Growing like Spain:

Transcription

1 Growing like Spain: Manuel García-Santana Université Libre de Bruxelles (ECARES) Enrique Moral-Benito Banco de España Josep Pijoan-Mas CEMFI and CEPR Roberto Ramos Banco de España May 4, 2015 Abstract Spanish GDP grew at an average rate of 3.5% per year during the expansion of , above the EU average of 2.2%. However, this growth was based on factor accumulation rather than productivity gains. In particular, TFP fell at an annual rate of 0.7%, while it increased at 0.4% in the EU and 0.7% in the US. Why did Spain fail to benefit from the growth of the technological frontier? We argue that deterioration in the allocative efficiency of productive factors across firms is at the root of the low TFP growth in Spain. Using administrative data of firms we show that within-sector misallocation of production factors increased substantially over the period in all industries, with most of the effects coming from inefficient capital and labor mix rather than inefficient size. We find that absent such deterioration, average TFP growth would have been around 0.8% per year, in line with the growth of the technological frontier. Finally, we provide empirical evidence that differences in the influence of the public sector across industries is a potential source of this deterioration. In contrast, sectoral differences in skill intensity, innovative content, or financial dependence are unrelated to changes in allocative efficiency. We also document that young and small firms were the most affected. JEL Codes: D24, O11, O47. Keywords: TFP, Misallocation, Spain. The opinions and analyses are the responsibility of the authors and, therefore, do not necessarily coincide with those of the Banco de España or the Eurosystem. We thank John Fernald for sharing the financial intensity data with us, and Eric Bartelsman for helpful discussion. We also thank seminar participants at Banco de España for useful comments.

2 1 Introduction The expansion was the longest in Spanish history (see Berge and Jordá (2013)). GDP grew at an average 3.5% per year, which compares very favourably to the EU average of 2.2% over the same period. 1 However, Spanish growth during this expansion was based on factor accumulation rather than productivity gains. In particular, annual TFP growth was -0.7%, which is low in comparison to other developed economies such as the US or EU. Such a dismal performance of productivity growth is surprising for a country that is so well integrated in a trade and monetary union with some of the World technology leaders. Did Spain fail to keep up with the technological frontier? In this paper, we argue that the source of negative TFP growth has been the increase in the within-sector misallocation of production factors across firms. We use a large administrative data set of Spanish firms to compute several measures of allocative efficiency for every year between 1995 and In particular, we compute the potential TFP gains due to factor reallocation as Hsieh and Klenow (2009) and the Olley and Pakes (1996) covariances. All measures show a severe deterioration of allocative efficiency over the period. Furthermore, we find the phenomenon to be present in all sectors of activity, which casts doubt on the widespread view that specialization in low productivity sectors such as construction was the main force behind Spanish low TFP growth. We thus argue that allocative efficiency of resources across firms is at the root of the low rates of TFP growth observed in Spain. Our results are very stark: had the level of within-sector allocative efficiency remained constant, TFP growth would have been around 0.8% per year. Therefore, our conclusion is that Spain did not fail to keep up with the technological frontier. Aggregate productivity stagnated because the economy increasingly allocated capital and labor in the wrong place across firms within each industry. 2 The deterioration of factor allocation across firms during an economic expansion is arguably a singular experience in Spain. Bartelsman, Haltiwanger, and Scarpetta (2013) find that misallocation remained roughly constant over the nineties and early 2000s in several developed countries such as US, UK, Germany or the Netherlands, while it clearly fell for the transitional economies of Central and Eastern Europe. Lewrick, Mohler, and Weder (2014) find that improvements in the within-industry allocation of resources is one of the main drivers of aggregate TFP growth in Swiss manufacturing. Using the Hsieh and Klenow (2009) framework, Bellone and Mallen-Pisano (2013) find that misallocation remained constant between 1998 and 2005 in France, while Dias, Robalo, and 1 EU average refers to the EU15 group, which includes Austria, Belgium, Denmark, Finland, France, Germany, Greece, Ireland, Italy, Luxembourg, the Netherlands, Portugal, Spain, Sweden and the United Kingdom. We take this reference group of developed countries similar to Spain because we have comparable growth accounting data from EU-KLEMS. 2 The prominent role of within industry misallocation as a hindrance to TFP growth has relevant policy implications. For instance, reallocation of workers across industries is generally more costly than reallocation within industries (see e.g. Shin (1997)). 1

3 Richmond (2014) show that misallocation increased in Portugal between 1996 and 2011, but this was a period of stagnation in Portugal. Finally, Chen and Irarrazabal (forthcoming) find that the Chilean economy experienced a substantial decline in misallocation over its expansion. In order to shed some light on the potential sources of this phenomenon in Spain, we evaluate the relationship between several sector-specific characteristics and the changes in allocative efficiency. In particular, we find that industries in which the influence of the public sector is larger (e.g. through licensing or regulations) experienced significantly larger increases in misallocation. In contrast, other sectoral characteristics such as skill intensity, innovative content or financial dependence are unrelated to changes in allocative efficiency. Turning to firm-specific distortions, we find that small and young firms in Spain might have faced higher market distortions than large and mature firms. As a result, these firms grew less than optimal and operated with capital-labor ratios smaller than optimal (i.e. those observed in the same industries in the US). Finally, we show that the increase in misallocation across firms is present in all Spanish regions, and that regional differences in average wage growth are uncorrelated with the increase in distortions. It remains to be understood why the Spanish economy accumulated capital and labor at such a fast pace despite the negative increase in aggregate productivity. Our view is that this was due to exogenous supply factors. The convergence process leading to the entry to the EMU reduced interest rates dramatically and created an unprecedented credit boom. As shown by Fernandez- Villaverde, Garicano, and Santos (2013), the total credit to GDP ratio tripled between 1994 to Along these lines, Gopinath, Kalemli-Ozcan, Karabarbounis, and Villegas-Sanchez (2015) rationalize low rates of TFP growth in South Europe by developing a model of heterogeneous firms that can generate misallocation of capital across firms as a result of financial frictions and investment adjustment costs. An alternative view of this supply side explanation is given by Díaz and Franjo (2014). These authors show that the large increase in capital accumulation over the period was largely due to capital structures, which they interpret as the result of government subsidies. The process of capital accumulation triggered by cheap credit could by itself increase the employment rate. However, there were also clear labor supply factors at play: the working-age population ratio increased over the period and females of new cohorts participated in the labor market at a much larger rate than females of the older cohorts. The rest of the article is organized as follows. Section 2 briefly shows the growth accounting results for Spain as well as some micro-based evidence motivating the paper. Section 3 describes the data. Then, Section 4 presents the main results regarding the increase in misallocation and Section 5 explores the potential sources of misallocation. Some concluding remarks are provided in Section 6. 3 These authors also argue that the credit boom might be the very reason behind an increase in misallocation of productive factors across firms. Banks face a signal-extraction problem to identify good firms. In bubble times the signal becomes more noisy and hence credit may be allocated less efficiently. This would be consistent with our finding that the increase in distortions was larger among young firms. However, we should also expect the increase in misallocation to be larger in sectors with higher financial dependence, a pattern that is absent in the firm-level data. 2

4 Finally, Appendix A reviews the theoretical framework of Hsieh and Klenow (2009) and Appendix B contains additional results. 2 The growth experience The Spanish economy grew by around 3.5% per year between 1995 and This expansion, the longest in the twentieth century, helped Spanish income per capita surpass the EU average in the early 2000s. Growth accounting exercises show that the boom was driven by factor accumulation (labor and capital) rather than increases in productivity. Using data from EU-KLEMS, Figure 1 clearly illustrates this pattern. Figure 1: The Spanish growth experience Macro evidence Production Labor Capital TFP The labor contribution to output (total hours worked) expanded 3.8 percent a year in This was the result of three main factors: a fast growing working age population, mainly due to migration flows, and an increasing labor force participation rate, mainly reflecting the incorporation of women into the labor market, and a decline of the unemployment rate since the high values achieved in The capital stock also grew at an unprecedented pace of 5.2 percent a year. The rise of the construction sector together with easy borrowing conditions played an important role in the expansion of the capital stock in Spain. Since both labor and capital grew more than final production, total factor productivity (TFP) was reduced by 0.7% per year. 3

5 These Spanish figures are in sharp contrast to other developed economies. In the average EU country, output growth was 2.2% per year with growth rates of 1.1% and 3.3% for labor and capital respectively. As a result, TFP growth in the EU was on average 0.4% per year in contrast to the Spanish annual rate of -0.7%. This difference is even more pronounced with respect to the US economy, which experienced TFP growth rates around 0.7% per year over the period. 4 Figure 2: The Spanish growth experience Micro evidence Manufacture of butter Manufacture of toys Joinery installation Change in share Change in share Change in share Relative TFP in Relative TFP in Relative TFP in 2001 Sale of textiles Retail sale of bread Retail sale of telecom Change in share Change in share Change in share Relative TFP in Relative TFP in Relative TFP in 2001 t statistics for industry specific regressions Notes: Relative TFP refers to the logarithm of firm-specific TFP relative to the industry average, ln(tfp i /TFP). Change in share refers to the difference in firm-specific market share measured in terms of sales. Turning to the micro-based evidence, a first glimpse at the data cleary illustrates the deterioration in the within-industry allocation of resources across firms. In the upper panel of Figure 2 we plot the change in market shares over the period against the level of TFP in 2001 for six selected 4-digit industries. 5 In all the six cases the relationship is negative, which means that firms with initial TFP below the industry average gained market share at the expense of firms with larger 4 See EU-KLEMS dataset at 5 For illustrative purposes we focus on the period to maximize the number of observations in the scatter plots. This is so because we use a balanced version of our panel dataset in order to compute changes in shares. 4

6 TFP. This negative relationship is negative and statistically significant for 80 per cent of the 356 industries considered as we can see in the bottom panel of Figure 2, which plots the distribution of the t-statistics resulting from the 356 sector-specific regressions. We interpret this evidence as an intuitive illustration of how more productive firms lost market share at the expense of less productive firms, which, in our view, is a clear indication of a deterioration in the allocation of resources across firms within each industry. However, in the remainder of the paper we focus on well-known indicators of allocative efficiency that facilitate the mapping between the micro and the macro evidence. 3 Data We use a firm-level dataset containing information of a representative sample of non-financial companies in Spain from 1995 to The sample contains an average number of 497,782 firms per year. This database is named Central Balance Sheet Data (Central de Balances) and is provided by the Banco de España. The database is comprised of two complementary datasets. The first one is based on a standardized voluntary survey handled to companies at the time of requesting compulsory accounting information. Each year around 9,000 companies fill this survey. The information gathered is very detailed, but the sample size is low and big firms are over represented. The second dataset contains the balance sheets of a much larger number of companies. It originates from the firms legal obligation to deposit their balance sheets on the Mercantile Registry. Therefore, coverage is much wider. The Bank of Spain Central Balance Sheet Office is in charge of collecting and cleaning these datasets. All of the variables contained in the second database are also included in the first one. For each firm, we observe its value added, total wage bill, employment, book value of the capital stock (both physical and intangible) and sector of activity at the 4-digit level (according to NACE rev. 2 classification). Since most of the variables are recorded in nominal terms, we employ sector-specific deflators for capital and value added, to compute real values with 2000 as base year. 6 Panel A of Table 1 illustrates the size distribution of firms in our raw sample for the year The table also compares this distribution with that obtained from the Central Business Register available from the National Statistics Institute. There are two important aspects to highlight. First, the coverage of our raw sample is remarkably large in terms of both the number of firms (56% of the operating firms in Spain) and the level of employment (54% of total employment). Second, our sample provides an excellent representation of the firm size distribution in Spain. In particular, small firms (less than 10 employees) account for 83.90% of the total number of firms and 20.47% of the employment in our sample versus 83.07% and 20.23% in the population. At the other extreme, large firms (more than 200 employees) represent less than 0.5% of the total number of firms both in our 6 The capital deflator is collected from Mas, Pérez, and Uriel (2013) and the value added deflator is taken from the National Accounts. Both deflators are constructed at the 2-digit NACE classification. 5

7 sample and in the population, while they account for 33.47% of the employment in our sample and 32.13% in the population. Table 1: Size distribution of firms in our sample and in the census. Central Balance Sheet Dataset Central Business Register Firms Labor Firms Labor Number of employees Total (#) Share (%) Total (#) Share (%) Total (#) Share (%) Total (#) Share (%) PANEL A: Raw Sample , , , ,718, , , , ,050, , , , ,400, , , , ,596, , ,540, , ,728, All 485, ,601, , ,494, PANEL B: Final Sample , , , ,718, , , , ,050, , , , ,400, , , , ,596, , ,528, , ,728, All 327, ,536, , ,494, Notes: Figures refer to the year Self-employed persons are not included. From this original sample we drop observations with missing or non-positive values for the number of employees, value added, or capital stock. We also eliminate observations at the top and bottom 1% of these variables. Since our misallocation measures are computed within each 4-digit industry, we also drop firms belonging to industries with less than 10 firms per year. Therefore, we are left with around 350,000 firms per year distributed in digit industries. Turning to this final sample in Panel B of Table 1, our screening strategy slightly over-samples larger firms because small firms with less than 10 employees are more prone to misreport their information in the Mercantile Registries. Note also that our final sample does not include firms with 0 employees because these firms represent mostly firms with no production, created merely for tax purposes. In any event, since those firms account for a small fraction of employment, the representativeness of our final sample in terms of employment remains noticeably good. 7 It is our view that the availability of information on small firms is crucial for measuring within industry misallocation at the 4-digit level as opposed to typical 7 The Amadeus database, commercialised by Bureau Van Dyck, also provides firm-level information extracted from firms balance-sheets on a set of variables for all European OECD members including Spain. For instance, Hsieh and Klenow (2014) exploited this dataset. However, Amadeus presents some well-known drawbacks. First, information on employment (typically a non-mandatory item in balance sheets) is only available for 40-50% of the firms in the sample (this implies that although listed in terms of identifier in the Amadeus data, 50-60% of the firms do not provide information about employment). Second, the large attrition bias generates a the lack of representativeness in terms of size, European Central Bank (2014). Third, the readily usable version of the Amadeus data currently starts in the year

8 datasets used in the literature that are restricted to samples of larger firms (e.g. with more than 10 or 20 employees). Indeed, using only large firms in our sample, the estimated increase in misallocation is two times smaller than that obtained from our full sample. 4 Misallocation and productivity in the Spanish boom Our main finding is illustrated in Figure 3. Applying the methodology of Hsieh and Klenow (2009) to our sample of Spanish firms, we find that potential TFP gains from reallocation steadily increased over the period. While TFP could have been around 24% higher in 1995, this figure doubled by Between 1995 and 2007, the allocative efficiency decreased by 20 percent (1.49/1.24-1), or a reduction of about 1.7 percent per year. These hypothetical increases in the level of aggregate TFP would result from fully equalizing TFPR across firms in each 4-digit sector, i.e., from reallocating resources from firms with low physical TFP towards firms with high TFP. As acknowledged by Hsieh and Klenow (2009), these counterfactuals do not allow for measurement error or model misspecification, which may cast doubt on the usefulness of these numbers without a reference point to compare. However, we do not focus on the level but on the change of potential TFP gains relative to the year Our implicit assumption is that neither measurement error nor model misspecification have changed over time. In any event, changes in dispersion of TFPR might arise from other frictions apart from idiosyncratic distortions of the type embedded in the Hsieh and Klenow (2009) theoretical framework. For instance, overhead labor or quasi-fixed capital (see Bartelsman, Haltiwanger, and Scarpetta (2013)). Based on Olley and Pakes (1996), we thus explore two covariances as alternative measures of misallocation. First, we compute a covariance term between firm-specific labor shares and labor productivity; and second we compute the covariance between firm-specific production shares and total factor productivity. Under an efficient allocation of resources, more productive firms should produce more and hire more workers. 9 Table 2 summarizes the different measures of misallocation for two sub-periods, namely, and In particular, we use the Hsieh and Klenow (2009) measure of potential TFP 8 Crespo and Segura-Cayuela (2014) consider a sample of French, Italian, German, and Spaninsh firms in order to compare TFP gains resulting from the HK methodology in selected years. According to their results, TFP gains in Spain are larger than those in France but smaller than those in Germany and Italy. Moreover, they also find an increase in Spanish TFP gains between the years 2002 and To be more precise, for industry j and year t, the covariance statistic for labor productivity (LPR) is given by: OP j,t = i (θ ij,t θ j,t )(ω ij,t ω j,t ) where i is the firm index, θ ij,t refers to the firm-specific labor share, ω ij,t is the firm-specific labor productivity, and θ j,t and ω j,t are the unweighted averages of industry j. The same covariance can be computed for TFP using firm-specific production shares as originally considered by Olley and Pakes (1996). 7

9 Figure 3: Potential TFP gains from reallocation gains (labeled as HK) together with the standard deviation of log TFP (labeled as STD), which measures the dispersion of (log) TFPR at the firm level as an alternative measure of misallocation in the HK framework. 10 We also report the two covariance terms as used by Bartelsman, Haltiwanger, and Scarpetta (2013) labeled as OP, one based on labor productivity (LPR) and labor shares, and the other based on total factor productivity (TFP) and production shares. All the statistics in Panel A of Table 2 clearly point to an increase in the degree of misallocation in the Spanish economy over the last expansion as documented in Figure 3. While the dispersion in TFP increased from 0.42 to 0.47, the OP covariance of LPR and labor shares was reduced from 0.30 to 0.21, and the covariance between TFP and production shares moved from 1.59 to This finding is present not only for the aggregate economy but also for the main four the sectors of the economy as shown in Panel B of Table Depending on the misallocation measure considered, the sector with the largest misallocation deterioration is either construction or services. However, the four measures of misallocation point to a deterioration in allocative efficiency in the four sectors. Moreover, Table A1 in Appendix B reports the corresponding results for disaggregated sectors at NACE rev 2-2 digits showing that this deterioration is prevalent among virtually all of the 58 2-digit sectors considered. This fall in the allocative efficiency of production factors is distinctive of the Spanish growth 10 Under joint log normality of A si, 1 τ Ysi, and 1 + τ Ksi, both measures are equivalent. 11 Note that the sector-specific results are based on misallocation within 4-digit industries aggregated using industry weights. 8

10 experience. In particular, Bartelsman, Haltiwanger, and Scarpetta (2013) find that misallocation remained roughly constant over the nineties and early 2000s in several developed countries such as US, UK, Germany or the Netherlands, while it clearly fell for the transitional economies of Central and Eastern Europe. Table 2: Misallocation in Spain over the period PANEL A: Total Economy HK STD TFP OP LPR OP TFP PANEL B: By sector HK STD TFP OP LPR OP TFP Manufacturing Construction Trade Services Notes: HK refers to the potential TFP gains if resources were allocated efficiently as proposed by Hsieh and Klenow (2009). OP refers to the Olley and Pakes (1996) covariance term. STD refers to standard deviation as a measure of dispersion. LPR refers to log labor productivity and TFP to log total factor productivity. We argue that the stark increase in within-sector misallocation over the Spanish boom is at the root of the bad performance of aggregate TFP. We next compute potential TFP growth under the assumption that the level of misallocation remains constant at its 1995 level. This counterfactual exercise provides the aggregate TFP that we would have observed during the expansion without increases in misallocation. To be more precise, we simply multiply the observed aggregate TFP by the year-specific percentage of TFP gains given by the HK exercise above (see Figure 3). Then, we plot the resulting potential TFP growth rates together with the observed ones in Figure 4. Annual growth rates of potential TFP growth would have been between 0.6% and 1.1% with an average of 0.8% under the assumption of constant within-sector misallocation. In contrast, observed TFP growth was -0.7% on average, ranging from -0.8% to -0.5%. We can also use the OP methodology to compute the potential TFP growth. Had the OP covariance term remained constant for the period, TFP would have grown at a 1.1% annual rate 12 which is similar to the counterfactual TFP growth 12 This number can be computed by calculating the percentage change in the OP covariance term between and as reported in Panel A of table 2, which is 24%, and obtaining the corresponding average annual rate over the 12-year span. 9

11 computed using HK methodology. Finally, we also compute an alternative counterfactual TFP growth based on aggregating sectorspecific TFP growth rates with weights given by the sector shares in This exercise aims to illustrate the role of between sector misallocation in the evolution of aggregate TFP. While we see that this counterfactual TFP growth is higher than the observed one, it is substantially smaller than the counterfactual based on constant within-sector misallocation. More specifically, it ranges from -0.8% to -0.1% with an average annual growth of -0.4% All in all, these counterfactual exercises speak in favor of the crucial role of within-sector misallocation in the evolution of aggregate TFP in Spain over the last expansion. This finding casts doubt on the traditional view that between-sector misallocation (i.e. specialization in low productivity sectors such as construction) was behind low TFP growth in Spain. Figure 4: Potential TFP growth under 1995 misallocation level Observed TFP growth Potential TFP growth constant misallocation Potential TFP growth constant sector shares 4.1 Robustness analysis In this Section we perform three robustness exercises related to the level of industry disaggregation, to the distinction between intensive and extensive margin only, and to the elasticity of substitution Industry classification Our baseline results are based on misallocation within 4-digit industries because the HK theoretical framework relies on the assumption that each industry represents a monopolistic competitive market 10

12 in which firms produce different varieties of the same intermediate good. Therefore, the greater the level of disaggregation the more plausible this assumption is when taken to the data. However, since the 4-digit level of disaggregation requires very large samples of firms to obtain meaningful figures for more than five hundred industries, we investigate if the deterioration in allocative efficiency documented for Spain at the 4-digit level is also present when considering 2- and 3-digit classifications. Table 3 shows the evolution of allocative inefficiency in Spain in terms of potential TFP gains from reallocation to an efficient allocation of resources across firms within each 4-, 3-, and 2-digit sectors in columns (1), (2) and (3). The increase in TFP gains, or the deterioration in allocative efficiency, is prevalent among all the three industry classifications. Moreover, the increases over the whole period are of the same magnitude, around 20% or 1.7% per year, in all the cases. In particular, the average increases are 1.7, 1.6, and 1.8 percent per year for the exercises at 4-, 3-, and 2-digit industries, respectively Balanced versus unbalanced panel Our baseline sample is an unbalanced panel including firms that might enter or exit at any time. The extensive margin may also play a role in shaping the evolution of allocative efficiency depicted above. However, the potential sources of misallocation might be different depending on the importance of this extensive margin relative to the intensive margin of misallocation of resources within established firms. In order to quantify the importance of the extensive margin in terms of efficient TFP and the evolution of allocative efficiency over time, we consider a balanced panel restricted to firms that were in the sample for the whole period ( ). In the balanced version of the panel we have only 5,419 firms per year, which precludes us from considering misallocation within 4-digit industries. Column (4) in Table 3 shows the resulting TFP gains from the balanced panel under the 2-digit disaggregation. We find that the deterioration in allocative efficiency over time still holds, although smaller in size: while the increase in misallocation is around 20% (1.49/1.24-1) or 1.7% per year when considering the unbalanced panel, the corresponding figures are 7% (1.28/1.20-1) and 0.6% under the balanced panel. These numbers suggest that about two thirds of the deterioration in allocative efficiency is due to the extensive margin Elasticity of substitution As as final robustness test we repeat the exercise with a higher elasticity of substitution: σ=5. This figure is also used by Dias, Robalo, and Richmond (2014) for Portugal and comes from the estimates for the Eurozone in Christopoulou and Vermeulen (2012). In column (5) of Table 3 we report the results. As expected, TFP gains increase for all years when σ=5. Moreover, the magnitude of the increase in misallocation over the period is similar to that of the case σ=3, a decrease of 21% (1.69/1.39-1) or 1.8 percent per year. 11

13 4.1.4 Measurement error Our estimated increases in TFP gains from reallocation might be driven by an increase in measurement error in our data as a result of the year-to-year increases in our sample size. While this concern is partially addressed in the balanced panel exercise, we also consider an alternative robustness check based on recording errors created by extreme outliers. In particular, following Hsieh and Klenow (2009) we trim the 2% tails of TFPR and TFPQ in order to avoid the potentially increasing influence of outliers in our sample. Column (6) of Table 3 shows the resulting TFP gains, which clearly point to a large deterioration in allocative efficiency of around 14% (1.42/1.25-1) or 1.1 percent per year Sample of large firms We now check the sensitivity of our findings to the size distribution of firms in our sample. In particular, we computed TFP gains resulting from removing idiosyncratic distortions in a subsample of large firms (more than 50 employees). We report the results in Column (7) of Table 3. While the deterioration in allocative efficiency still arises, its magnitude is substantially smaller, 0.6% percent per year against the baseline of 1.7% per year. Moreover, the levels of TFP gains are substantially smaller than those of the full sample in the baseline case. Our interpretation of this result is that datasets of large firms typically used in the literature might under-estimate the magnitude of withinindustry misallocation. 12

14 Table 3: Misallocation in Spain over the period Robustness analysis TFP gain from reallocation Baseline 3-digit 2-digit Balanced σ = 5 M. error Large firms (1) (2) (3) (4) (5) (6) (7) Notes: Baseline in column (1) refers to our benchmark results based on misallocation within 4-digit industries, σ=3, and the unbalanced panel. Columns (2) and (3) report the results when considering indutries at 3- and 2-digit classifications (NACE 2 rev. 2). Column (4) is based on the balanced version of our panel. Column (5) reports the TFP gains when considering σ=5 instead of σ=3. Column (6) refers to the trimming of the 2% tails of TFPR and TFPQ in order to alleviate the influence of measurement error. Finally, column (7) is based on a sample of large firms (more than 50 employees). 5 Sources of misallocation s evolution In the previous sections we have uncovered a large decrease in within-industry allocative efficiency, which we have shown to be the main source of low TFP growth observed over the period in Spain. Given this finding, the challenge is to identify the economic forces that led to the increase in the misallocation of production factors across firms. In this section, we make a first step by providing some descriptive evidence together with some tentative interpretations regarding the potential sources of the Spanish increase in misallocation over the last expansion. In particular, we estimate the relationship between different sector- and firm-specific characteristics with the observed sector and firm-specific changes in allocative efficiency. 5.1 Size versus capital distortions In this section we explore the firm-specific measures of distortions implied by the Hsieh and Klenow (2009) exercise of Appendix A. In particular, we have two measures of distortions, one labeled as τ K that distorts the capital-labor ratio, and one labeled as τ Y that distorts the size of the firm. Intuitively, a firm faces a high distortion in the capital-labor ratio (i.e. τ K is high) when the ratio of 13

15 labor to capital compensation is high relative to what one would expect in the absence of distortions (i.e. when there is no within-industry variation on the ratio of labor to capital compensation). On the other hand, a firm faces a high distortion in its size (i.e. a high τ Y ) when it is smaller than it should be, in other words, when the labor compensation of the firm is low compared to what one would expect given the industry elasticity of output with respect to labor. Potential TFP gains increased from around 0.25 in 1995 to around 0.50 in 2007 (see Figure 3). Figure 5 plots the resulting TFP gains when eliminating variation in one distortion at a time. By switching down the capital-labor distortion, i.e., fixing τ Ki = 0 i, TFP gains increased from around 0.01 in 1995 to around 0.03 in Therefore the level of TFP gains due to size distortions is very low. In contrast, the level of TFP gains remain relatively high with an increase from 0.11 in 1995 to 0.20 in 2007 when the size distortion is switched off (i.e. τ Yi = 0 i) and the τ K distortion is present. All in all, the τ K distortion to the capital-labor ratio seems to be the most important distortion in explaining the evolution of aggregate misallocation in Figure 3. Moreover, the interaction between both distortions also explains a significant part of the misallocation level and increase. A possible rationale for this finding is that firms operating with bad input mixes (large capital distortion) also tend to be larger than optimal (large size distortion), worsening the misallocation problem. For instance, Smagghue (2015) shows that size-dependent regulations might also generate an inefficient reallocation of resources from capital intensive to labor intensive firms. Indeed, the within industryyear correlation between both distortions is 0.40 in levels and 0.24 in growth rates. Figure 5: Potential TFP gains from reallocation by type of firm-specific distortion Overall TFP Gain Due to capital distortion Due to interaction Due to size distortion 14

16 5.2 Sector-level analysis We first analyze the type of sectors in which the deterioration in allocative efficiency between 1995 and 2007 was more pronounced. We have information on several characteristics of 58 sectors at the 2-digit NACE rev. 2 classification (see Table B1 in the Appendix). In order to exploit this information we consider here changes in allocative efficiency at the 2-digit level. 13 Figure 6 plots the level of potential TFP gain from reallocation in 1995 against its change between 1995 and Most of the sectors (51 out of 58) experienced increases in such TFP gains, i.e. deterioration in allocative efficiency, which explains the overall deterioration in Figure 3. In particular, the industries Warehousing and support activities for transportation, Electricity, gas, steam and air conditioning supply, and Activities of head offices; management consultancy activities worsened the most while Manufacture of furniture, Manufacture of beverages, and Motion picture, video and television programme production, sound recording experienced slight improvements in allocative efficiency. Moreover, there is no relationship between the initial level of allocative efficiency and the change over the period. Therefore, we next investigate to what extent observable sector-specific characteristics can generate such differences in allocative efficiency between sectors. In particular, we consider four different dimensions that might be related to the evolution of allocative efficiency. Figure 6: TFP gain and its change by 2-digit sector Change in TFP gain ( ) TFP gain in 1995 by sector (NACE 2 digits) 68 First, we explore the role of skill intensity differences across sectors. There are several reasons 13 Note that the results are based on misallocation within 4-digit sectors that is aggregated to the 2-digit level using production-based weights as oppossed to computing misallocation within each 2-digit sector. 15

17 why this may matter. For instance, firing costs have been long blamed as a possible source of misallocation of workers across firms (see Dolado, Ortigueira, and Stucchi (2011)). Firing costs on open ended contracts are high in Spain. However, the share of employment under flexible fixedterm contracts was large and stable over the period. 14 It has been argued that fixed-term contracts, because they preclude human capital accumulation on the job, are more prevalent among low skilled occupations. Hence, if firing costs are a source of misallocation, we should expect a larger increase in misallocation in high-skill industries. Skill intensity in US sectors is taken as our baseline proxy because it is expected to be exogenous to the evolution of allocative efficiency in Spanish sectors of activity. As a robustness check, we also consider the share of skilled workers in Spain taken from PITEC (Panel de Innovación Tecnológica), which is based on a survey of innovative firms conducted by the National Statistics Institute. 15 Second, differences in external financial dependence across sectors may affect the resource allocation process. The sharp expansion in bank lending during the period originated a stock of loans from credit institutions to non-financial corporations of 90% of GDP in 2007 while it was 38% in The increasing abundance of new credit to firms together with a loose screening process by banks can generate a deterioration in allocative efficiency if bad firms are able to survive hampering the reallocation process towards better firms. In order to check this potential channel, we consider a sector-specific finance intensity variable constructed by Fernald (2014) for the US. Exploiting I-O tables, this finance intensity variable is given by nominal purchases of intermediate financial services as a share of industry gross output. Again, using US sector characteristics ensures exogeneity with respect to the evolution of allocative efficiency in Spanish industries. As an alternative measure of sector-specific financial dependence, we consider the ratio of sector s total liabilities as a percentage of its total assets computed using firm-level data from the Central Balance Sheet Data. Third, more dynamic industries can be expected to produce better allocations of resources. For instance, more innovative sectors have usually larger shares of innovative and young firms that can easily adapt to shifts in demand or actions taken by competitors. Cecchetti and Kharroubi (2012) argue that credit booms (such as the one witnessed in Spain over ) undermine R&D intensive sectors, which might be related to the deterioration in TFP growth. Along these lines, we consider Fernald (2014) IT intensity variable at the sector level in the US, which consists on the payments for IT as a share of income (taken from the Bureau of Labor Statistics). As an alternative measure of sector-specific IT content, we exploit the Spanish PITEC to construct shares of R&D investment over total investment. Finally, some sectors may be less competitive because business success is related to state licensing or regulation. If this is the case, we could expect some firms in such sectors to operate with size or 14 The share of fixed-term contract was around 35% of employment in those years, but it witnessed a large increase before See for more details. 16

18 input mix far from optimal and still survive. To explore this hypothesis, we define a dummy variable taking value 1 for those sectors considered to be crony sectors, which we define as those sectors susceptible to monopoly or requiring licensing or highly dependent on government regulation. 16 As an alternative measure, we also consider the Bribe Payers Index constructed by Transparency International with information at the sector level. The 2011 Bribe Payers Survey, on which the index is based, asked business executives how common bribery was in the sectors with which they have business relations. 17 Table 4 shows some correlations between the sector characteristics just described and the changes in allocative efficiency. In particular, we regress the change in sector-specific potential TFP gains on the different characteristics measured as the average over the period. Columns (1)-(5) are based on linear regressions with different covariates. Column (6) is based on weighted-average least squares (WALS), a model averaging approach that provides standard errors incorporating not only parameter uncertainty but also model uncertainty These sectors are, casinos, coal, palm oil and timber, defense, deposit-taking banking and investment banking, infrastructure and pipelines, ports, airports, real estate and construction, steel, other metals, mining and commodities, utilities and telecoms services. In our dataset, we label as crony the following 2-digit sectors: 24, 35, 37, 38, 39, 41, 42, 50, 51, 61, and 68 (see Table B1 in the Appendix). 17 The survey asked how often three different types of bribery were perceived to occur in each sector: firstly, bribery of low-ranking public officials; secondly, improper contributions to high-ranking politicians to achieve influence; and thirdly, bribery between private companies. Answers were given on a 5-point scale. This was then converted to a 10-point scale where 0 indicates that companies in that sector are perceived to always pay bribes and 10 to never pay bribes. 18 Model uncertainty results from the lack of theoretical guidance on the particular regressors to include in the empirical model. When model uncertainty is present, traditional standard errors would under-estimate the real uncertainty associated to the estimate of interest because variation across models is ignored. In order to account for both levels of uncertainty, model averaging techniques (e.g. WALS) estimate all possible combinations of regressors and constructs a single estimate by averaging all model-specific estimates (see Moral-Benito (2015) for an in-depth analysis of model averaging). 17

19 Table 4: Misallocation and sector-specific characteristics. Dep. Variable: TFP Gain (1) (2) (3) (4) (5) (6) OLS OLS OLS OLS OLS WALS High-skill intensity (US share) (0.219) (0.271) (0.210) Innovative content (US IT intensity) (0.445) (0.501) (0.408) Financial dependence (US financial intensity) (0.029) (0.032) (0.027) Public sector influence (crony dummy) 0.226*** 0.209** 0.150** (0.081) (0.086) (0.077) Constant 0.219*** 0.216*** 0.148** 0.197*** ** (0.069) (0.046) (0.066) (0.034) (0.078) (0.068) Observations R-squared Notes: TFP Gain refers to the change over the period in the ratio of optimal TFP in the absence of misallocation to observed TFP. We fail to find any statistically significant relationships between skill intensity, innovative content, or financial dependence with the change in allocative efficiency (see Column (1)-(3) in Table 4). Furhtermore, the R-squared indicates that variation in these characteristics can only account for less than 0.5% of the variation in misallocation changes. In contrast, Column (4) in Table 4 indicates that the deterioration in allocative efficiency was 22.6 points larger in crony than in non-crony sectors. This statistically significant difference implies that the eleven industries in which success in business depends more on relationships between firm managers and public sector officials were the industries experiencing the largest increases in misallocation over the period. In addition, the crony dummy is able to account for 11% of this variation. When all the variables are jointly included in the regression in column (5), the magnitude and significance of the crony dummy remains virtually unaltered; however, partial correlations of skill intensity, innovative content, and financial dependence are statistically indistinguishable from zero. Finally, column (6) reports WALS estimates confirming the conclusion from column (5) even when we also account model uncertainty. 18

20 Table 5: Misallocation and sector-specific characteristics Alternative proxy regressors. Dep. Variable: TFP Gain (1) (2) (3) (4) (5) (6) OLS OLS OLS OLS OLS WALS High-skill intensity (Spain share) (0.155) (0.157) (0.131) Innovative content (Spain R&D share) * (0.249) (0.249) (0.215) Financial dependence (Spain debt burden) (0.097) (0.092) (0.086) Public sector influence (BPI index) *** *** ** (0.086) (0.090) (0.084) Constant 0.187*** 0.257*** 0.228*** 2.015*** 2.033*** 1.553*** (0.053) (0.040) (0.051) (0.058) (0.623) (0.585) Observations R-squared Notes: See notes to Table 4. Table 5 shows the same set of results but considering alternative proxy variables for each sectorspecific characteristic. For the influence of public sector we consider the BPI index described above and for the other three dimensions we evaluate the Spanish counterparts of the US indicators. Again, public sector influence is significantly related to changes in misallocation; in particular, sectors in which the incidence of bribery is larger experienced large increases in misallocation over the period (note that the lower the BPI index the higher the bribery incidence). Turning to skill intensity, innovative content, and financial dependence in columns (1), (2), and (3), respectively, we again fail to find statistically significant correlations in all the three cases. Using these alternative proxies, columns (5) and (6) of Table 5 also confirm the results in the corresponding columns of Table Firm-level analysis We now turn to exploit firm-level data on distortions. Formally, the two measures of firm-specific distortions provided by the theoretical model, τ Ki and τ Yi, are given by equations (17) and (18), respectively. We aim to investigate the characteristics of firms behind the increases in misallocation over the period; hence, we define the firm-specific growth rates of τ Ki and τ Yi to facilitate the interpretation of our estimates. 19 We then regress the firm-specific changes in distortions on two firm characteristics (size and age) including a set of 4-digit industry dummy variables as well as a set of year dummy variables to ensure that identification is based on across firms variation within each industry-year. 19 More specifically, ln(1 + τ Ki,t ) = ln(1 + τ Ki,t ) ln(1 + τ Ki,t 1 ) and ln(1 τ Yi,t ) = ln(1 τ Yi,t ) ln(1 τ Yi,t 1 ). We do not use the firm-specific growth rate over the entire period because it would drastically reduce the number of observations available for estimation. 19