thus, the results of employment distribution by age of the Enterprise Surveys are consistent with empirical prediction. The standard errors and the asymmetric percentile bootstrap confidence intervals of age cohort coefficients are reported in Tables 5 and 6. Age cohort coefficients of establishments with age 25 or more have higher standard errors than establishments with age less than 10; establishments with age have higher standard errors than establishments with age 10 or less. Because of their larger standard errors, older establishments have a wider percentile interval of age cohort coefficients than younger establishments. Since the percentile confidence intervals do not included zero within intervals, age cohort coefficients are significant except for Bulgaria, Slovak Republic, and Slovenia in 2005/2006 and Botswana, and Slovak Republic in 2009/2010. These results suggest that employment level is not similar across establishments at a given age cohort. In addition, establishments with age 25 or more have significantly more employees than establishments with age less than 10 because the lower limit of percentile intervals for age cohort 3 is greater than one. By considering the lower limit of percentile intervals, establishments with age do not significantly differ from the youngest cohort. This could be because of the youngest age cohort length of the Enterprise Surveys is bigger than the Hsieh and Klenow (2012). The age cohort coefficients, the bootstrap estimate of standard errors, and the bootstrap percentile confidence intervals have a similar pattern across countries in the Enterprise Surveys, with few exceptions. These age cohort coefficients of the Enterprise Surveys are comparable with the Hsieh and Klenow (2012) estimates in India and Mexico. These results can be summarized using different characteristics of the countries in the Enterprise Surveys. Specifically, employment and productivity point estimates can be used in understanding cross-country income differences. 15

17 3.1.3 Low- and Middle-Income Countries To generalize employment distribution by establishment age in low- and middle-income countries, I use the estimated age cohort coefficients of establishments. Using the calculated per capital income weights and country s age cohort coefficients, I calculate the weighted average of age cohort coefficients for low- and middle-income country. In addition, I compare the results of the Enterprise Surveys with the Hsieh and Klenow (2012) estimates in the U.S. as a base case scenario. For calculating percentile confidence interval, I calculate weighted average of age cohort coefficients using country s age cohort coefficients for each bootstrap replication. As reported in Figures 1.6 and 1.7, in low-income countries, 25 year or older establishments have two and half times more employees than establishments with less than 10 year; establishments with age 25 or more in middle-income countries have three times more employees than establishments with age less than 10 year. It implies that older establishments in middle-income countries have more employees than low-income countries. In addition, age cohort coefficients of establishments of the Enterprise Surveys in low- and middle-income countries are lower than the U.S. Hsieh and Klenow (2012) estimates; U.S. establishments with age 25 or more have four times more employees than with age less than 10. Furthermore, the percentile confidence intervals include the Hsieh and Klenow (2012) estimates for the U. S., however, this results drive with few countries. Specifically, establishments with age 25 or more in around 22 countries out of 29 countries have employment levels far below with same cohort in U.S. establishments (Tables 5 and 6). This low employment of older establishments in low-income countries could be explained income differences between lowand middle-income countries. Thus, the Hsieh and Klenow (2012) results of the employment distribution over the life cycle for India and Mexico hold more broadly in countries of the Enterprise Surveys dataset. In order to justify the argument that age cohort coefficients can be used as a measure of cross-country income differences, I plot age cohort coefficients for age 25 or more against per capita income for both year 2005/2006 and 2009/2010 in Figure 1.8. It shows that the age cohort coefficient for age 25 or more is positively correlated with per capita income. 16

18 This implies that age cohort coefficients can be used as a proxy to measure of cross-country income differences. 3.2 Employment and Productivity Growth This section presents employment and productivity growth distributions by establishment age using the two-year panel observations. Using the revenue productivity instead of the physical productivity due to lack of data in the Enterprise Surveys, I estimate the productivity distribution by establishment age. According to Hsieh and Klenow (2012), the revenue productivity of 35 year or older establishments in Mexico is roughly the same as that of new establishments; whereas employment and physical productivity grow at the same rate. Similar predictions are also hold in U.S. manufacturing establishments. Theoretically, they also derive the life cycle predictions 8 of establishments employment and productivity distributions. Using the Longitudinal Research Database (LRD) of establishments, Bartelsman and Dhrymes (1998) find no systematic relation between the revenue productivity of an establishment and its age in U.S. manufacturing sector. Thus, it implies that all age cohorts coefficients relative to youngest cohort for revenue productivity would be close to one; however, employment and physical productivity could grow at the same rate. With limited data in the Enterprise Surveys, I follow two steps for calculating the employment and productivity growth distributions by establishment age. First, I calculate the growth rate of establishments. Second, I estimate three age cohort coefficients by regressing the growth rate (employment and productivity) of establishments on age cohort dummies of 2005/2006 using sampling weights. In addition, I report age cohort coefficients relative to the youngest cohort coefficient as the coefficient of youngest cohort is normalized to one. Since the Enterprise Surveys does not provide the sampling weights of 2005, I use sampling 8 Hsieh and Klenow (2012) derive the equilibrium employment of establishments, if the the ratio of capital and labor distortions does not vary across establishments, is proportional to: L ai Aσ 1 ai τai σ Where τ ai ( PaiYai K ai ) α ( PaiYai L ai ) 1 α is revenue productivity. Where i indexes the establishment, a refers to the establishment s age, Y ai is the value added of plant i of age a and σ>1istheelasticity of substitution between varieties. An establishment s employment is increasing in its physical productivity, A ai and decreasing in its revenue productivity,τ ai. 17

19 weights of 2009 and 2010 in GMM estimation. Due to insufficient panel observations, I estimate the age cohort coefficients of establishments by estimating a GMM regression on age cohort dummies of 2005/2006 for each country. In addition, I drop observations whose value added is negative or zero. The employment and productivity growth distributions by establishment age are reported in Tables 7 and 8. In Mexico, the age cohort 3 coefficient of employment growth is 1.63, which implies that employment growth for establishments with age 25 or more have 1.63 times more employment growth than establishments with age less than 10. The results are similar to the employment distribution estimates using cross-section data in terms of employment level. Furthermore, the revenue productivity for the age cohort 3 (0.90) is roughly the same as youngest age cohort in Mexico. For comparing the estimates of the Enterprise Surveys dataset for Mexico with the Hsieh and Klenow (2012) estimates, the employment and productivity growth distributions by establishment age are reported in figures 1.9 and The age cohort coefficients of employment and productivity growth are consistent with the Hsieh and Klenow (2012) estimates. Since employment and physical productivity growth rate could be same, establishments with age 25 year or more have physical productivity growth 1.63 times more than establishments with less than 10 year. A major problem with the age cohort coefficients for Mexico is that the standard errors and the percentile confidence interval of the estimates are quite large. Because the percentile intervals include zero, age cohort coefficients are insignificant. Furthermore, the lower values of percentile confidence intervals are less than one. Using similar approach of Mexico to estimate the employment and productivity growth distributions, I estimate the employment and productivity growth distributions for other countries in the Enterprise Surveys dataset, as reported in Tables 7 and 8. Some age cohort coefficients are negative, implying that the employment level of establishments decreases as they age. This could be explained by the global recession of Due to insufficient observations, the bootstrap standard errors and the percentile confidence intervals of age cohort coefficients, using employment and revenue productivity growth distributions, are quite large. This could be due to misallocation of resources across countries and differences in entry and adoption of technology. Because the percentile intervals include zero, the age 18

20 cohort coefficients are insignificant. The employment growth of establishments grows over establishment ages with larger variation at older cohorts than younger cohorts. Thus, the Hsieh and Klenow (2012) results for India and Mexico do not hold more broadly in low- and middle-income countries. However, this could be because of the global recession. Specifically, the period of the dataset in the Enterprise Surveys for employment and productivity growth distributions was in the global recession. An important question remains; what exactly are the barriers facing manufacturing establishments with lower employment and productivity distributions by establishment age. Based on the bootstrap estimate of standard errors and the asymmetric bootstrap percentile confidence intervals, older establishments have higher variation in employment and productivity. This variation of employment and productivity could be due to misallocation of resources (intensive margin) and entry and adoption of technology (extensive margin). This variation in establishments as establishment ages suggests manufacturing establishments in low- and middle-income countries face multitude barriers. Hsieh and Klenow (2012) provide suggestive evidence for these barriers like as bigger contractual frictions in hiring non-family labor, higher tax enforcement on larger firms, financial frictions, difficult in buying land or obtaining skilled managers, and costs of shipping to distance markets. 19

21 4 Conclusions Employment and productivity distributions by establishment age play an important role in understanding the difference in aggregate TFP across countries. Generally, establishments grow with age as they invest in new technologies, develop new markets, and produce a wider array of higher quality products. This evolution differs by country; however, there is no study to estimate this evolution pattern more broadly in a larger set of countries. In order to establish this pattern, I estimated employment and productivity distributions by age using the dataset from the Enterprise Surveys. To check a general consistency for Mexico estimates, I compared the Enterprise Surveys estimates with Hsieh and Klenow (2012) estimates. I establish a fact that the employment distribution over the life cycle estimates of establishments, with few exceptions across countries, demonstrates a systematic pattern. In low-income countries, establishments with age 25 year or older have two and half times the number of employees than the establishments with age less than 10; when marking same comparison in middle-income countries, the establishments have three times larger employees. In particular, the relative size of older establishments is much smaller in developing countries than in developed countries. My estimates can be used to explain cross-country income differences. In roughly 22 countries out of 29 countries, establishments with age 25 or more have the employment level far below with same cohort in U.S. establishments. This implies that Hsieh and Klenow (2012) results for India and Mexico hold more broadly in a larger set of countries. The results further showed that the age cohort coefficient for age 25 or more is positively correlated with per capita income. This further confirms that age cohort coefficients can be used as a proxy to measure of income differences across countries. An important question remains; what are the causes of lower employment and productivity for older establishments in developing countries. The results showed that the older establishments have higher standard errors and wider confidence intervals. These variations suggest establishments in developing countries face multitude barriers that caused lower employment and productivity paths over the life cycle. Potential areas of future research would be to identify the causal effect of the barriers on establishment dynamics. 20

22 Table 1: The number of observations in the Enterprise Surveys (2005/2009) Status of Country Country Before Cleaning After Cleaning Panel * Panel * Low-income Kyrgyzstan Armenia Lower middle-income Moldova Moldova Bosnia and Herzegovina Bulgaria Upper middle-income Macedonia, FYR Poland Romania Russian Federation Slovenia High income Slovenia Republic Slovenia * - Panel shows the number of panel observations in 2005 and Notes: Some countries are dropped from 2005/2009 dataset because of insufficient observations to calculate weighted average of age cohort coefficients. Table 2: The number of observations in the Enterprise Surveys (2006/2010) Status of Country Country Before Cleaning After Cleaning Panel * Panel * Low-income DRC Angola Bolivia Lower middle-income Ecuador El Salvador Guatemala Paraguay Argentina Botswana Chile Colombia Upper middle-income Mexico Panama Peru Uruguay Venezuela * - Panel shows the number of panel observations in 2006 and 2010.! 21

23 Table 3: Mean and standard deviation of establishments employment and age in 2009/2010 Status of Country Low-income Middle-income Country Employment Age Mean Std. Dev. Mean Std. Dev. Angola Armenia Azerbaijan Bolivia DRC Ecuador El Salvador Guatemala Kyrgyzstan Moldova Paraguay Argentina Bosnia and Herzegovina Botswana Bulgaria Chile Colombia Macedonia, FYR Mexico Panama Peru Poland Romania Russian Federation Serbia Slovak Republic Slovenia Uruguay Venezuela