Misallocation, Aggregate Productivity and Policy Constraints: Cross-country. Evidence in Manufacturing

Transcription

1 Misallocation, Aggregate Productivity and Policy Constraints: Cross-country Evidence in Manufacturing Addisu A. Lashitew University of Groningen, P.O. Box 800, Nettelbosje 2, 9747 AE Groningen The Netherlands. March 2012

2 ABSTRACT Cross-country differences in GDP per capita are to a large extent driven by differences in total factor productivity (TFP). Recent studies suggest that low-income countries have low levels of aggregate TFP because of distorting policies that induce resource misallocation. This study uses a unique, international, firm-level dataset to measure the effect of misallocation on aggregate TFP, and to explain the role of policy constraints therein. The results show that hypothetically removing misallocation leads to considerably higher aggregate TFP. Misallocation is also substantially higher in countries where policy constraints such as firing costs and barriers to external competition are higher. JEL Classification: D24, O40, L60 Key Words: Aggregate TFP, Productivity, Misallocation, Policy constraints

3 INTRODUCTION The most enduring conclusion of decades of research in the development accounting literature is that TFP contributes at least as much as factor inputs in explaining income per capita. Most reviews of the literature contend that TFP accounts for 50% - 70% of per capita income differences across countries (Hsieh and Klenow, 2010; Caselli, 2005). With the assumption that TFP measures underlying production technology, the standard explanation for cross-country income differences then becomes the existence of barriers to technology adoption in less developed countries (Parente and Prescott, 1994). A relatively new strand of literature, on the other hand, emphasizes that resource misallocation across heterogeneous firms could be an important determinant of aggregate productivity. Since firms in the same industry often exhibit very large productivity differences (Syverson, 2011; Foster et al., 2001), the way resources are allocated among them can have a substantial effect on aggregate productivity. A large number of empirical studies find that input reallocation towards more productive firms contributes significantly to aggregate productivity growth (Eslava et al., 2004; Scarpetta et al., 2002; Foster et al., 2001). Theoretical studies in this literature try to establish the mechanism through which reallocation affects aggregate productivity. An influential study by Melitz (2003) used a dynamic industry model with heterogeneous firms to show the reallocative benefits of trade. In this model, trade openness increases aggregate TFP by facilitating the exit of unproductive firms thus reallocating inputs to more productive plants. Building on this, several studies have tried to measure the magnitude of misallocation and the relevance of policy constraints therein (Restuccia and Rogerson, 2008; Alfaro et al., 2008; Bartelsman et al., 2009). 1

4 Hsieh and Klenow (2009) extend Melitz's (2003) monopolistic competition model to measure the effect of resource misallocation on aggregate TFP. They find that idiosyncratic price distortions cause sizeable gaps in marginal products of labor and capital across plants within four- digit manufacturing industries in China and India. Their result shows that TFP is significantly higher when distortions are hypothetically removed so that the marginal products of labor and capital are equalized. If China and India could improve the efficiency of factor allocation in their manufacturing industries to the level of the United States, they could increase their manufacturing TFP by 30% and 60% respectively (or by 87% and 128% in absolute terms). They propose that public ownership of firms in China and licensing restrictions in India could be mechanisms that foster misallocation. Although TFP appears to be the most important proximate determinant of per capita income, policy and institutional factors are the ultimate determinants of income through their effect on the level of productivity. Among others, Hall and Jones (1999) conclude that policy and institutional factors such as rule of law and openness to trade play an important role in explaining cross-country productivity differences. However, little been done to examine how policy and institutional factors affect misallocation, and hence TFP and income. The broader literature on productivity analysis has so far been exclusively confined to explaining TFP growth with limited effort to separately identify the drivers of misallocation. The more recent reallocation literature is also limited to establishing the theoretical link between policy constraints and aggregate TFP. Empirical evidence on the relative importance of policy differences across countries on misallocation is hence lacking. Moreover, the effort to measure the importance of misallocation has been so far been limited to a small number of countries, partly because of lack of comparable firm-level data. 2

5 The aim of this study is to address these gaps in the literature. The paper measures the magnitude of misallocation using the World Bank s Enterprises Survey dataset, which is unique for its coverage of more than 20,000 manufacturing firms in 77 countries. The first contribution of this paper to the literature is measuring the magnitude of misallocation across a large number of countries using a common approach. Secondly, the paper contributes to the literature by examining the role of policy constraints in factor and product markets on misallocation. The large country coverage of the WBES dataset and the presence of substantial policy heterogeneity across the countries it covers give us an excellent opportunity to link misallocation with policy factors. The paper tries to provide robust measures of misallocation by utilizing two different approaches of measurement. Firstly, I measure misallocation using the monopolistic competition framework proposed by Hsieh and Klenow (2009). In this model, distortions in input and output prices create wedges between the marginal products of capital and labor for firms in the same industry, thus lowering aggregate TFP. This framework allows us to calculate the hypothetical, efficient TFP that could be realized by fully removing distortions so that marginal products are equalized within industries. The gap between actual TFP and the hypothetical, efficient TFP measures the degree of misallocation that is induced by price distortions. This indicator of TFP gap will be our first measure of misallocation. The second approach used for measuring misallocation in this paper is based on more general measures of dispersion. Specifically, I use the standard deviation and interquartile range of total factor productivity and the marginal products of labor and capital within industries. The dispersion of productivity will be higher in economies where allocative efficiency is low, for example, because of low competitive pressure that allows less productive firms to survive. 3

6 Productivity dispersion has been, therefore, used as an indicator of allocative efficiency in a number of studies (Ito and Lechevalier, 2009; Arnold et al., 2008). Similarly, higher dispersion of the marginal products of labor and capital is indicative of low allocative efficiency. An important advantage of dispersion-based measures of misallocation is that, compared to our first approach, they are based on fewer assumptions and are easier to measure and interpret. The results show that misallocation has a significant negative effect on aggregate productivity. On average, aggregate TFP could be increased by around 120% if distortions are removed so that marginal products are equalized within industries. Hence output in the average country could be more than doubled for the same amount of inputs if resources were most efficiently allocated. This suggests that misallocation is an important determinant of aggregate TFP and thus income. The monopolistic competition framework provided by Hsieh and Klenow (2009) also allows us to separately measure the effects of two types of distortions on aggregate TFP. Output distortions are taxes (subsidies) on production that make the size of the firm in a given industry to be smaller (larger) than what is optimal. Input distortions, on the other hand, affect the input mix decision of the firm so that it deviates from the optimal input mix. Firms facing high distortion on the price of capital (relative to labor) use less capital (again relative to labor) than what would be optimal, and vice versa. Among the two types of distortions, the results show that output distortions that affect the size of the firm constitute the largest part of total misallocation. There is also substantial difference in the level of misallocation among countries; the interquartile range of the TFP gap across countries is close to two. To shed light on the driving forces behind these differences, I systematically examine the relative importance of different policy constraints that have been suggested to affect misallocation in the literature. Specifically, I 4

7 look into four groups of policy constraints in factor and product markets: barriers to external competition, entry and exit barriers, financial frictions, and labor and capital market distortions. The relatively large country coverage of the dataset and the policy heterogeneity across the countries it covers proves useful for investigating the effect of policy variables on misallocation in cross-country regression framework. The results reveal that the cost of firing, barriers to trade and foreign investment, and high cost of entry all have significant positive effect on misallocation. High cost of firing has a particularly adverse effect on misallocation; lowering the cost of firing from the 75 th to the 25 th percentile in the sample can lead to TFP gain of 20%. I also find that openness for foreign investment is likely to improve allocative efficiency only in countries where R&D spending and educational attainment are sufficiently high. This is in line with previous findings that FDI improves domestic productivity only when a minimum threshold of absorptive capacity is reached (Wang and Wong, 2009). These results are robust for alternative ways of measuring policy constraints. The rest of the paper is organized as follows. Section I provides the theoretical framework for measuring misallocation by sketching the monopolistic competition model of Hsieh and Klenow (2009). Section II deals with data and measurement issues, and provides descriptive results for our measures of misallocation. Section III describes how policy constraints can affect misallocation and provides results for the regression analysis in which policy constraints are linked to misallocation. Section IV provides a robustness test for the regression analysis and section V concludes. 5

8 I. THEORETICAL FRAMEWORK Several recent studies try to establish theoretically the mechanism through which reallocation affects aggregate productivity. Some of these studies focus on measuring the magnitude of misallocation by introducing a general type of distortion. Hsieh and Klenow (2009) is a prime example of these types of studies. Similarly, Restuccia and Rogerson (2008) use a growth model with heterogeneous firms to show that distortions can substantially lower aggregate TFP. Alfaro et al. (2008) use a monopolistic competition model and find that resource misallocation across firms plays an important role in explaining cross-country income differences. Bartelsman et al. (2009) report large differences across countries in the relationship between firm productivity and size which is indicative of differences in allocative efficiency. A number of other studies use growth models to calibrate the effect of specific policy distortions on misallocation. Buera et al. (2011), Buera and Shin (2010), and Erosa and Cabrillana (2008) show that financial frictions lower aggregate TFP. Barseghyan and DiCecio (2009) show how entry costs induce higher misallocation, and Fang (2010) examines the effect of financial frictions and entry costs on TFP levels. Lagos (2006) studies the effect of labor market frictions on aggregate TFP. This section will introduce the theoretical framework for measuring misallocation which is based on the monopolistic competition model of Hsieh and Klenow (2009). This model allows us to measure misallocation using the gap between actual TFP and the hypothetical, efficient TFP that could be realized if distortions were absent. Since the full description of the model can be found in the original article, I only provide a brief sketch of the model below. At the highest level of aggregation, final output Y is produced by combining output from manufacturing industries Ys using Cobb-Douglas production technology: 6

9 (1) Y = S θ S ( Ys ), s= 1 where θs is the value added share of sector s, and S is the total number of manufacturing industries in the economy. Industry output Y S is a CES aggregate of M S differentiated products: (2) Ms Y s = i = 1 Y σ 1 σ si σ σ 1, where Y si is a differentiated product by firm i in industry s, and σ is the elasticity of substitution. Each differentiated product is produced by firms with heterogeneous productivity (A) using labor (L) and capital stock (K) with Cobb-Douglas technology: (3) Y si = A K si αs 1 α S si Lsi. The main feature of the model is that firms are not only heterogeneous with respective to their productivity as in Melitz (2003), but they also face idiosyncratic distortions on their input and output prices. Two types of distortions are introduced: output distortions that affect quantity of production while leaving the input mix unaffected, and capital distortions that affect the use of capital relative to labor. Output distortion is modeled as a tax on production that is independent of factor use because it distorts the marginal products of capital and labor in equal proportions. Capital distortion, on the other hand, is a form of tax on capital that affects the input mix decision. Since capital distortion is calculated relative to labor, it can also be understood as labor distortion (i.e. a positive capital distortion implies tax on capital or subsidy to labor). Note that both distortions are exogenous and are implied from the data as discussed below. In this framework, profits depend not only on prices and quantities, but also on distortions: (4) π si = Psi (1 τ Ysi) Ysi wlsi (1 + τ Ksi ) RKsi, 7

10 where w is the wage rate, R is the rental price of capital, τ ysi is output distortion and τ ksi is capital distortion. Profit maximization leads to the standard condition that the firm s output price is a fixed markup over its marginal cost: (5) P si = [ σ ( σ 1) ] w ( 1 α ) 1 α s α s α s [ ] [ R α ] [(1 + τ ) (1 τ ) A ], s s Ksi Ysi si where the term (σ/ σ-1) is the markup. In addition to factor prices, both output and capital distortions appear in the price equation with a positive effect. The marginal revenue product of labor (MRPL) and the marginal revenue product of capital (MRPK) are given by the respective partial derivatives of the revenue function multiplied by the inverse of the markup to correct for rents: [ 1 α ][ ( σ 1) σ ][ P Y L ] = w ( 1 τ ). (6) MRPLsi s si si si Ysi [( σ 1) σ ][ P Y K ] = [ R ( 1+ τ ) ( 1 τ )]. (7) MRPK si α s si si si Ksi Ysi Equations 6 and 7 show that the marginal revenue products of labor and capital are determined not only by the wage rate and the rental price of capital but also by distortions. Capital distortions raise only the marginal revenue of capital whereas output distortions raise both the marginal revenue product of labor and capital. In other words, output distortions lead to smaller size by reducing the use of both labor and capital, whereas capital distortion decreases the use of capital relative to labor. All differences in marginal revenue products are thus attributed to exogenous distortions. To link the two measures of distortion with aggregate productivity, it is important to note the distinction between the total factor productivity of revenue (TFPR) and total factor productivity of quantity (TFPQ). TFPQ is a measure of total factor productivity after accounting 8

11 for firm-level price differences, whereas TFPR is a measure of productivity that is not separated from price (i.e. TFPR = TFPQ * P). (8) TFPQ si A si = Y si K α S 1 α S si Lsi. (9) TFPR si P si A si = P Y si si K α S 1 α S si Lsi. Equation 9 defines TFPR as a product of prices and physical productivity (TFPQ). By substituting the value of prices given by Equation 5 in Equation 9, TFPR can be expressed as a function of distortions and factor prices. Since all distortions are reflected in factor marginal products, TFPR can also be alternatively expressed as a function of the marginal revenue products of capital and labor: (10) TFPR si = = [ σ ( σ 1) ] w ( 1 α ) 1 α s α s α s [ ] [ R α ] [(1 + τ ) (1 τ )] s 1 α s 1 α s [ MRPL si ( 1 α )] [ MRPK ( α )]. s si s s Equation 10 shows that all differences in TFPR within an industry are caused by output and capital distortions. Note that no physical productivity (TFPQ) term features in the equation, and thus TFPR has no relationship with productivity. Although firms with high physical productivity (TFPQ) have high revenue productivity by definition (Equation 9), they also charge lower prices since they are cost efficient (Equation 5). Thus the positive effect of TFPQ on TFPR is offset by its negative effect on prices. This relationship allows us to use TFPR as a term that captures the effects of both types of distortions. Similarly, industry-level revenue productivity TFPR S can be shown to be a function of distortions: Ksi Ysi (11) TFPR s Ms = 1 s s Ms [ σ ( σ 1) ] w ( 1 α s ) ( 1 τ Ysi ) k si R α s {( 1 τ Ysi ) ( 1 + τ Ksi )} k si, α α where the weighting term k si is the output share of firm i in industry s. 9

12 Aggregate productivity at industry level is given by the following equation: (12) Ms TFP s = i= 1 σ 1 σ 1 ( A TFPR TFPR ), si s s i 1 where A si is physical productivity (i.e. TFPQ), and TFPR s is the industry-level revenue productivity. To calculate the potential TFP gain that can be realized by removing distortions, we should first derive the hypothetical, efficient TFP. If all firms face zero output and capital distortion, firm-level TFPR (Equation 10) will be the same across firms in the same industry. The industry average TFPR (Equation 11) will be equal to the firm-level TFPR, and thus identical among all firms. In this case, the industry-level TFP (equation 12) will depend only on firm productivity A si, giving us the efficient TFP: (13) Ms σ 1 TFP = Efficient S A si i= 1 1 σ 1. For each industry, TFP gap is calculated as the ratio of efficient TFP (Equation 13) to actual TFP (Equation 12). Efficient (14) TFPGap S = TFPS TFPS. The TFP gap, calculated as the ratio of efficient to actual TFP, defines the level of misallocation. Hsieh and Klenow (2009) show that the TFP gap is a function of the variances and covariance of output and capital distortions. The TFP gap given by Equation 14 will thus be the basis of our first measure of industry-level misallocation. To get a country-level measure of misallocation, I aggregate the industry-level TFP gap using the Cobb-Douglas aggregator given by Equation 1: (15) TFPGap = M S S = 1 M S θ θ S Efficient S ( TFPGap ) = ( TFP TFP ). s S = 1 S S 10

13 For each country in the dataset, the TFP gap is calculated separately using the ratios of efficient to actual TFP given by equations 14 and 15. To simplify interpretation, I convert the ratios into percentage terms by deducting one and then multiplying them by 100. The two measures of TFP gap, expressed in percentage terms, are thus our first measures of misallocation at industry and country level. The TFP gap can be interpreted as the extent to which actual TFP is lower than the efficient TFP due to misallocation, or as the potential TFP gain that can be realized by removing distortions. II. DATA AND MEASUREMENT Data sources The main data source for this analysis is the World Bank s Enterprises Survey (WBES) dataset. The WBES is an ongoing survey that collects firm-level data worldwide. The major advantage of the WBES survey is that data collection is conducted systematically using standardized survey instruments. The dataset thus provides comparable data that is unique for its extensive country coverage. Sampling for the WBES is conducted using stratified sampling procedure to ensure representation. First, the number industry groups to be covered across each major sector (services, manufacturing and non-agriculture primary activities) is determined. For manufacturing, industry grouping is based on 2-digit ISIC classification. The number of industry groups to be covered in each country is determined according to the size of the total economy which is taken as a proxy for the universe of firms. Once the number of industries is decided, industry groups that contribute relatively more to the total economy in terms of total production or employment are selected. In the second stage, a sampling equation is used to determine a 11

14 representative sample size per industry group. Finally, further stratification is made based on firm-size and geographical location to select the firms that are covered by the survey. 1 Data collection started in 2002 and different countries have been covered in subsequent years. Panel data is available for some countries; however, the country coverage of the panel dataset is limited. The analysis in this paper is performed using a cross-section dataset that combines data from different years for all countries to maximize total available data. When multiple years of data are available for a country, the cross-section dataset contains only the year for which the largest number of observations are available. I started compiling the cross-section data by removing repeated (panel) observations, nonmanufacturing firms, and observations with missing or incomplete data. Then I checked for the presence of outliers using the share of labor in value added and the capital-value added ratio. I removed observations below the 1 st and above the 99 th percentiles of the two ratios to reduce bias from measurement error. After cleaning, the cross-section dataset has 20,205 manufacturing firms in 77 countries. The number of observations per country, the year of data collection, and summary statistics for other key variables is provided by Table A1 in the appendix. There is a large difference in sample size across countries; whereas large countries such as India, Brazil and China have well above a thousand observations, smaller ones such as Albania and Armenia have only around 30 observations. The dataset covers mostly low- and middle-income countries with an average per capita GDP of USD 5,550 in The country with the highest income is Czech Republic (USD 20,362) and the one with the lowest income is Congo Democratic Republic (USD 266), all in constant PPP prices. 12

15 I use data on total production, cost of intermediate inputs, capital stock and labor inputs for measuring productivity. Market value of production is not available for most firms, and so the more widely available data of total sales is used. Value added is measured as the difference between sales and the cost of intermediate inputs. Cost of intermediate inputs is calculated by adding up three major cost categories: energy consumption (fuel, electricity and other energy costs), cost of raw materials and overhead and other expenses. To account for differences in hours worked and human capital, labor input is measured using labor cost rather than employment. Measuring misallocation The TFP gaps from equations 14 and 15, which measure the ratio of efficient TFP to actual TFP at industry and country level, will be our first measures of misallocation. Large TFP gap indicates higher level of misallocation since it means that actual TFP is much lower than its potential due to distortions. This section discusses the measurement of the TFP gap and presents some descriptive results. A number of parameters are required to calculate the TFP gap given by equations 14 & 15. First, we need to specify values for the wage rate and the rental price of capital in order to measure the marginal products of labor and capital. For every country in the dataset, I set the wage rate to the average value of the observed wage rate among firms within the country. For all countries, I set the rental price of capital R to 0.10, assuming a real interest rate of 5% and a depreciation rate of 5% as in Hsieh and Klenow (2009). It is noteworthy that incorrectly measuring the wage rate and the rental price of capital does not affect our measures of TFP gap. 13

16 This is because the error will be reflected in the marginal products of labor and capital of all firms, thus affecting the distortion of all firms in equal amount. Second, we need to assign a value for the elasticity of substitution (σ) among products. As can be seen from equations 12 and 13, the elasticity of substitution affects the level of actual and efficient industry TFP and hence the TFP gap. Again, I follow Hsieh and Klenow (2009) and choose an elasticity parameter of three. Although this parameter is chosen on the basis of the relevant micro literature (Hsieh and Klenow, 2009), I also test the sensitivity of the TFP gap for higher level of substitution between products by setting σ to five. Finally, benchmark values of the output elasticity of capital and labor are required in order to measure distortions. It is necessary to apply similar industry-specific elasticity parameters for all countries in our dataset in order to get comparable measures of distortions and TFP gap. These benchmark elasticity parameters should come from data that are not distorted and thus reflect the true characteristics of each industry s technology. Since it is likely that our data is tainted by distortions (which is the reason why we want to measure misallocation), we cannot uncover the elasticity parameters from our own data. Again following the precedence of Hsieh and Klenow (2009), I use elasticity parameters from the US as benchmark values since input and output prices from the US are relatively less distorted. For each industry in our data, the output elasticity of labor (1-α s ) is calculated using the value added share of labor in the respective US industry. Using constant returns to scale assumption, the output elasticity of capital is measured as 1 minus the labor share of the industry. Table A2 in the appendix gives the elasticity parameters used for our analysis and the data source. Once the above-mentioned parameters are determined, output and input distortions can be easily calculated using derivations from the model. Using the definition of MRPL given by 14

17 Equation 6, the output distortion of a firm is measured as the gap between its labor share (multiplied by the markup to adjust for rents) and the labor share of a representative US firm in the same industry: (16) τ Ysi σ wl si / Psi Y = 1 1 σ 1 α s si. If a firm faces high MRPL, this will show as low labor share in value added for a given wage rate. This lowers the ratio of the firm s labor share to the labor share of the representative US firm, resulting in a high output distortion. Again using the definitions of MRPL and MRPK given in equations 6 and 7, the capital distortion is imputed from the gap between the firm s capital-labor ratio and the capital-labor ratio of the US industry-representative firm: α s wl si (17) τ = 1 1 Ksi α s RK si. The implication here is that if a firm has lower capital-labor ratio compared to the US industry benchmark, it is facing higher capital distortion. Using Cobb-Douglas technogy assumed in Equation 3, firm-level productivity is meaured as follows: ( P Y ) σ σ 1 si si (18) Asi = αs 1 K L si α S si. Physical productivity (TFPQ) has to be imputed using the above equation because, as most other datasets, our dataset only contains revenue data but not quantity and price data. In the absence of quantity and price data, we impute physical productivity by deriving quantities from revenues using a demand function that establishes the relationship between quantity and prices. 15

18 The exponent in the numerator of Equation 18 is the derviation of the elasticity parameter that is used to convert revenues to quantities. Once productivity and distortions are calculated, we are able to measure misallocation. First, the TFP gap values given by equations 14 & 15 are calculated for each country by removing both output and capital distortions. 2 These measures of TFP gap indicate overall misallocation since they are measured by comparing the actual TFP with the efficient TFP that could be realized if all distortions are removed within industries. To get more insight on the relative importance of output and capital distortions, two additional measures of TFP gap are calculated by removing the two types of distortion turn by turn. The second measure of TFP gap is calculated using the level of efficient industry TFP (Equation 13) in which only output distortions are removed. In other words, instead of allowing marginal products of both labor and capital to be equalized within industries as in the first case, only the marginal product of labor is equalized for calculating the efficient industry TFP. The second measure of TFP gap thus indicates misallocation caused by size-relate distortions. The third measure of TFP gap is calculated using the efficient industry TFP that can be achieved when only capital distortions are removed. In other words, our last measure of TFP gap indicates misallocation exclusively due to capital distortions that affect the optimal input mix. Figure 1 presents the average values of the three measures of TFP gap across all countries. The TFP gap calculated by removing both distortions is high, with an average value of 120%. Since the TFP gap between efficient and actual TFP s is expressed as a percentage of actual TFP, this value indicates that the efficient TFP is 120% higher than the actual TFP. Thus, output could be more than doubled using the same amount of inputs if misallocation across plants within industries is fully removed. 16

19 This apparently large TFP gap reflects the presence of substantial heterogeneity among firms with respect to their productivity and marginal products. Since some level of heterogeneity is inevitable even in relatively efficient markets such as the US (Hsieh and Klenow, 2009; Syverson 2011, ), the TFP gap should not be strictly interpreted as measuring the TFP loss caused by misallocation. The TFP gap values are more meaningful as measures of misallocation when compared across countries than in absolute terms, since they are calculated using a benchmark which is not necessarily achievable. 3 Figure 1 also reveals that the TFP gap due to output distortions is much larger than the TFP gap due to capital distortions. Thus distortions that affect firms ability to grow are more important than those affecting their input mix decisions. Hsieh and Klenow (2009) also find that the larger part of the gap between actual and efficient TFP is the result of size-related distortions rather than capital distortions. Note that the TFP gaps caused by each type of distortion does not sum up to the TFP gap in which both distortions are removed because the distortions have a positive correlation with each other. --- [Figure 1 about here] --- Table A1 in the appendix presents the three measures of TFP gap for all countries in the dataset. The gap between actual and efficient TFP, expressed as a percentage of actual TFP, varies widely across countries, with a standard deviation of 66. The five countries with the highest level of TFP gap in our dataset are Ethiopia (303%), Turkey (286%), Egypt (266%), Botswana (266%) and Peru (257%), whereas the five countries with the smallest level of TFP gap are Guinea (12%), Chile (15%), Czech Republic (19%), Kyrgyzstan (20%), and Burkina Faso (23%). The 25 th and 75 th percentiles of the TFP gap are, respectively, 80% (in Morocco), and 146% (in Honduras), implying a large interquartile range of

20 The last three columns of Table A1 also present the three measures of TFP gap under a higher level of elasticity of substitution (σ = 5). The results reveal that the TFP gap now appears much larger, indicating that our measures of misallocation depend heavily on assumptions regarding the elasticity of substitution [Figure 2 about here] --- Figure 2 compares the TFP gap for China and India according to our own calculation using the WBES dataset and the TFP gap according to the original calculation by Hsieh and Klenow (2009). Although the differences are not substantial, the original calculations by Hsieh and Klenow (2009) appear smaller than our calculations, especially for China. This is likely because Hsieh and Klenow (2009) use 4-digit level industry classification whereas the WBES dataset allows for at best 2-digit level classification. Thus our TFP gap calculations based on the WBES dataset can capture structural differences among industries along with misallocation. Figure A1 in the appendix also reports comparison of TFP gap across different years for countries for which panel data is available. With the exception of a few countries, the measures of TFP gap do not substantially vary over the years, indicating that misallocation is persistent. The apparent similarity of the TFP gap indicators over time for most countries also suggests that they are consistent measures of misallocation. Measures of misallocation: correlation results The previous sub-section has discussed the measurement of the first group of indicators of misallocation as well as providing descriptive results. As indicated earlier, I also use alternative measures of misallocation as robustness check. To be specific, the standard deviation and interquartile range of the log-transformed values of the revenue productivity (TFPR), and the 18

21 marginal revenue products of labor (MRPL) and capital (MRPK) are used as additional indicators of misallocation. Table 1 provides correlation results between our three measures of TFP gap, the standard deviations of TFPR, MRPL, MRPK, and GDP per capita. Firstly, it can be noted that our first measure of TFP gap in which both types of distortions are removed is more strongly correlated with the TFP gap in which only output distortions are removed (coefficient = 0.8), than with the TFP gap in which only capital distortions are removed (coefficient = 0.5). This highlights that output distortions contribute disproportionately more to overall misallocation compared to capital distortions. --- [Table 1 about here] --- Secondly, our first measure of TFP gap and the dispersion of revenue productivity have a strong correlation of 0.6. This shows that a simple measure misallocation such as the dispersion of revenue productivity can capture a large part of total misallocation. Finally, Table 1 also reveals that GDP per capita has negative correlation with most measures of misallocation. Although the significance of the correlation coefficients is mixed, their negative sign suggests that misallocation contributes to lowering per capita income. In the following sections, I will explore which policy factors increase misallocation, hence lowering per capita income. III. MISALLOCATION AND POLICY CONSTRAINTS Measuring the magnitude of misallocation goes halfway to understanding its role on aggregate TFP. Shedding light on the sources of misallocation, however, requires a systematic examination of the relative importance of different potential policy constraints. The results presented so far indicate that misallocation reduces aggregate TFP substantially. Correlation 19

22 results also show that misallocation is higher in low-income countries, suggesting that it contributes towards lowering per capita income. Moreover, there are large differences in the magnitude of misallocation across countries, the interquartile range of the TFP gap across countries being close to 2. The large dispersion of misallocation possibly reflects differences in the policy environment that determines allocative efficiency. The goal of this section is to identify policy constraints that have been associated with misallocation in previous studies, and to conduct regression analysis in order to test their effect on misallocation. OLS regressions of economic outcomes on institutional and policy variables are often subject to endogeneity problems. One major source of endogeneity stems from the fact that institutional design could be responsive to economic outcomes, thus leading to serious reverse causation problems (Hall and Jones, 1999). Fortunately, this issue is not of major concern for our analysis since our outcome variable, misallocation, is measured using micro data, and thus is not directly observable to policy makers. Measuring policy constraints The list of policy and institutional factors that are likely to affect the efficiency of resource allocation across firms is long (Arnold et al., 2008). In general, policy constraints can affect allocative efficiency in two different ways. Some policy constraints reduce competitive pressure by lowering the entry of new firms, thus reducing the possibility of reallocation of inputs from inefficient incumbents to more productive new-entrants. Others induce misallocation by protecting inefficient existing plants (such as public firms) so that inputs are not reallocated towards more productive incumbents (Dollar and Wei, 2007). 20

23 Policies that affect allocative efficiency are also likely to affect technical efficiency. For example, increased competitive pressure not only facilitates efficient allocation of inputs across producers, but also pushes producers to use resources more efficiently and/or to adopt more efficient technologies. The effect of most policy variables on aggregate productivity is hence twofold; directly they determine the level of technical efficiency of producers, and indirectly they influence the allocation of inputs across producers. This sub-section reviews four groups of policy constraints that can affect allocative efficiency, and discusses their measurement. i. Openness to external competition. Exposure to external competition because of openness to trade and foreign direct investment can enhance allocative efficiency. As highlighted by the seminal work of Melitz (2003), trade openness intensifies competition and increases aggregate productivity by allowing more productive firms to expand and the least efficient firms to exit. There is an extensive empirical literature supporting the reallocative effect of trade. Among others, Bernard et al. (2006) show that low-productivity plants are more likely to die in industries with falling trade costs, thus contributing to higher productivity growth. Eslava et al. (2004) report that trade openness contributed to higher productivity in Colombia by facilitating reallocation of inputs from low- towards high-productivity businesses. Increased FDI has practically the same effect on local firms which face more competition following the entrance of multinationals (Crespo and Fontoura, 2007). Openness to trade and FDI can also have a direct effect on the technical efficiency of firms. Facing more intense competition, domestic firms are forced to use their existing resources more efficiently. In addition, FDI and trade openness expose domestic firms to new technologies, allowing them to improve their productivity through spillovers and imitation (Crespo and Fontoura, 2007). 21

24 In this paper, trade openness is measured using the level of trade intensity, calculated as imports plus exports expressed as a percentage of GDP. Although this measure is likely to understate trade in relatively large countries which rely less on inter-country trade and more on within-country trade, it is nonetheless the best available measure of openness to trade. Openness to foreign investment is measured as net FDI inflows as a percentage of GDP, which includes initial equity capital to acquire at least 10% stake in domestic enterprises as well as reinvestments of earnings. Both measures are taken from the World Development Indicators database, and are averaged over the years to reduce year-to-year fluctuations. As shown in the descriptive statistics given by Table 2, the average level of trade openness in our sample stands at 82.5% of GDP and the average net FDI inflow is 4.3% of GDP. --- [Table 2 about here] --- ii. Level of domestic competition: entry and exit barriers. Regulatory constraints can exacerbate misallocation by hindering competition among existing firms or by discouraging the entry of new firms. Market regulations, often imposed to address public interest issues such as externalities and monopoly, can end up hampering competition and reducing allocative efficiency (Djankov et al., 2002). Examples of regulatory barriers that reduce the entry of new firms include high registration costs, licensing restrictions, and government ownership of firms. Several studies have found a negative relationship between entry barriers and aggregate TFP. Barseghyan (2008) finds that entry costs have a negative effect on TFP and GDP per capita, and Barseghyan and DiCecio (2009) show how high entry costs can induce misallocation. Fang (2010) shows that financial frictions amplify the negative effect of entry costs on TFP. Scarpetta et al. (2002) find that entry regulations influence productivity by affecting entry. Klapper et al. (2006) report that costly regulation in European countries has reduced the rate of new firm entry 22

25 and forced new entrants to be larger. Arnold et al. (2011) find that product market regulations that curb competitive pressures tend to reduce the productivity performance of firms. Fisman and Allende (2010) find that the level of entry regulations affect factor allocation within industries by determining whether growth opportunities are utilized by new entrants or existing firms. In this paper, I use two indicators related to regulatory constraints to measure entry and exit barriers. The first indicator measures the level of entry barriers with the cost of starting a new business, expressed as a percentage of GDP. Originally developed by Djankov et al. (2002), this indicator includes all official fees, and fees for legal or professional services that are required for starting up a new business. This measure of entry barriers has been used in previous studies (see Barseghyan, 2008). The second indicator measures exit barriers using the level efficiency of insolvency proceedings. This indicator was initially developed by Djankov et al. (2008) and measures the percentage of financial assets that can be recovered by creditors from their total claims upon the closure of a business due to bankruptcy. The indicator is inversely proportional to the cost and length of insolvency proceedings, and has been found to be a robust measure of the efficiency of bankruptcy laws (Djankov et al., 2008). 5 By encouraging investors to use legal proceedings, efficient bankruptcy laws allow bankrupt firms to exit the market or reorganize at lower cost, hence reducing exit barriers and improving allocative efficiency. Both measures of entry and exit barriers are taken from the World Bank s Doing Business Indicators database and are averaged over the years to reduce year-to-year fluctuations. The average level of entry cost in our sample is around 85% of GDP, which is almost twice as large as the average value Djankov et al. (2002) report for their sample of 85 countries. Democratic Republic of Congo has the highest level of entry cost in our sample (1201%), whereas Lithuania has the lowest level of entry cost (3.36%). With respect to exit 23

26 barriers, the average rate of recovery during bankruptcy proceedings is 25%. This is almost half as large as the average recovery rate of 52% Djankov et al. (2008) find in their data, suggesting that countries in our sample have relatively less efficient insolvency proceedings. Since this measure has a significant correlation of 0.47 with GDP per capita, the low efficiency of insolvency proceedings in our data reflects the fact that our sample is dominated by low-income countries. Mexico has the most efficient insolvency proceeding in our sample, with 64% recovery rate. Rwanda, Laos, Burundi and Cape Verde have the least efficient debt enforcement with zero recovery rate. iii. Financial frictions. Financial frictions are among the most widely studied determinants of factor allocation (Arnold et al., 2008). The level of financial development in a market can affect allocative efficiency in two ways (Buera and Shin, 2010). Firstly, efficient financial markets lower the cost of financial capital by pooling risks and providing efficient intermediation. Lower cost of capital can boost the entry of new firms, thus intensifying competition and forcing inefficient incumbents to exit. Secondly, well-developed financial markets are more capable at identifying profitable firms and reallocating capital towards them. The allocation of capital based on market forces will drive inefficient firms out of the market, including publicly-owned firms that would thrive due to preferential credit access from public banks. The resulting reallocation of capital towards more efficient firms will improve allocative efficiency and boost aggregate TFP. Several influential studies find that improvements in financial access enhance the entry of new firms which are potentially more productive (Buera et al., 2011; Rajan and Zingales, 1998; Aghion et al., 2007; Greenwood et al., 2010). To measure financial frictions, I use an indicator of financial development. Specifically, I use the size of the financial sector measured as credit extended for the private sector as a 24

27 percentage of GDP. Since financially developed markets with lower financial frictions in general have a large financial intermediary sector, this measure is used in the literature as an indicator of financial development (Levine, 2005). The average level of private sector credit as a percentage of GDP in our data, averaged over , is 33%. Congo Democratic Republic has the lowest level of financial development (2%) whereas South Africa has the highest level of financial sector development (138%). iv. Labor and capital market distortions. In addition to financial frictions, regulatory constraints related to labor and capital markets can also distort factor allocation. Hopenhayn and Rogerson (1993) show that labor market frictions such as firing costs hinder the creation of new jobs. Lagos (2006) shows that labor market distortions can lower TFP. Eslava et al. (2004) find that labor market reform contributed to the reallocation of inputs towards more productive firms in Colombia. Bassanini et al. (2009) report that firing regulations have a negative effect on productivity growth. Likewise, a number of studies have looked into possible distortions in capital markets due to corporate taxes. High corporate tax rates can distort investment in the economy and affect the re-investment decisions of multinationals (Mooij and Ederveen, 2008). Economic theory and empirical evidence also reveal that, when capital is perfectly mobile, the incidence of the corporate tax is almost entirely borne by labor (Nicodème, 2008). In this paper, I consider two indicators related to labor and capital market regulations. I use an indicator of the cost of firing to measure regulatory constraints that induce labor markets frictions. First developed by Botero et al. (2004), this measure includes various costs of terminating redundant workers such as advance notice requirements, severance payments, and other penalties. The cost of firing is expressed in terms of weekly wages so as to make it comparable across countries. Secondly, I use the profit tax rate as an indicator of capital market 25

28 distortions. Both factors, firing costs and the profit rate, affect the cost of labor relative to capital and hence can induce misallocation by forcing firms to make suboptimal input mix decisions. Data for both measures is taken from the World Development Indicators database. The cost of firing is averaged over the years , but the profit rate for the year 2010 is used since data is not available for other years. Firing workers costs an amount equivalent to 56 weeks of wage on average, which varies from 4 weeks in Oman and Jordan to 192 weeks in Sri Lanka. The profit tax rate averages around 17%, and ranges from zero in Liberia, Lithuania and Bolivia to almost 60% in Congo Democratic Republic. Regressions results This section presents regression results that link the policy constraints discussed above with our measures of misallocation. Table 3 provides six regression results in which the dependent variables are the three measures of TFP gap that are discussed in the measurement section. The TFP gap indicators used in the first three regressions measure misallocation at country level and are calculated based on Equation 15. In regression 1, the dependent variable is the TFP gap in which both output and capital distortions are removed. Regression 2 is based on the measure of TFP gap in which only output distortions are removed, and regression 3 is based on the measure of TFP gap in which only capital distortions are removed. These partial measures of misallocation capture the separate effect of output and capital distortions on aggregate TFP. The dependent variables in the last three regressions are industry-level measures of TFP gap that are calculated using Equation 14. Although all policy constraints are measured at countrylevel, their effect on misallocation could vary across industries. It is thus important to test if the effect of policy constraints is the same when misallocation is measured at industry level and 26