Job Losses in the Great Recession

When Credit Dries Up: Job Losses in the Great Recession March 2017 Abstract We study whether the solvency problems of Spain’s weakest banks in the Great Recession have caused employment losses outside the …nancial sector. Our analysis focuses on the set of banks that were bailed out by the Spanish authorities. Data from the credit register of the Bank of Spain indicate that these banks curtailed lending well in advance of their bailout. We show the existence of a credit supply shock, controling for unobserved heterogeneity through …rm …xed e¤ects, and assess its impact on employment. To this aim, we compare the changes in employment between 2006 and 2010 at client …rms of weak banks to those at comparable …rms with no signi…cant pre-crisis relationship to weak banks. Our estimates imply that around 24% of job losses at …rms attached to weak banks in our sample are due to this exposure. This accounts for one-half of the employment losses at those that survived and one-third of employment losses at those that closed down. Keywords: Job losses, Great Recession, credit constraints. JEL codes: D92, G33, J23. 1 Introduction Do shocks to the banking system have real e¤ects and, if so, do they cause job losses outside the …nancial sector? Both questions have strongly resurfaced in the wake of the economic and …nancial crisis that started in 2008. The renewed interest in the real e¤ects of credit supply shocks is motivated by the exceptionally strong and persistent contraction of economic activity in the countries that su¤ered a banking crisis, like the US and several peripheral countries in Europe. In this study we use data from the credit register of the Bank of Spain to analyze the link between the unprecedented drops in bank lending and employment in Spain. Our identi…cation strategy exploits large di¤erences in bank health observed at the onset of the crisis. The entire banking sector su¤ered the consequences of the collapse of the construction sector and the bursting of a housing bubble, and most banks were cut o¤ from wholesale funding for a while, but the main problems were concentrated in savings banks (cajas de ahorros). We focus our attention on the ones that were bailed out by the Spanish government as part of a large-scale restructuring process that involved 32 savings banks and one commercial bank. We show that this set of weak banks started to curtail lending relative to the other banks almost two years ahead of the bailout process. Our main objective is to explore how this credit supply shock a¤ected employment growth in 2006-2010 at the …rms that maintained a pre-crisis relationship with any of these weak banks. The Spanish economy provides an ideal setting to analyze this issue. Spanish …rms are more reliant on bank credit than their counterparts in most advanced economies. In 2006, the stock of loans from credit institutions to non-…nancial corporations represented 86% of GDP compared to 62% in the European Union (European Central Bank, 2010). On the contrary, funding through …nancial markets is rarely used: on average only …ve large corporations per year issued publicly traded debt between 2002 and 2010, and the number of companies listed in the stock market is tiny. Finally, the vast majority of …rms in Spain are small and medium-sized enterprises (SMEs), and many of them became highly leveraged prior to the recession. We are obviously not the …rst to estimate the real e¤ects of credit supply shocks. 1 A number of recent studies have demonstrated the adverse impact of such shocks on employment and investment (see Section 2). Most of this work is however based on either incomplete data on bank loans to the corporate sector –such as syndicated loans– or information about banking relationships rather than loans. One of the main contributions of our study is to use a comprehensive data set that includes loan level data. The credit register of the Bank of Spain provides exhaustive information about all bank loans to …rms in the non-…nancial sector and these data are matched to balance sheet data for all banks and nearly 150,000 …rms. We are thus able to trace all credit ‡ows to a large sample of …rms, for which we also have information on their credit history, such as loan defaults, and their applications for a …rst loan from other banks. Hence, we observe credit demand by …rms that apparently need to establish a new banking relationship and the data from the credit register allow us determine whether the loan application was granted or not. Our high-quality data let us perform more tests than related studies carry out and explore the presence of heterogeneous e¤ects along many dimensions, but above all they allow us to resolve several key identi…cation issues. The …rst challenge is the need to disentangle changes in credit supply from concurrent changes in credit demand. Between 2007 and 2009, the European Central Bank (ECB) Bank Lending Survey indices for Spain show a simultaneous increase of around 40% in the bank lending standards applied to non-…nancial …rms and a similarly-sized drop in their loan demand (Banco de España, 2015). To obtain clean identi…cation of the credit supply shock we use the standard procedure of Khwaja and Mian (2008) to analyze credit growth at the bank-…rm level. The results clearly demonstrate that weak banks curtailed credit vis-à-vis healthy banks that lent to the same …rm. Thus, controling for …rm …xed e¤ects, we show that there was indeed a di¤erential credit supply shock. Moreover, we also establish that the a¤ected …rms could not …nd new lenders to fully compensate this reduction in credit supply, which implies that the impact found at the bank-…rm level was partially transmitted to the level of the …rm. Finally, when we analyze employment growth we …nd that these …rm-level credit shocks caused relatively large employment losses at …rms with a pre-crisis relationship to weak banks. To capture these shocks, our baseline 2 di¤erence-in-di¤erences speci…cation includes a dummy variable for …rms with a weakbank loan-to-asset ratio above a minimum threshold. The second main challenge is to control for selection e¤ects. Before the crisis, healthy banks may have worked with better …rms than weak banks, as indicated by di¤erences in the observable characteristics between …rms above and below the threshold for the weak bank loan-to-asset ratio. Failing to control for these di¤erences would bias our estimates upwards if the …rms above the threshold su¤ered a stronger contraction of demand or if they were more vulnerable to reductions in the availability of credit than the …rms below the threshold. Moreover, selection may also occur on unobservables and in this case the introduction of …rm controls need not su¢ ce to avoid the bias. As already indicated, in the analysis of credit growth at the bank-…rm level this problem is dealt with by introducing …rm …xed e¤ects. Moreover, we also estimate a …rm-level speci…cation in which the …xed e¤ects are replaced by …rm control variables. This di¤erence-in-di¤erences speci…cation in growth rates includes exhaustive controls for di¤erential trends in credit growth by …rm characteristics and it includes a full set of …xed e¤ects by industry-municipality pairs to control for local demand e¤ects. The resulting estimate of the impact of weak-bank attachment on credit growth is virtually identical to the one obtained in the within-…rm speci…cation. This reveals that unobservables do not play a signi…cant role as far as access to credit is concerned, but they could still generate di¤erent patterns of employment growth. In our analysis of the impact of weak-bank attachment on …rm-level employment we perform several tests to deal with the potential problem of selection on unobservables. First, we show, using the normalized di¤erence test of Imbens and Wooldridge (2009), that selection on observables is not that important, because …rms attached to weak banks only di¤er signi…cantly from the rest in three of the 14 controls that we use: they have a lower capital-to-assets ratio, a higher debt ratio, and they work with more banks. In addition, we estimate a panel …xed-e¤ects model with yearly observations for employment growth and we use matching techniques to directly compare …rms within narrowly-de…ned cells. In neither case do we …nd any signi…cant di¤erence with our baseline. We nevertheless also apply the procedure in Oster (2017) to place an upper bound on the impact of unobservables on our coe¢ cient of interest. Lastly, we exploit 3 a legal change in the regulation of savings banks in December of 1988 to construct an instrument that generates exogenous variation in weak-bank attachment. The results of this instrumental variable model must be interpreted with caution, as explained below, but they con…rm that weak-bank attachment exerted a signi…cant negative e¤ect on employment growth. In the rest of the analysis we consider several extensions of our baseline setup to analyze the transmission mechanism of the credit shock and to explore the role of …rms’ …nancial vulnerability. We start by estimating a separate e¤ect of weak-bank attachment on credit lines, …nding that weak banks strongly reduced access to credit lines relative to healthy banks. A direct impact of working capital on employment therefore seems to prevail over potential indirect e¤ects via reduced investment. Furthermore, we o¤er a novel decomposition of job losses through adjustments along the internal and external margins and we estimate the e¤ect of weak-bank attachment on changes in the wage bill and in the share of employees with a temporary contract. To the best of our knowledge, we are the …rst to o¤er a comprehensive analysis of these margins of adjustment. Lastly, we interact the treatment variable with an extensive set of …rm characteristics that capture di¤erent dimensions of …rms’…nancial vulnerability. Our baseline result is that weak-bank attachment caused employment losses of about 2.8 percentage points. This estimate is large, accounting for 24.4% of the total fall in employment among exposed …rms in our sample. Surviving …rms account for about onehalf of the overall loss, while the remaining half corresponds to job losses in exiting …rms. Nonetheless, weak-bank exposure explains a larger share of job losses in downsizing …rms than in exiting …rms (around one-half vs. one third, respectively), and we show that temporary employees bore the brunt of the employment adjustment. Lastly, we …nd that …nancially vulnerable …rms su¤ered the largest job losses. The rest of the paper is organized as follows. In Section 2 we review previous empirical work on the topic and in Section 3 we provide background information on the Spanish economy before and during the …nancial crisis. Section 4 describes our data and Section 5 presents our empirical strategy. In Section 6 we show our estimates of the e¤ect of weak-bank attachment on credit growth and in Section 7 our baseline employment e¤ect estimates. Selection e¤ects are dealt with in Section 8 and Section 9 4 presents results on treatment heterogeneity. Various margins of adjustment are studied in Section 10. Section 11 contains our conclusions. Two appendices provide information on weak banks and securitization, as well as details on the variables used. 2 Literature review In recent years there has been a surge of studies exploiting quasi-experimental techniques to estimate the real e¤ects of credit supply shocks.1 The closest ones are the seminal contribution of Chodorow-Reich (2014) and the article by Greenstone et al. (2014).2 Both studies exploit cross-sectional di¤erences in lender health at the onset of the recent crisis to study the link between credit supply shocks and employment. In Greenstone et al. (2014) this link is indirect, as they do not have access to loan-level data or information about …rms’ banking relationships. To circumvent this problem, they construct a county-level credit supply shock from the product of the change in US banks’small-business lending at the national level and their predetermined credit market share at the county level. Using con…dential data from the Bureau of Labor Statistics Longitudinal Database (LBD), they …nd that this measure predicts the reduction in county-level credit to small, standalone …rms and their employment levels over 2008-2009. But even assuming that the entire reduction in lending is due to a drop in credit supply, the estimated e¤ect is small, around 5% of the employment fall. Chodorow-Reich (2014) does have access to loan-level data from the Dealscan syndicated loan database. He constructs a …rm-speci…c credit supply shock that is equal to the weighted average of the reduction in lending that the …rm’s last pre-crisis syndicate 1 Several studies exploit the variation induced by large external shocks to the banking system (Chava and Purnanandam, 2011, Benmelech et al., 2012). Almeida et al. (2012), Benmelech et al. (2012), and Boeri et al. (2013) exploit di¤erences in the debt maturity structure of …rms. Garicano and Steinwender (2015) compare the response of di¤erent types of investment at foreign-owned and nationally-based manufacturing …rms in Spain. 2 Popov and Rocholl (2016) analyze the impact of German savings banks’ exposure to the US subprime crisis on …rms’labor demand. Fernandes and Ferreira (2017) explore the impact of …nancing constraints on the choice between permanent and …xed-term contracts. Berton et al. (2017) apply the Greenstone et al. (2014) approach to examine the impact of credit constraints in Italy on di¤erent types of workers. Other papers that study the employment e¤ects of the …nancial crisis are DuyganBump et al. (2015), Hochfellner et al. (2015), and Siemer (2014). Some papers include employment among a broader set of real outcomes. Acharya et al. (2016) study the real e¤ects of the sovereign debt crisis in Europe using syndicated loan data. Balduzzi et al. (2015) use a survey of Italian …rms to analyze the real e¤ects of ‡uctuations in banks’cost of funding during the 2007-2009 …nancial crisis and the ensuing sovereign debt crisis. 5 imposes on other …rms during the crisis. These data are matched to employment records from the LBD data set for a sample of just over 2,000 …rms. In line with Greenstone et al. (2014), he …nds that SMEs with pre-crisis relationships with less healthy banks faced stronger credit constraints after the fall of Lehman Brothers and reduced their employment more compared to clients of healthier banks, attributing between one-third and one-half of job losses in SMEs to this factor. By contrast, there are no signi…cant e¤ects for the largest companies in the sample. In this paper we also exploit di¤erences in lender health to uncover the employment e¤ects of credit supply shocks, but the access to credit register data represents a substantial improvement on the existing work in this …eld. First, we are able to reconstruct the entire banking history of …rms and to trace back all credit ‡ows and not only syndicated loans.3 Second, the representative nature of our large sample of …rms is important to gauge the overall e¤ect of the credit shock on employment. Studies relying on data for relatively large …rms may substantially underestimate the impact of credit shocks if larger …rms are more able to …nd substitutes for bank credit than smaller …rms. We do not …nd compelling evidence of such di¤erences by …rm size, but in other countries they may be important when large …rms have access to well-developed markets for private debt. Third, access to detailed micro data allows us to perform a wider range of robustness checks and to explore the presence of heterogeneous e¤ects along more dimensions than most existing studies. Our analysis pays particular attention to the role of …rms’ …nancial vulnerability. Apart from standard indicators, such as …rm size or age, our analysis also includes controls for …rms’degree of bank dependence, the term structure of their bank debt, and their credit history. This analysis reveals that a bad credit history in the form of past defaults triples the negative e¤ects associated with a pre-crisis relationship with a weak bank and these e¤ects come on top of the almost 20 pp reduction in credit growth for all …rms with a bad credit history. Similarly, for bank-dependent …rms with 3 Cingano et al. (2016) study the e¤ects of banks’exposure to the interbank market on the investment decisions of their clients using data from the Italian credit register, also presenting results for employment growth. Jiménez et al. (2016) include an estimate of the impact of the size of banks’ capital bu¤ers on employment growth in Spain, but they focus on the impact of counter-cyclical capital bu¤ers on credit supply over the business cycle. These two studies con…rm the negative employment e¤ects of impaired access to credit. 6 a ratio of bank debt to total debt above the median, job losses are …ve times bigger than the average treatment e¤ect. These strong di¤erences in the intensity of the e¤ects con…rm the …nding in Paravisini et al. (2015) that it is key to compare …rms within very narrowly-de…ned cells to avoid omitted variable bias. The theoretical literature has identi…ed several potential transmission mechanisms through which shocks to the banking system might a¤ect employment in non-…nancial …rms (see the review in Boeri et al., 2013). First, mismatch between the timing of payments to workers and the generation of cash ‡ow may force …rms to …nance salaries as part of their working capital. Second, turnover costs in the labor market transform labor into a quasi-…xed factor of production, creating a link between employment and external …nance that is similar to the well-known link with investment. Third, …nancial frictions may alter the optimal mix of permanent and temporary jobs, as the latter are cheaper to destroy, and this may in turn have important implications for the cyclical volatility of employment. Lastly, the availability of external …nance may indirectly alter the use of labor if capital and labor are complements in production.4 While we cannot identify the relevance of these mechanisms, we try to shed some light on them in several ways. On the one hand, we explore the relative importance of weak-bank attachment on short-term funding, which indirectly informs us about the purpose of the loans. Next, we consider three alternative margins of employment adjustment. First, we o¤er a decomposition of job losses along the internal and external margins, showing that weak-bank attachment leads to a signi…cant increase in …rms’ exit probability. This …nding helps to understand the persistence of the e¤ects of credit shocks, since it is cheaper and quicker to create jobs at ongoing businesses than to rebuild …rms once the economy recovers. And second, for the sample of surviving …rms we estimate the e¤ect of weak-bank attachment on the size of the wage bill and the share of temporary jobs. Popov and Rocholl (2016) and Fernandes and Ferreira (2017) o¤er comparable results on the importance of wage cuts and changes in the composition of employment, respectively, but we are the …rst to consider all three margins jointly. 4 See Wasmer and Weil (2004) and Petrosky-Nadeau and Wasmer (2013) on frictions, and Caggese and Cuñat (2009) on temporary jobs. 7 3 The …nancial crisis in Spain The Spanish economy experienced a severe credit crunch in the Great Recession. In this section we brie‡y document the magnitude of this credit crunch, but we start by de…ning the set of weak banks, so that we can compare the evolution of lending by weak and healthy banks. 3.1 The bank restructuring process During our sample period, the Spanish Government intervened a total of 33 banks (see Table A1). First two small banks were nationalized and quickly resold. Later on the Government fostered either bank mergers (26 weak banks) or the takeover of ailing banks by other banks (5 banks). Most of these operations entailed State support, in the amount of 11.6 bn euro, i.e. about 1.1% of Spanish GDP (Banco de España, 2014). We classify a bank as weak if it was nationalized, it participated in a merger with State funding support or it was insolvent and bought by another bank. Banks that absorbed other banks are considered to be healthy. During our sample period, all weak banks except the two small nationalized ones were run by their incumbent managers and all the institutions participating in mergers remained separate legal entities. Further consolidation operations and the bulk of the nationalizations took place in 2011-2012 (see Appendix 1 and International Monetary Fund, 2012), and savings banks were forced to convert into commercial banks. These operations fall outside the scope of our analysis and the same is true for the recapitalization of the banking sector that took place in 2012 with …nancial assistance from the European Financial Stability Facility. Finally, it is important to stress that savings banks were subject to the same regulation and supervision by the Bank of Spain as commercial banks, though they had a di¤erent ownership and governance structure. Not being listed in the stock market, they were less exposed to market discipline than commercial banks but also quite limited in their ability to raise capital in response to the crisis. Furthermore, they were de facto controled by regional governments, which led to delays in their restructuring and may have a¤ected their credit allocation prior to the crisis.5 5 See Cuñat and Garicano (2010), Fernández-Villaverde et al. (2013), and Santos (2014). 8 3.2 The di¤erences in lender health Table 1 illustrates the di¤erences in lender health at the onset of the crisis. In the last two columns we report the t-statistic and the normalized di¤erence test of Imbens and Wooldridge (2009), for which Imbens and Rubin (2015) suggest a heuristic threshold of 0.25.6 Both tests yield the same results. In 2006, weak banks were on average larger than healthy banks and held less capital and liquid assets. By contrast, both their rate of return on assets and their share of non-performing loans were comparable, but this apparent similarity hides latent losses at weak banks, which surfaced later, as witnessed by their vastly larger ratio of non-performing loans in 2012.7 The ratio of securitized loans to assets is also larger for weak banks, but not signi…cantly so, suggesting that it was not key in explaining the di¤erential evolution of credit during the crisis between the two sets of banks. Below we conjecture that the comparatively large share of loans to construction companies and real estate developers –henceforth real estate industry or REI– is a key source of the surge in loan non-performance and the comparatively strong contraction of lending by weak banks. Loans to the REI make up 68% of all loans of weak banks to non-…nancial …rms compared to 37% for healthy banks. The split between weak and healthy banks allows us to analyze the compound e¤ect of these initial di¤erences in lender health on credit supply, including latent losses not o¢ cially recognized until much later. The weak bank label should therefore be interpreted as a proxy for the relatively strong deterioration of the balance sheets and the lending capacity of the most vulnerable banks. 3.3 The credit collapse Figure 1 depicts the real value of the annual ‡ow of new credit to non-…nancial …rms by month and bank type (average over the past 12 months). It reveals that the ‡ow of new credit grew signi…cantly more at weak than at healthy banks during the boom –60% vs. 12% from 2002 to 2007–and it also fell more in the slump –46% vs. 35% from 2007 to 2010. This evolution stems from both the intensive and extensive margins. The 6 See a brief explanation of the use of this statistic in large samples in Section 4.2. In 2012 the authorities carried out stringent stress tests on banks, supervised by the ECB, the European Commission, and the International Monetary Fund. 7 9 latter is captured in Figure 2, which represents acceptance rates for loan applications by potential clients (non-client …rms). Initially acceptance rates were higher for weak than for healthy banks, they then converged, subsequently both fell precipitously, and at the end of the period they were lower at weak banks. This evolution re‡ects the di¢ culties faced by Spanish …rms trying to switch to a new lender during the crisis. Figure 3 shows that the average interest rates charged by the two sets of banks closely follow the ECB policy rate. Until the end of our sample period the di¤erence between them was always below 30 basis points. This suggests that interest rates were scarcely used by weak banks to ration credit demand during that period.8 We can therefore safely focus on the di¤erential evolution of credit volume at the two sets of banks during the crisis. The same graphs also show that the consolidation operations and nationalizations during 2011-2012 did not restore the credit ‡ow to weak banks. The gaps between bank types regarding new credit ‡ows and acceptance rates continued to grow and weak banks also started to ration credit by charging substantially higher average interest rates than healthy banks. 4 Data In this section we describe our data set, the sample selection procedure, and the construction of the treatment and control groups. For further details see Appendix 2. 4.1 Data sources We construct a matched bank-…rm data set with detailed information on all bank loans to non-…nancial …rms. Even though our analysis focuses on the period 2006-2010, we collect data starting in 2000. The loan data is obtained from the Central Credit Register (CIR) of the Bank of Spain, which records all bank loans to non-…nancial …rms above 6,000 euros (around 7,900 dollars at the end of 2006). Given the low threshold, these data can be taken as the census. The CIR provides the identity of the parties involved 8 Stiglitz and Weiss (1981) explain why imperfect information leads to credit rationing rather than interest rate di¤erences, and Petersen and Rajan (1994) show that US banking relationships operate more through quantities than through prices. 10 in a loan, the share of collateralized loans by …rm, its maturity structure, the identity of its main bank –i.e. the one with the largest value of outstanding loans–, and indicators of its creditworthiness, such as the value of the …rm’s non-performing and potentially problematic loans. It does neither record the interest rate nor the purpose of the loan. We therefore have to rely on the distinction between credit lines and loans to indirectly establish a shortage of working capital or of funds for investment.9 Apart from the information on new and outstanding loans, we also have access to loan applications from non-clients.10 By matching the records on loan applications with the CIR we infer whether the loan materialized. If not, either the bank denied it or else the …rm obtained funding elsewhere (Jiménez et al., 2012). Since the application data set only provides information on borrowing for …rms with a credit history, we exclude …rms with no prior loans. We gather economic and …nancial information for more than 300,000 private, non…nancial …rms from the mandatory annual balance sheets and income statements that Spanish corporations submit to the Mercantile Registers. Our source is the Iberian Balance Sheet Analysis System (SABI) produced by INFORMA D&B in collaboration with Bureau Van Dijk and the Central Balance Sheet Data O¢ ce (CBSO) of the Bank of Spain. We match the data on loans, banks, and …rms through …rms’tax ID. Employment is measured as the annual average of employees, where temporary workers are weighted according to their weeks of work. SABI also provides information on variables like the …rm’s age, size, and indebtedness, though for most …rms we only observe an abridged balance sheet with no liability breakdown. Lastly, we observe the …rm’s industry and use a two-digit breakdown into 78 industries. To disentangle job losses in surviving …rms from those due to …rm closures, we use 9 As indicated, there was signi…cant merger and acquisitions (M&A) activity in the Spanish banking industry in the sample period. In most empirical speci…cations we only use loans with banks in 2006. In one speci…cation however, in Section 6.1, we use …rm-bank relationships as units of analysis throughout the period 2006-2010. To ensure that we follow credit relationships after a bank is acquired or merged into a new entity, we arti…cially reconstruct the banking relationships from 2010 backwards. So if, from 2006 to 2010, bank A acquires bank B and a …rm had a loan from bank B, a single relationship will appear in 2006 with bank A, which will encompass both banks, while keeping track of the weak identity of either bank if that is the case. Still, bank merger activity was only relevant from 2010 onwards. 10 Banks receive monthly information from the CIR on their borrowers’total indebtedness and defaults vis-à-vis all banks in Spain, but they can also get it on “any …rm that seriously approaches the bank to obtain credit”. 11 the Central Business Register (DIRCE), which allows us to ensure that …rms that are in the sample in 2006 but which disappear from it in subsequent years have indeed closed down.11 Lastly, we exploit two databases on banks. The …rst one, used for supervisory purposes, records their …nancial statements. It includes 239 banks, comprising commercial banks, savings banks, and credit cooperatives. The second one contains historical data on the location of bank branches at the municipal level, which is used for the …rst time for research purposes. 4.2 The treatment and control groups To analyze the employment e¤ects of the credit reductions by weak banks, we divide the sample of …rms into two groups depending on the strength of their pre-crisis relationship to weak banks. We measure weak-bank attachment through the ratio of the total amount of loans from weak banks to the …rm’s asset value. It is the product of the …rm’s ratio of debt with weak banks to total debt –i.e. the weight of weak banks in debt–and the ratio of total debt to asset value –leverage–and it is measured in 2006. Our baseline treatment measure is a dummy variable denoted W Bi , which takes the value 1 if the weak-bank loan-to-asset ratio for …rm i is above the …rst quartile of the distribution of …rms with non-zero exposure. The chosen threshold for our weak-bank dummy has the advantage of excluding …rms with marginal attachment to weak banks from the treatment group, while still providing a conservative estimate of the impact of such attachment, since the threshold is set at the lower end of the distribution.12 We will nevertheless show the robustness of our …ndings to the use of the continuous weak-bank loan-to-asset ratio measure and to di¤erent cut-o¤ levels for the weak-bank dummy. Given the size of our data set, we adopt stringent sample selection rules. To avoid the problem of reverse causality –so that …rms’troubles drive banks’problems rather than the opposite– we exclude …rms in the REI or in two-digit industries selling at 11 We do not observe M&A. However, the CBSO sample of …rms above 50 workers contains such information and in 2012 only 3% of all …rm closures according to the DIRCE resulted from M&A. Since M&A usually take place among large …rms and in our sample only 5% of …rms are above that threshold, we expect to have a much lower fraction. A …rm may be closed down with one ID and then opened with another one, but this type of transaction cannot be identi…ed. 12 The …rm at the median exposure ratio in the overall distribution, which could be an alternative threshold for the treatment, has no loans from weak banks. 12 least 20% of their value added to the REI in 2000 (see Appendix 2). This date is chosen to minimize potential endogeneity through credit decisions taken in the boom.13 Throughout the analysis we work with a balanced sample and we only include …rms in our sample for which we have reliable observations on all variables from 2006 to 2010. In particular, we exclude …rms that do not deposit their accounts after 2006 but still appear in the Central Business Register. Hence, only …rms that are missing in both registers are classi…ed as having closed down. Moreover, since we are interested in bank credit, we exclude …rms with no loans in 2006. This leaves us with a …nal sample of 149,458 …rms. We choose 2006 as the base year because both GDP and real credit were growing very quickly, at 4.1% and 19% p.a., respectively, so that neither the recession nor the credit crunch were generally anticipated then. However, in one speci…cation we set 2007 as the base year to check the robustness of our results to this dating. In 2006 the …rms in our sample represented 19% of …rms, 28% of value added, and 42% of private sector employees in the industries included in our analysis. Most …rms are very small, indeed, 98.7% of them are SMEs according to the European Commission de…nition (i.e. having less than 250 employees and either turnover below 50 million euros or a balance sheet total below 43 million euros). On average these …rms reduced employment by 8.1% during the sample period. This …gure is very close to the aggregate reduction in employment for the industries we cover. Table 2 presents descriptive statistics for our treatment and control groups. About 69% of …rms have either no credit from weak banks or a weak-bank loan ratio to assets below the …rst quartile –which is equal to 4.8%. For …rms above this threshold, the average share of credit from weak banks is equal to 68.5% and their ratio of weak-bank credit to assets to 22.8%. Compared with the control group, …rms in the treatment group seem to be on average younger and smaller, they have more temporary workers, and they are as likely to be exporters. The data also re‡ect the worse …nancial pro…le of …rms in the treatment group: they are less pro…table, hold less capital and liquid assets, and they are more indebted to banks, although the average maturity of their loans is higher. In addition, 13 The housing bubble is thought to have started around mid-2003 (Ayuso and Restoy, 2006). 13 they work with more banks and over 2002-2006 they defaulted more often on their loans. They also applied for loans more often and had a higher acceptance rate. These di¤erences are signi…cant according to the t-ratio, but in very large samples this statistic has the problem that it increases with the number of observations. Imbens and Wooldridge (2009) propose a scale-free alternative measure, the di¤erence in averages scaled by the square root of the sum of the variances. Indeed, as shown in the last two columns of Table 2, the results from these two tests diverge in our case. Using the normalized di¤erence test and the heuristic threshold of 0.25 suggested by Imbens and Rubin (2015), the only signi…cant di¤erence is that …rms exposed to weak banks have a lower degree of capitalization and a higher level of indebtedness, and they work with more banks than non-exposed …rms. Their lower liquidity and higher number of past loan applications are close to being signi…cant. This implies, …rst of all, that selection on observables is not a very important problem, once we take into account the large sample used. Second, this also means that we should control for those three …rm-level characteristics –or at most …ve–in our empirical analysis. Nevertheless, we will use all 17 variables listed in the table below the employment level as controls in our …rm-level credit and employment estimation to be sure that we are conditionally comparing very similar …rms. At this stage we need to deal with the potential objection that our treatment is de…ned in terms of an outcome, a bank‘s bailout, that is realized after the crisis broke out. Using an ex-post criterion does not invalidate our results, however, as long as the outcome was unexpected. To study whether …rms could have anticipated in 2006 the future solvency problems of weak banks, we analyze the risk premia charged to Spanish banks’ securitization issues prior to the recession. We use data on tranches of mortgage backed securities and asset backed securities in 2006, grouping the ratings into prime (AAA), investment grade (AA+ to BBB-), and speculative (BB+ to D). We have 303 observations (deal-tranches) from Dealogic, with ‡oating rate, quarterly coupon frequency, and referenced to the 3-month Euribor, from 24 issuer parents. Without any controls, weak banks actually paid 7 basis points less than healthy banks. To control for issue characteristics, we regress coupon di¤erentials in basis points on a set of variables capturing the type of securitization, risk category, month 14 of issue, years to maturity, collateral type, and guarantor type. Standard errors are clustered by issuer parent. The estimated coe¢ cient associated with the weak-bank dummy is positive but non-signi…cant: 2.8 basis points, with a p-value of 0.55 (see Table A2). Hence we cannot reject the hypothesis that …nancial markets failed to recognize the buildup of di¤erential risk at weak banks in 2006.14 It seems safe to assume that private …rms, with a lower capacity to process available information than …nancial markets, could not possibly have predicted it either. 5 Empirical strategy Our identi…cation strategy proceeds in two steps. The …rst step consists of establishing the presence of a credit supply shock associated with the troubles of weak banks. In this part we …rst estimate a credit equation at the bank-…rm level and then at the …rm level. The second part consists of estimating the e¤ects of weak-bank attachment on employment. 5.1 Identi…cation of the credit supply shock We start by estimating the following credit growth equation for bank-…rm pairs:15 log(1 + Creditib ) = where i + W Bb + Zib0 + Sb0 + ib (1) is a -year di¤erence with respect to the year 2006, Creditib is total credit committed by bank b to …rm i –both drawn and undrawn, so as to minimize potential endogeneity–, i is a …rm …xed e¤ect, W Bb is a dummy variable that takes the value 1 if bank b is weak, Zib is a vector of bank-…rm controls that includes the length of their relationship and a dummy variable for past defaults, Sb is a vector of bank controls, and ib is a random shock. Our coe¢ cient of interest is . Speci…cations like (1) are the standard procedure to identify credit supply shocks. The …xed e¤ects absorb any di¤erences in observable and unobservable …rm characteristics. As a result, they perfectly control for potentially confounding demand e¤ects, 14 Financial markets operators may have been aware of the concentration of risks in savings banks, but they may have also anticipated an implicit bailout guarantee. Either way, the risk perceived by funders is not statistically di¤erent. 15 Khwaja and Mian (2008) label this as the local analysis and the subsequent …rm-level equation the aggregate analysis. 15 allowing us to test whether the same …rm experiences a larger reduction in lending from weak banks than from healthy banks once we control for di¤erences in Zib and Sb . This equation can however be estimated only for …rms that work with more than one bank. Since many …rms in our sample work with a single bank, we also estimate a between…rm variant of equation (1) in which the …rm …xed e¤ects are replaced by a vector Xi of …rm controls. As originally explained by Khwaja and Mian (2008), this ordinary least squares (OLS) speci…cation may yield biased estimates of in the presence of both credit demand and supply shocks, but we show that this risk can be minimized by introducing a rich set of …rm controls and industry times municipality dummies to control for demand e¤ects. Next, even if we con…rm the presence of a credit shock at the bank-…rm level, we still need to check whether the a¤ected …rms managed to o¤set the reduction in credit supply by weak banks with additional loans from other banks. For this purpose we estimate the following …rm-level equation: log (1 + Creditij ) = + W Bi + Xi0 + j + vij (2) which links credit growth to our treatment variable W Bi , where Xi contains the 17 variables listed in Table 2, and j is a vector of industry (78) times municipality (2,749) dummies that control for local credit demand conditions. Here our coe¢ cient of interest is , which will typically be smaller than to the extent that …rms managed to obtain credit from healthy banks when weak banks curtailed their supply. 5.2 The employment impact of credit constraints We then proceed to estimate the impact of the credit supply shock on employment. To ensure that our estimates do not capture the e¤ect of di¤erences in observable characteristics of …rms rather than the e¤ect of credit supply, we adopt a di¤erence-indi¤erences (DD) speci…cation in growth rates that has the same structure as equation (2) (Wooldridge, 2010): log (1 + nij ) = + W Bi + Xi0 + j + uij (3) where nij is employment in …rm i in industry-municipality cell j and uij is a random shock. All regressors are again measured in 2006. This estimate is an average treatment 16 e¤ect on the treated (ATT). To measure the employment adjustment in both surviving and closing …rms, we set nij to zero for …rms that are present in 2006 but have closed down years later. Estimating in di¤erences implies that we are including an aggregate trend and differential trends by industry-municipality cells and by …rm characteristics, which is substantially more demanding than the standard DD speci…cation in levels. Unbiased estimation of the causal impact of weak-bank attachment on employment however relies on the unconfoundedness assumption, which requires the assignment of …rms to the treatment and control groups to be completely random conditional on the controls for observables. Moreover, unlike for credit, this speci…cation cannot perfectly control for confounding factors through the introduction of …xed e¤ects. The main threat to identi…cation is the non-random assignment of …rms to banks before the crisis. Aggregate shocks may di¤erentially a¤ect …rms depending on their pro…tability, product quality or …nancial vulnerability. If, for example, the crisis hit more …nancially vulnerable …rms harder and these …rms were over-represented among the clients of weak banks, then the DD speci…cation would tend to overestimate the real e¤ects of the credit shock. Our …rm controls, Xi , are meant to absorb potential di¤erences in both …rms’performance and their …nancial vulnerability and creditworthiness, but selection may take place on both observables and unobservables. This may not be an overriding problem, since only three of the 17 …rm characteristics that we use are statistically di¤erent, but we still devote Section 8 to deal with this issue. The other main concern is that demand e¤ects may bias our estimation (Mian and Su…, 2014). Before the crisis, lending grew especially in the real state industry and it was more concentrated in certain areas, where in the recession we might observe both a larger drop in demand by households and a higher density of (non-REI) …rms exposed to weak banks. In these circumstances employment reductions would stem from lower consumption demand rather than from less credit. The fact that small …rms tend to be …nanced by local banks (Petersen and Rajan, 2002; Guiso et al., 2013) would additionally contribute to the presence of local demand e¤ects. For this reason we allow for di¤erential trends in the j cells de…ned by the product of 2-digit industry and municipality dummy variables. 17 6 The di¤erential evolution of credit The …rst step is to validate our claim that the di¤erential evolution of the volume of lending by the two sets of banks re‡ects a credit supply shock, which we carry out through both bank-…rm and …rm-level analyses. 6.1 Bank-…rm analysis Table 3 reports the results for our baseline equation (1) and alternative speci…cations for the change in credit between 2006 and 2010. Robust standard errors are corrected for multiclustering at the …rm and bank level. The speci…cation with …rm controls and industry-municipality dummies yields an estimated di¤erential reduction in credit of 23.2 percentage points (pp) for weak banks vis-à-vis healthy banks (col. 1). Restricting the sample to …rms with multiple banking relationships reduces the estimate to -25.6 pp (col. 2), which is virtually identical to the estimate for our baseline speci…cation with …rm …xed e¤ects, -25.5 pp, estimated with the same sample of multi-bank …rms (col. 3). The similarity between these two estimates suggests that unobservables do not play a signi…cant role in access to credit, despite the di¤erences in three of the observable …rm characteristics. Moreover, a Hausman test fails to reject the null hypothesis of orthogonality between the …rm …xed e¤ects and W Bb with a p-value of 0.372. These results con…rm that our weak bank dummy variable captures changes in credit supply rather than in credit demand. Was credit rationing by weak banks stronger on short-term funding? To answer this question we interact W Bb with an indicator for …rms that enjoyed a credit line in 2006. The results in Table 3 indicate that weak banks reduced credit to …rms with credit lines by 10.6 pp more than healthy banks (col. 4). A natural interpretation is that credit lines were the easiest loans to cut for banks in distress and, possibly, that the credit constraint a¤ected more working capital than investment. In Appendix Table A3, we present the results when we also include dummies for maturities of 1-3 years, 3-5 years, above 5 years, and for unknown maturity (a¤ecting 4% of the loans), with loans below one year being the reference category, and their interactions with the weakbank dummy variable. The estimates do not indicate a di¤erential e¤ect of the credit 18 supply shock for di¤erent loan maturities, as none of the coe¢ cients of the interactions is statistically signi…cant. Next, we show the estimate for bank-…rm relationships which were still alive in 2010 (col. 5). In this case, the di¤erence in the reduction of credit supply is equal to 7.9 pp, indicating that adjustments at the internal margin –i.e. reductions in loan volume to existing lenders–account for a small share of the reduction in lending by weak banks, so that the external margin dominates –i.e. committed credit is reduced to zero or renewal of expired loans is denied. The di¤erential e¤ect captured by the weak-bank dummy indicates that standard measures of bank health do not fully capture the deterioration of weak banks’assets. A natural explanation for this …nding relies on di¤erences in banks’pre-crisis exposure to the REI. We measure it by the 2006 share of each bank’s loans to …rms in the REI and create a dummy variable that takes the value 1 for banks in the upper quartile of the cross-sectional distribution. The associated coe¢ cient is 7.5 pp smaller and less signi…cant than in our baseline (col. 6). Hence, latent losses elsewhere in the balance sheets of weak banks may have been imperfectly correlated with their exposure to the REI. Lastly, our identi…cation procedure relies on the absence of di¤erent pre-crisis trends in access to credit for …rms in the treatment and control groups, which would lead to biased estimates. To check this issue, panel A of Figure 4 shows the yearly coe¢ cients, from 2004 to 2010, for our baseline speci…cation with …rm …xed e¤ects. The coe¢ cient of W Bb is not signi…cantly di¤erent from zero between 2004 and 2007; indeed, except for 2004 the point estimates are equal to zero. The treatment e¤ect becomes signi…cant in 2008 and it grows over time from -10 pp in 2008 to over -25 pp in 2010. This shows that weak-bank exposure has no signi…cant impact on access to credit prior to 2007 once we control for …rm-…xed e¤ects.16 In sum, we have shown that weak banks reduced credit more than healthy banks, not just in the aggregate but also at the level of individual bank-…rm relationships. 16 In the case of single-bank …rms we cannot include …rm …xed e¤ects, but introducing …rm controls leads to similar results. 19 6.2 Firm-level analysis We now study credit rationing at the …rm level. The dependent variable is the log di¤erence between the …rm’s total credit outstanding in 2006 and 2010, and the weakbank indicator of the bank-…rm analysis is replaced by our treatment dummy W Bi . The estimated drop in credit supply at the …rm level of 5.3 pp in col. (1) of Table 4 indicates that treated …rms managed to o¤set around two-thirds of the fall in credit supply by weak banks. The corresponding estimate for multi-bank …rms is -3.1 pp (cols. 2). Using the weak bank measure based on REI exposure yields a smaller e¤ect on credit, although it is not statistically di¤erent from the baseline estimate (col. 3).17 Furthermore, while multi-bank …rms su¤ered a stronger credit supply contraction at the local level than the average client …rm of weak banks, the reverse is true at the …rm level. A pre-crisis banking relationship with more than one bank thus provided some insurance against the shocks that subsequently hit weak banks. The average change in credit is equal to -23.1% for unattached …rms and to -31.3% for attached …rms. Out of this 8.2 pp di¤erence, 2.8 pp are due to the attachment to weak banks, which therefore explains 34% of the fall in credit for attached …rms. The large di¤erence between the bank-…rm and …rm-level estimates may seem surprising, but weak banks may have predominantly severed their relationship with those …rms with a marginal attachment to weak banks that are included in the control group. This would be consistent with our estimates for the …rm-level e¤ects being close to the estimated e¤ects at the bank-…rm level for continuing relationships (Table 3, col. 5). It is also worth noting that our results are at variance with those in Jiménez et al. (2014), who –using the same CIR data–…nd a positive credit shock for banks that securitized mortgages in the years of the Spanish boom (2004Q4 to 2007Q4), but then …nd no transmission of this positive shock at the …rm level. At that time the credit market was booming and acceptance rates for loan applications were high for all banks, as shown in Section 3.3. On the contrary, the steep fall in acceptance rates during our sample 17 If there was no o¤setting, …rms attached to weak banks should su¤er a reduction that is proportional to their share of loans from weak banks in total loans, which is 68.5% (Table 2). Since non-attached …rms are also exposed, with a share of 8.5%, the di¤erential degree of exposure of attached …rms is 60%. Thus, in the no-o¤setting case their di¤erential credit reduction would be 15.3 pp (25.5 pp times 60%). 20 period made it much harder for …rms to o¤set credit rationing by weak banks through new loans from healthy banks. Our results are qualitatively similar to those in Cingano et al. (2016), who use data on Italian …rms to estimate the real e¤ects of the bank lending channel exploiting the 2007 liquidity drought in interbank markets as a source of variation in banks’ credit supply. The comparison is not straightforward, since they de…ne treatment based on the ratio between a bank’s interbank market loans and its asset value. Contrary to Jiménez et al. (2014) and to our case, they …nd similar coe¢ cient estimates at the bank-…rm and …rm levels. Their estimate implies that a 1 pp increase in the interbank-to-asset ratio leads to a reduction in credit growth of 2.4 pp for a …rm with a degree of exposure that is 1 standard deviation above the mean. Gobbi and Sette (2014) also study credit growth at the …rm level for Italian …rms during the Great Recession, …nding that the number of banking relationships has a negative impact on credit growth, with the e¤ect being strongest when …rms move from one to two banks. In contrast, in the bank-…rm analysis we …nd that the negative e¤ect of weak-bank attachment is larger for multi-bank …rms than for single-bank …rms, whereas in the …rm-level analysis the opposite holds. The latter is the most relevant evidence regarding the e¤ect of the credit crunch, and when estimated for all …rms (Table 4, col. 1) the coe¢ cient on a control variable for the number of banking relationships is equal to 0.022 (s.e. 0.006), which means that single-bank …rms fared worse in terms of getting credit, and more so if attached to weak banks.18 7 Main results This section presents the empirical results for our baseline speci…cation for the employment e¤ects of credit constraints and for a set of robustness checks. 7.1 Di¤erence in di¤erences Table 5 presents the estimation results for our baseline DD equation (3). We report robust standard errors corrected for multiclustering at industry, municipality, and main bank level. In order to illustrate the sensitivity of our estimates to changes in the set 18 The full estimation details are available upon request. 21 of control variables, we subsequently add more controls until we arrive at our baseline speci…cation. If we only include industry and municipality …xed e¤ects, we obtain that employment in …rms attached to weak banks falls by 7.4 pp relative to employment in unattached …rms, while allowing for di¤erential trends at the industry times municipality level leads to a treatment e¤ect of -7.6 pp (cols. 1 and 2). Next, including the only three …rmlevel controls that are signi…cant according to the normalized di¤erence test –namely capitalization, bank debt, and the number of banks– reduces the treatment e¤ect to 3.5 pp, while adding all remaining …rm controls brings down the e¤ect marginally to -2.8 pp (cols. 3 and 4), with half of the di¤erence being due to the inclusion of the next two variables with the largest normalized di¤erence, i.e. liquidity and number of past loan applications (not shown). Thus, while the estimate in col. (3) is signi…cantly di¤erent from the one obtained without any controls (col. 2), further including the remaining 14 control variables does not yield a statistically di¤erent estimate. Finally, including main bank …xed e¤ects does not alter the results (col. 5), which is further proof that our weak bank indicator captures the relevant dimensions that explain the reduced access to credit for treated …rms. We therefore adopt the speci…cation in col. (4) as our baseline. Our identi…cation relies on the assumption of parallel pre-crisis trends for treated and control …rms. The validity of this assumption is tested by running a placebo regression with 2002 as the pre-crisis year and 2006 as the post-crisis year. As required, this speci…cation test delivers a coe¢ cient that is not signi…cantly di¤erent from zero (col. 6).19 Further evidence is provided in Figure 5. It depicts the estimated coe¢ cients of W Bi and the con…dence intervals for the period 2002-2010 and it shows that the treatment e¤ect is signi…cantly negative from 2008 onwards. Before that time, weakbank attachment does not produce signi…cant di¤erences in employment in …rms in the treatment and control groups. Hence, the timing of the real e¤ects coincides with the timing of the credit constraints at the local level (Figure 4a). Credit rationing at the …rm level also follows the same pattern (Figure 4b), but these e¤ects are less precisely estimated, which helps to explain why the treatment e¤ect at the …rm level does not 19 We also estimated the same model for every year from 2002 to 2005 getting the same result. 22 become signi…cant until 2009. The coincidence between the timing of the changes in credit supply and employment is reassuring, but not su¢ cient to establish a causal relationship. We need to demonstrate that the reductions in credit supply drive the di¤erential evolution of employment at the …rm level. Moreover, it would be incorrect to limit the analysis to a year-to-year comparison between the extent of credit rationing and the size of the employment adjustment. When the crisis erupted –and in particular after the fall of Lehman Brothers in September 2008–expectations about access to credit changed dramatically and …rms in the treatment group may have rationally anticipated a further tightening of credit conditions in later years. To some extent, real e¤ects may therefore be observed before actual credit rationing shows up in the data. Indeed, in a survey of banks undertaken by the ECB, the net balance of banks expecting an increase in the supply of credit to non-…nancial …rms and banks expecting a decrease went from roughly zero in 2007Q2 to -40% already in 2007Q4, remaining there for the subsequent four quarters (Martínez-Pagés, 2009). A similar survey of …rms was launched by the Bank of Spain in March 2009 (Banco de España, 2009). When asked about their ability to obtain funding from banks over the preceding six months, 40% of …rms up to 50 employees –which are the majority in our sample–, reported that funding was obtained only in part or from credit institutions other than their usual ones, and 30% reported that the could not obtain any bank credit. In these two groups of …rms, 65% reported that the main reason for not obtaining the funding was a change in attitude of credit institutions. Taking due account of the above observations, we now proceed with a formal test of the direct link between access to credit and changes in employment at the …rm level using an IV setup that spans our entire sample period 2006-2010. 7.2 The credit channel We now decompose the estimated e¤ect of the credit supply shock into two parts: the impact of weak bank attachment on the amount of credit obtained by …rms, i.e. a credit volume e¤ect, and the pass-through of this measure of the credit shock to employment, 23 using the following IV model: log (1 + nij ) = log (1 + Creditij ) = log (1 + Creditij ) + Xi0 + + + W Bi + Xi0 + j + vij j + "ij (4) in which W Bi acts as an instrument for access to credit and the …rst stage coincides with equation (2). Therefore captures the di¤erential impact of weak bank attachment on committed credit, while captures the pass-through from credit to employment. Thus, the product is equivalent to parameter in equation (3). The exclusion restriction is that working with a weak bank alters employment growth only through credit. Although the di¤erence between the average interest rate charged by weak and healthy banks is quite small, it is non-zero and higher for weak banks. The presence of an interest rate response, albeit small, would contradict the exclusion restriction and as a result our second stage coe¢ cient should be interpreted as an upper-limit. Table 6 reports the estimates. The …rst stage coincides with our equation for credit rationing at the …rm level. For the entire sample of …rms it delivers a di¤erential drop in credit due to weak-bank attachment of 5.3 pp, while the elasticity of employment with respect to credit is estimated at 0.519 (col. 1), i.e. that the elasticity of employment to the volume of credit is about one-half. This yields a compound e¤ect on employment of -2.8 pp, which coincides with the baseline of the previous section. Next, for multi-bank …rms we obtain a smaller impact of W Bi on credit growth, -3.1 pp, but interestingly the pass-through is estimated to be larger than in the full sample, 0.797, yielding an overall impact of -2.5 pp (col. 2). The estimates are highly signi…cant and the F -statistics con…rm the absence of a weak instrument problem. In comparison, Cingano et al. (2016), who explore the e¤ects of the 2007 liquidity drought in interbank markets on Italian …rms, …nd that a 10 pp reduction in credit growth reduces employment by 1.4 pp, whereas the equivalent …gure in our case would be 5.2 pp, indicating a much higher elasticity than theirs. 24 7.3 Alternative speci…cations In this section we perform a wide range of speci…cation tests to check the robustness of our results. We start by considering their sensitivity to alternative de…nitions of the treatment variable. So far we have used a discrete treatment measure and the threshold for assignment to the treatment group was set at the …rst quartile of the distribution of the weak-bank loan-to-asset ratio. In our …rst exercise, we replace the treatment dummy by the ratio itself, which allows the intensity of credit constraints to depend on the normalized size of …rms’debt with weak banks in 2006. The corresponding coe¢ cient, reported in the …rst column of Table 7, is -9.2 pp. Evaluated at the average ratio (22.8%), this delivers an overall e¤ect of -2.1 pp. Then we report the estimates when the threshold for our discrete treatment measure is set, respectively, at the median and the third quartile of the distribution. As we raise the threshold, the estimated treatment e¤ect becomes stronger, going from -3.0 pp to -3.3 pp (cols. 2 and 3). Neither of these estimates is statistically di¤erent from our baseline, but this exercise reveals that the magnitude of the impact increases with the degree of exposure, indicating that our choice of the …rst quartile as the threshold is quite conservative. The aim of the next exercise is to separate employment adjustments along the intensive and the extensive margins. Restricting the sample to surviving …rms, the estimated treatment e¤ect drops to -1.4 pp, which is exactly half the size of our baseline estimate (col. 4). A further robustness check is to use a di¤erent de…nition of the dependent variable that also allows us to account for both exit and entry, originally proposed by Davis et al. (1996) to study establishment-level data, namely (nijt nijt 1 )=(0:5(nijt + nijt 1 )). The associated coe¢ cient is -3.4 pp, which is larger but not statistically di¤erent from our baseline estimate (col. 5). We now consider alternative procedures to control for local demand e¤ects. Mian and Su… (2014) argue that local demand e¤ects should only a¤ect output in non-traded goods sectors, whereas credit supply shocks should a¤ect traded good sectors as well. We therefore aim at …ltering out local demand e¤ects by restricting attention to traded sectors. Mian and Su… (2014) use two classi…cations, based on either ad-hoc tradability criteria or geographical concentration. We choose the latter, because more concentrated 25 industries are likely to be more traded and hence less dependent on local demand conditions.20 We follow these authors in computing the Her…ndahl concentration index for 3-digit industries and 50 provinces, and label as tradable the goods in the highest quartile. This sample selection yields an e¤ect on employment of -5.8 pp (col. 8). It is statistically di¤erent from our baseline estimate, presumably because these …rms sell in a wider geographical area and may therefore rely more on bank credit or be more sensitive to changes in credit supply, or be more sensitive to the cycle, or a combination of these factors. For our purposes what matters is that these estimates are not the result of local demand shocks. We next check the impact of the alternative de…nition of weak bank, already used in Section 6.1, where weak banks are de…ned as those in the upper quartile of the distribution of exposure to the REI. For this de…nition the measured impact is equal to -3.0 pp, which is again very similar to our baseline (col. 7). In the two …nal checks we alter the reference period. First, we rede…ne the pre-crisis year to 2007. This choice is motivated by the fact that aggregate employment in Spain kept growing until the third quarter of 2007. Surprisingly, the estimated weak-bank e¤ect drops to -1.9 pp (col. 8). This result suggests that the slowdown in 2007 altered the mix of employment in …nancially vulnerable and resilient …rms in the treatment and control groups, though once again this estimate is not di¤er statistically from the baseline. Lastly, we measure weak-bank attachment and all other variables in 2002 (col. 9). Hence, …rms in this sample are at least …ve years old at the start of the crisis. The table shows that the treatment e¤ect survives, though at -1.5 pp it is signi…cantly smaller than for the 2006 sample, suggesting that older …rms were less a¤ected by the credit supply shock, which will be con…rmed in Section 9. 8 Selection Our baseline speci…cation includes an exhaustive set of controls for observable …rm characteristics, many of which moreover do not signi…cantly di¤er across exposed and non-exposed …rms. However, this does not completely rule out the possibility of selection e¤ects. Our list of …rm controls may still be incomplete and our estimation 20 As found by Mian and Su… (2014) for the US and by Ramos and Moral-Benito (2015) for Spain. 26 strategy does preclude selection on unobservables. Selection on unobservables would not be a concern if the same unobservables that are relevant for credit demand fully captured the unobservable demand e¤ects relevant for employment growth during the crisis. There are however no strong reasons to believe that this should hold. For example, …rms facing a low product demand may demand less credit to the extent that they need to produce less; however they may also have higher demand for credit since they have lower cash ‡ow and may need more resources to pay back outstanding liabilities. Given the importance of the selection e¤ects we devote Section 8 to this issue. We can informally check the sensitivity of our estimates to the inclusion of the observable controls, so as to derive bounds for the possible bias arising from unobservable variables, as in Oster (2017). If the value of the regression R2 increases when the controls are included but the coe¢ cient of interest does not vary much, then it is expected that the inclusion of unobservables would not alter it either. We compute a bias-adjusted estimated coe¢ cient, following Oster (2017) in making the heuristic assumption that the maximum R2 would be 30% higher than the R2 that would be obtained if all potential determinants were included. The estimate of the e¤ect of W Bi is equal to -1.1 pp, which places a lower bound on the e¤ect of interest. More formally, we now perform three further tests to corroborate our claim that the di¤erential evolution of …rm-level employment is not driven by selection. 8.1 Panel estimates Our DD model is based on a cross-section and cannot therefore include …rm …xed effects. To rule out the di¤erential evolution of employment being driven by unobservable characteristics, we estimate the following panel …xed e¤ects model (Wooldridge, 2010): log(1 + nijt ) = where 0 i 0 i + W Bi0 dt 0 + Xi0 dt 0 + dt j + dt + vijt (5) is a set of …rm …xed e¤ects, dt a vector of time dummies for t=2007,...,2010, and vijt a random shock. The rest of the variables are de…ned as before. This model includes industry times municipality times year …xed e¤ects and both the treatment dummy and the vector of time-invariant …rm characteristics are interacted with year 27 dummies. The equivalent of in equation (3) is the element of the coe¢ cient vector 0 corresponding to 2010 –whose value is relative to 2007. As reported in Table 8, in this panel …xed e¤ects speci…cation the treatment e¤ect amounts to -2.7, which is indistinguishable from the baseline. Interestingly, the treatment e¤ect is statistically signi…cant in 2008 and monotonically increasing in absolute value over time. These estimates indicate that unobservables do not play a signi…cant role in the transmission of the credit supply shock once we …lter out any trends at the industry-municipality level and across …rms with di¤erent observable characteristics. 8.2 Matching estimates We have already pointed out the presence of some degree of heterogeneity between treated and control …rms. Matching techniques allow us to directly compare similar …rms in both groups. This avoids problems derived from a possible lack of overlap between the characteristics of …rms in the two groups and it improves the e¢ ciency of our estimates. For the sake of completeness, we use both propensity score and exact matching for our discrete and continuous treatment variables, thus obtaining four di¤erent estimates. The propensity score matching estimates are derived from …rst estimating a probit model for the probability that a …rm borrows from a weak bank –which includes the same controls as the baseline regression– and then estimating our baseline model using the weights coming from the sample balanced on all the observables used for the propensity score. In exact matching we compare treated and non-treated …rms within industry times municipality and …rm control cells. For the latter we use the coarsened exact matching method (Iacus et al., 2011) where all characteristics are entered as 01 dummy variables (see Appendix 2 for details). We end up with 6,556 strata with observations that can be matched across treated and control …rms, out of a total of 13,520 strata, so that only 2,122 …rms (5.1%) in the treatment group are left without a matching control …rm, and the treatment e¤ect is estimated using the method of weighted least squares. Table 9 reports the results. For the discrete treatment measure W Bi , the estimated e¤ects with propensity score and exact matching are, respectively, -3.2 pp and -1.6 pp 28 (cols. 1 and 2).21 With the continuous measure, the corresponding coe¢ cients are -6.5 pp and -5.2 pp. Since average exposure among attached …rms is 22.8%, the average treatment e¤ects are, respectively, -1.5 pp and -1.2 pp (cols. 3 and 4). Thus, in both cases propensity score matching delivers larger e¤ects and, in line with our previous results, the continuous measure implies somewhat lower e¤ects than the discrete treatment dummy. Importantly, all four estimates are signi…cantly di¤erent from zero and they lie within the con…dence intervals of their respective baseline estimates. 8.3 Exogenous variation in weak-bank attachment Ultimately, weak bank attachment is an endogenous choice. Private …rms may not have been able to predict the solvency problems of weak banks during the crisis, but the most vulnerable …rms may have ended up working with weak banks due to relatively lax credit standards or poor risk management at weak banks. Thus, to further try to rule out selection e¤ects we need an exogenous source of variation in weak-bank attachment. For this purpose we exploit a change in banking regulation. Until 1988 savings banks could not open more than 12 branches outside their region of origin, but at the end of December 1988 all location restrictions were lifted (Real Decreto-ley 1582/1988). We calculate for each municipality the share of bank branches at the start of December 1988 –i.e. right before the adoption of the reform– that belonged to any of the weak banks and we use this indicator for local weak-bank density as an instrument to explain the weak-bank attachment of …rms in 2006. In other words, we are assuming that …rms are more likely to work with weak banks when these banks have traditionally held a strong market position in the municipality of the …rm as re‡ected by the local weak-bank density in 1988. The exclusion restriction is that local weak-bank density only a¤ects a …rm’s employment through its attachment to weak banks, which cannot be tested. Having neighboring …rms that are more likely to have borrowed from a weak bank may a¤ect a given 21 Post-matching balance tables, for both exact and propensity score matching, show no trace of signi…cant di¤erences in the control variables across treated and control …rms according to the normalized di¤erence test of Imbens and Wooldridge (2009). The mean distance for each of the matching variables is smaller than 3:2e 14 and the L1 statistic, computed by the cem command in Stata and introduced in Iacus et al. (2012), which is a comprehensive measure of global imbalance based on the L1 di¤erence between the multidimensional histogram of all pre-treatment covariates in the treated group and the control group, has a value of zero in both cases. These results are available upon request. 29 …rm, which would violate the exclusion restriction. Using geography-based variation therefore means that the estimates may include local general equilibrium e¤ects from having a weak bank. We can only imperfectly control for local demand e¤ects because our IV varies by municipality and so we cannot include trends at that level. However, we do include a dummy for Spain’s coastal provinces among the regressors, to capture the comparatively strong growth in credit and housing prices in these provinces prior to the crisis.22 To be a valid instrument, local weak-bank density in 1988 would have to be as good as randomly assigned vis-à-vis the set of …rms that existed in that area in 2006. In terms of our control variables, …rms in municipalities with local weak-bank density above the median (0.25) are very similar to …rms below the median in all respects. As can be seen in Table 10, none of the …rm characteristics, other than the share of loans with weak banks, is signi…cantly di¤erent across the two groups of …rms according to the normalized di¤erence test of Imbens and Wooldridge (2009). It should be stressed that the overwhelming majority of weak bank branches in December 1988 belonged to savings banks that were founded in that particular region. In other words, our IV analysis uses weak bank attachment in the home region of the savings banks. This point is relevant because the available evidence suggests that the expansion of the savings banks beyond their region of origin was accompanied by a deterioration of their client pool (Illueca et al., 2013). The estimates in Table 11 show that high weak-bank density in 1988 signi…cantly predicts weak-bank attachment 18 years later and that the associated employment e¤ect amounts to -7.6 pp with the W B dummy variable and -7.3 pp with the continuous measure. Both estimates are larger than the foregoing ones, but they should not be taken at face value due to the above-mentioned caveats about the potential violation of the exclusion restriction. 9 Financial vulnerability The literatures on relationship lending and …nancial accelerators indicate that smaller, less transparent, and …nancially weaker …rms should be more vulnerable to changes 22 They include provinces along the Mediterranean Coast and in the Balearic and Canary Islands. 30 in credit market conditions. To …nd out if these features alter the impact of credit constraints, we estimate a triple di¤erence (DDD) model, again implemented with fouryear di¤erences. We interact the treatment dummy with an exhaustive set of …rm controls that capture the …rm’s credit history (captured by past defaults and rejected loan applications), …nancial strength (own funds and liquidity), bank dependence (bank debt over total debt), short-term liabilities (the share of bank debt that expires within 12 months), credit lines, number of banks (single- or multi-bank), productivity (service sector and export status), and other variables like age and size, that are traditionally used to proxy for the strength of agency problems.23;24 The estimation results appear in Table 12. As expected, having a bad credit record, as evidenced by rejected loan applications and, especially, past loan defaults, entails higher job losses during the crisis. More traditional …nancial indicators also attract the expected signs. The contraction of employment is stronger for …rms with more debt and a higher share of short-term debt, and lower capital and liquidity. All these e¤ects are stronger for …rms exposed to weak banks, except for the case of liquidity. Lastly, …rms working with a single bank reduced employment less than …rms with multiple banking relationships, but as in the case of liquidity there are no signi…cant di¤erences between …rms in the treatment and control groups. As to other results, larger …rms su¤er slightly lower losses and this marginal e¤ect is the same for …rms in both groups of …rms. Older …rms, on the contrary, su¤er higher losses, but the e¤ect is cut by half for …rms attached to weak banks. The positive sign on export status, both alone and interacted with W Bi , may have a couple of explanations. Exporting …rms may have su¤ered a smaller drop in demand than …rms that produce for the Spanish market –even though international trade also experienced a sizeable but short-lived drop at the start of the global …nancial crisis. Indeed, while real internal demand fell by 3.1% from 2006 to 2010, real exports increased by 4.5%. Alternatively, export status may capture cross-sectional di¤erences in productivity –even 23 The dummy variable for any application rejected is the complement of the one for all applications accepted in the baseline. The dummy for short-term debt above 50% replaces the share of short-term debt that was present in the baseline. The dummy for …rms indebted with only one bank replaces the variable for the number of banking relationships. 24 To avoid having to weigh estimates by the variables’ average values, regressors are in deviations from their means. 31 within narrowly de…ned sectors, the most productive …rms are typically the ones that manage to sell their products in international markets, while the least productive …rms only sell in the domestic market. As expected, …rms with a higher share of temporary jobs su¤er higher job losses and the e¤ect is stronger for …rms in the treatment group. Lastly, we do not observe any signi…cant di¤erence by broad sector. 10 Margins of adjustment and aggregate e¤ects We end by probing deeper into the margins of adjustment available to …rms. Our analysis in the previous section quanti…ed the additional job losses due to the creditrationing by weak banks, but it is also important to know what types of jobs are most at risk when …rms face credit constraints and whether …rms have explored alternative margins of adjustment, like changes in wages, to alleviate the impact on employment. Lastly, we estimate the causal impact of weak-bank attachment on …rm exit. 10.1 The contribution of temporary jobs to job losses Our DDD estimates indicate that labor shedding was higher the larger the shares of temporary workers that …rms had in 2006. To further gauge the importance of temporary employment in the transmission of the credit shock, we analyze how weak-bank attachment a¤ected the share of temporary workers at surviving …rms.25 The breakdown of employment by type of contract is observed for 91% of these …rms. The …rst column of Table 13 shows that for this sample the temporary employment share fell by -0.5 pp between 2006 and 2010, as a result of the stronger credit constraints faced by the treated …rms. Since weak-bank attachment caused a 1.4 pp drop in total employment for this sample of …rms (Table 7) and the initial share of temporary jobs among surviving …rms was equal to 21%, this means that temporary employment fell by 3.7 pp, contributing 0.8 pp to the overall employment loss in this sample of treated …rms. In other words, temporary jobs made up around one-quarter of pre-crisis employment but accounted for 56% of employment adjustment in treated …rms.26 25 Financial market frictions may also have distorted the ex-ante mix of contract types in favor of temporary jobs, as shown in Caggese and Cuñat (2008), but this issue is outside the scope of our analysis. 26 Berton et al. (2016) …nd a similar e¤ect for …rms in the Italian region of Veneto. 32 The over-representation of temporary workers in job losses is also observed in the aggregate (Bentolila et al., 2012) and is commonly attributed to the lower …ring costs of these contracts. Nevertheless, in the recent crisis credit rationing may also have played a role. Weak banks curtailed the funding of working capital through credit lines more than healthy banks. It is therefore natural to expect …rms in the treatment group to concentrate their downsizing among temporary workers. 10.2 Changes in the wage bill Our next experiment analyzes how weak-bank attachment a¤ected the evolution of the total wage bill. What we would like to estimate is to what extent …rms in the treatment group managed to alleviate the stringency of the credit constraints through a reduction in wages. Since we do not observe individual wages and the information on the characteristics of workers in our data is minimal, however, we have to conduct the analysis on the basis of the total wage bill of surviving …rms. Our estimate, in the second column of Table 13, indicates that weak-bank attachment is associated with a 1.6 pp drop in the wage bill of the treated …rms. Accordingly, we conclude that the average wage of the employees of treated …rms fell by 0.2 pp compared to the corresponding wage of the employees of the …rms in the control group. The di¤erential drop in the average wage is signi…cant but also very small, and it could be driven by composition e¤ects rather than by an adjustment of nominal wages, since we cannot control for worker characteristics. At any rate, our results indicate that wage adjustments have not played a meaningful role in alleviating the impact of credit constraints faced by Spanish …rms during the crisis, which is consistent with the well-documented high real wage rigidity in Spain (Font et al., 2015). 10.3 Probability of …rm exit After a recession, employment may recover more slowly if a large fraction of …rms have close down than if they have only downsized. For this reason we estimate the e¤ect of credit constraints at the intensive and extensive margins. The former is given by reestimating our baseline DD equation (3) for surviving …rms alone, which as already reported yields a coe¢ cient of -1.4 pp (Table 7, col. 4), a signi…cantly lower …gure 33 than for the full sample. The extensive margin is tackled by estimating the e¤ect of weak-bank attachment on …rms’ exit probability. We start with a linear probability model for exit in 2010 with respect to 2006, using the same speci…cation as in equation (3) but now for a binary dependent variable. As seen in the …rst column of Table 14, weak-bank exposure leads to a marginal increase in the exit probability of weakbank dependent …rms of 1.1 pp, which represents an increase of 10.8% with respect to the baseline exit rate of 10.2%. We also try an alternative speci…cation including the continuous treatment variable in place of the dummy variable. The estimated e¤ect is 5.9 pp. This means, ceteris paribus, that a …rm located at the ninth decile –with a ratio of weak-bank debt to assets of one-quarter– has a 1.5% higher probability of closing down, compared to a …rm with a ratio of weak-bank debt to assets at the …rst decile –which is roughly nil. This amounts to 14.5% of the baseline exit rate. 10.4 Aggregate e¤ects Are the foregoing estimates large or small? We should clarify that these microeconomic estimates cannot be directly extrapolated to the aggregate economy. In general equilibrium there are further potential e¤ects (Chodorow-Reich, 2014). For example, a drop in aggregate demand generally reduces labor demand by both constrained and unconstrained …rms, but product demand may be shifted from the former to the latter, thus inducing an increase in their labor demand. With this caveat in mind, we can estimate job losses due to weak-bank attachment for each individual …rm and then add them up over all …rms in the sample.27 Using the estimate from our DD baseline estimate of 2.8 pp (Table 3, col. 4), exposure to weak banks accounts for 24.4% of the total fall in employment among exposed …rms in our sample. Given the employment share of exposed …rms, this represents 7% of total job losses in our full sample. If we use the lower bound resulting from the Oster (2017) conjecture, which yielded an estimate of 1.1 pp, then the …gures would respectively be 9.8% and 2.8%. Using our estimates we can compute separately employment losses at surviving 27 From equation (1), (1 + nijt )=(1 + nijt 4 ) = exp( 4 log(1 + nijt )), where t =2010. The estimated employment growth rate is then equal to: 4 n bijt = (1 + nijt 4 )[exp(b + b W Bi + Xi b + bj ) 1]. Estimated job losses due to weak-bank attachment then equal: ( 4 n bijt j W Bi = 1) ( 4 n bijt j W Bi = 0) = (1 + nijt 4 )[exp(b + b + Xi b + bj ) exp(b + Xi b + bj )]. 34 and closing …rms. On the one hand, we can use the previously cited estimate for the surviving …rms. On the other hand, we calculate the number of …rm closures from the estimated probability of exit in Table 14 (col. 1) and the employment reduction so induced. Adding up both estimates, the overall job loss at exposed …rms is 33.8%, which is higher than the baseline estimate, so that credit constraints explain 12% of total job losses in the sample. Survivors account for 48% of this overall loss and the rest corresponds to exiting …rms. However, the estimated e¤ect on each one di¤ers markedly. Weak-bank exposure accounts for a full 54.2% of job losses at surviving …rms, whereas it explains 33.8% of jobs lost due to …rms closing down. The latter suggests that shocks other than credit supply restrictions play a larger role in the exit decision of …rms. 11 Conclusions In this paper we have analyzed the link between the solvency problems of Spain’s weakest banks and the severe drop in employment su¤ered by this country during the Great Recession. We achieve identi…cation by exploiting di¤erences in lender health at the start of the crisis, as evidenced by public bailouts of savings banks. We proceed by comparing employment changes from the expansion to the recession between …rms that are exposed to weak banks and those that are not. Our exercise is challenging, since the bank solvency problems are linked to corporate loan portfolios. While we are not the …rst to study the link between external funding and employment outcomes, we do provide the …rst exhaustive analysis of these links on the basis of loan data from an o¢ cial credit register. For practical purposes this data set can be considered as the census of loans to non-…nancial …rms of all sizes, but mainly to the small and medium-sized, for which credit restrictions are strongest according to standard theory. Our exceptionally large and high-quality matched bank-loan-…rm data set allows us to control exhaustively for ex-ante characteristics of …rms and for potential endogeneity, as well as to perform a wide range of robustness checks. It also allows us to obtain more precise estimates and to re…ne the analysis in more directions than any existing study in the …eld. Our results show that the …rms attached to weak banks indeed destroyed more 35 jobs than very similar …rms working with healthier banks. For the average …rm, the additional job losses due to weak-bank attachment are around 2.8 percentage points. This estimate implies that weak-bank exposure accounts for around 24% of the total fall in employment among exposed …rms in our sample. The extraordinary strength of the credit crunch in Spain is illustrated by the …nding of sizeable e¤ects even for the largest …rms in our sample, whereas the evidence for the US only points to employment losses at the smallest …rms. Furthermore, our analysis uncovers striking di¤erences in the intensity of credit restrictions depending on …rms’ creditworthiness and the structure of their banking relationships. We also show that the brunt of the job losses due to credit constrains was borne by temporary employees, with little adjustment in wages. Separate estimates for employment losses at surviving and closing …rms indicate that for survivors 52% of job losses at exposed …rms are explained by weak-bank attachment, while 34% of the losses originated by …rm closures are. Our paper is the …rst to o¤er this type of decomposition, which carries relevant implications for the speed of recovery after slumps. 36 A Appendix A. Additional tables Table A.1. Spanish savings banks’restructuring process (end of year) Note. SIP: Institutional Protection System (see Section 3). Source: Banco de España (2014). 37 Table A2. Returns on securities issued by Spanish banks in 2006 Dependent Variable: Coupon di¤erential (basis points) Weak Bank Mortgage Backed Security Years to Maturity Investment Grade (AA+ to BBB-) Speculative Grade (BB+ to D) Collateralized Debt Obligation Customer Loan Corporate Loan Residential Mortgage No Guarantor Private Sector Bank Guarantor State/Provincial Authority Guarantor Supranational Guarantor R2 Coe¢ cient 2.84 15.55 0.83 24.37 131.01 0.32 2.76 5.55 -18.90 -5.65 13.33 -4.41 4.65 St. error 4.74 0.29 0.13 2.35 25.17 17.61 7.95 14.16 8.82 6.96 13.43 10.56 5.43 0.44 255 No. of observations Notes. OLS estimates of coupon di¤erentials of all asset and mortgage backed securities issued by Spanish banks in 2006 with reference to the 3-month Euribor. Reference group: Asset Backed Security, Prime Risk (AAA), Auto Receivables as collateral, Central Government as guarantor. Data for 24 issuer parents drawn from Dealscan. Month of issue dummies are included. Standard errors are adjusted for 24 clusters in issuing bank. 38 Table A3. Credit rationing at the …rm-bank level. Loan maturity Dependent variable: 4 log (1 + Creditijb ) I (Credit Lineib ) I (Credit Lineib ) W B i I (Maturityib 1-3 years) I (Maturityib 1-3 years) W B i I (Maturityib 3-5 years) I (Maturityib 3-5 years) W B i I (Maturityib 5 years) I (Maturityib 5 years) W B i I (Maturityib indeterminate) I (Maturityib indeterminate) W B i Firm …xed e¤ects Firm controls Bank …xed e¤ects Bank controls Firm-bank controls Industry Province …xed e¤ects Several banks Balance-sheet data R2 0.057 -0.082 -0.035 0.045 -0.115 0.034 0.012 0.091 0.201 0.170 (0.016) (0.040) (0.020) (0.046) (0.022) (0.048) (0.021) (0.048) (0.147) (0.247) yes – yes yes yes – yes yes 0.453 236,691 72,287 No. obs. Notes. OLS estimates for 2010. Bank controls: log of total assets, leverage ratio, liquidity ratio, rate of return on assets and provisions normalized by net interest income. Firm-bank controls: length of …rm-bank relationship in months and past defaults. Firm control variables: see Table 2. “yes/no/–”indicates whether the corresponding set of variables is either included, not included or redundant. Robust standard errors corrected for multiclustering at the …rm and bank level appear between parentheses. Signi…cance levels: * p<0.10, ** p<0.05, *** p<0.01. 39 B Appendix 2. De…nitions of variables Employment. Computed as the average level over the year, weighing temporary employees by their weeks of work. Treatment. The weak-bank dummy (0-1) is equal to 1 if the ratio of the …rm’s loans from weak banks to total assets in 2006 is above the …rst quartile of the distribution. Weak Bank Intensity is the ratio itself. Municipality (2,749). They correspond to …rms’headquarters. They need to have at least two …rms in the sample. Industry (78). Excluded industries (share of output sold to Construction and Real Estate in 2000 shown between parentheses): Extraction of Non-metallic Minerals (35.2%), Wood and Cork (21.1%), Cement, Lime, and Plaster (46.4%), Clay (60.1%), Nonmetallic Mineral Products n.e.c. (85.4%), Fabricated Metal Products except Machinery and Equipment (23.3%), Machinery and Electric Materials (19.2%), and Rental of Machinery and Household Goods (26.2%). Industry dummies (…rms’ main activity in 2006): Standard 2-digit NACE rev. 1.1 classi…cation, see www.ine.es/daco/daco42/clasi…caciones/cnae09/estructura_en.pdf. Control variables (stocks are book values in December). Temporary Employment (temporary employees/total number of employees), Age (current year–year of creation), Size (Total assets in million euros), Exporter (indicator for selling abroad), Own Funds (own funds/total assets), Liquidity (liquid assets/total assets), Return on Assets (earnings before interest, taxes, depreciation and amortization/Assets), Bank Debt (bank debt/total debt), Short-Term Bank Debt (debt up to one year/total bank debt), LongTerm Bank Debt (debt of …ve years or more/total bank debt), Non-Collateralized Bank Debt (non-collateralized loans/total bank debt), Credit Line (1 if at least one), Banking Relationships (number of banks with outstanding loans), Current Loan Defaults (1 if any non-performing loan in 2006), Past Loan Defaults (1 if any non-performing loan in 2002-2005), Past Loan Applications (number in 2002-2005), and All Loan Applications Accepted (0-1 dummy). Cells for matching estimation. Province (1 if East coast of Spain and Balearic or Canary Islands). Industry (1 if Agriculture, Farming, Fisheries, and Extractive). Value of 1 if above the median: Temporary Employment, Age, Size, Own Funds, Liquidity, Return on Assets, Bank Debt, Non-Collateralized Bank Debt, and Past Loan Applications. Value of 1 if above 50%: Short-Term Bank Debt and Long-Term Bank Debt. Value of 1 if variable equal to 1: Banking Relationships. Already 0-1 dummies in baseline speci…cation: Exporter, Current Loan Defaults, Past Loan Defaults, Credit Line, and All Past Applications Accepted. 40 Acknowledgements Bentolila is also a¢ liated with CEPR and CESifo, Jansen is also a¢ liated with Fedea and IZA. This paper is the sole responsibility of its authors. The views presented here do not necessarily re‡ect those of the Banco de España or the Eurosystem. We are very grateful to three anonymous referees and to Joshua Angrist, Manuel Arellano, Stéphane Bonhomme, Gabriel Chodorow-Reich, Juan J. Dolado, Markus Demary, David Dorn, Pietro Garibaldi, Tullio Jappelli, Juan F. Jimeno, José Liberti, Fabian Lange, Pedro Mira, Rafael Repullo, Javier Suarez, John van Reenen, Ernesto Villanueva, Etienne Wasmer. We are also grateful for comments by seminar audiences at BBVA Research Department, Banco de España, CEMFI, European Central Bank, and IMT Lucca; at conferences at Banco de Portugal, CSEF (Naples), De Nederlandsche Bank, IMÈRAAMSE (Marseille), IZA/CEPR European Summer Symposium on Labor Economics, Kiel Institute for the World Economy, and Simposio de la Asociación Española de Economía; and at the following universities: Alcalá de Henares, Alicante, Autònoma de Barcelona, Autónoma de Madrid, Ca’ Foscari Venezia, Complutense de Madrid, European University Institute, and Mainz. We also wish to thank Ana Esteban and José I. González-Pascual, from the Statistics O¢ ce of the Banco de España, for help with the data. All errors are our own. Jansen acknowledges funding from the Ministerio de Economía y Competitividad Grants ECO2012-37742 and ECO2015-69631-P. 41 References [1] Acharya, V. V., T. Eisert, C. Eu…nger, and C. Hirsch (2016), “Real E¤ects of the Sovereign Debt Crisis in Europe: Evidence from Syndicated Loans”, mimeo. [2] Almeida, H., M. Campello, B. Laranjeira, and S. Weisbenner (2012), “Corporate Debt Maturity and the Real E¤ects of the 2007 Credit Crisis”, Critical Finance Review 1, 3-58. [3] Ayuso, J. and F. Restoy (2006), “House Prices and Rents: An Equilibrium Asset Pricing Approach”, Journal of Empirical Finance 13, 371–388. [4] Balduzzi, P., E. Brancati, and F. Schiantarelli (2015), “Financial Markets, Banks’ Cost of Funding, and Firms Decisions: Lessons from Two Crises”, CESifo Working Paper Series 5669. [5] Banco de España (2009), “Survey of Non-…nancial Corporations on Conditions of Access to Credit”, Economic Bulletin, July, 147-157. [6] Banco de España (2014), “Background Note on Public Financial Assistance in the Restructuring of the Spanish Banking System (2009-2013)”, http://www.bde.es. [7] Banco de España (2015), “Encuesta sobre Préstamos Bancarios en España: Enero de 2015”, Boletín Económico, January, 13-29. [8] Benmelech, E., N. K. Bergman, and A. Seru (2012), “Financing Labor”, mimeo. [9] Bentolila, S., P. Cahuc, J. J. Dolado, and T. Le Barbanchon (2012), “Two-Tier Labour Markets in the Great Recession: France Versus Spain”, Economic Journal 122, F155–F187. [10] Berton, F., S. Mocetti, A. F. Presbitero, and M. F. Richiardi (2017), “Banks, Firms, and Jobs”, IMF Working Paper 17/38. [11] Boeri, T., P. Garibaldi, and E. Moen (2013), “Financial Shocks and Labor: Facts and Theory”, IMF Economic Review 61, 631-663. [12] Caggese, A. and V. Cuñat (2008), “Financing Constraints and Fixed-Term Employment Contracts”, Economic Journal 118, 2013-2046. [13] Chava, S. and A. Purnanandam (2011), “The E¤ect of Banking Crisis on Bankdependent Borrowers”, Journal of Financial Economics 99, 116-135. [14] Chodorow-Reich, G. (2014), “The Employment E¤ects of Credit Market Disruptions: Firm-level Evidence from the 2008-09 Financial Crisis”, Quarterly Journal of Economics 129, 1-59. [15] Cingano, F., F. Manaresi, and E. Sette (2016), “Does Credit Crunch Investments Down? New Evidence on the Real E¤ects of the Bank-lending Channel”, Review of Financial Studies 29 (10), 2737-2773. 42 [16] Cuñat, V. and L. Garicano (2010), “Did Good Cajas Extend Bad Loans? Governance, Human Capital and Loan Portfolios”, mimeo. [17] Davis, S. J., J. Haltiwanger, and S. Schuh (1996), Job Creation and Destruction, MIT Press, Boston, MA. [18] Duygan-Bump, B., A. Levkov, and J. Montoriol-Garriga (2015), "Financing Constraints and Unemployment: Evidence from the Great Recession", Journal of Monetary Economics 75, 89-105. [19] European Central Bank (2010), EU Banking Structures, Frankfurt. [20] Fernandes, A. P. and P. Ferreira (2017), “Financing Constraints and Fixed-Term Employment: Evidence from the 2008-9 Financial Crisis”, European Economic Review 92, 215-238. [21] Fernández-Villaverde, J., L. Garicano, and T. Santos (2013), “Political Credit Cycles: The Case of the Euro Zone”, Journal of Economic Perspectives 27, 145166. [22] Font P., M. Izquierdo, and S. Puente S (2015), “Real Wage Responsiveness to Unemployment in Spain: Asymmetries along the Business Cycle”, IZA Journal of European Labor Studies 4, 1-13. [23] Garicano, L. and C. Steinwender (2015), “Survive Another Day: Using Changes in the Composition of Investments to Measure the Cost of Credit Constraints”, Review of Economics and Statistics, forthcoming. [24] Gobbi, G. and E. Sette (2014), “Do Firms Bene…t from Concentrating their Borrowing? Evidence from the Great Recession”, Review of Finance 18, 527-560. [25] Greenstone, M. A. Mas, and H.–L. Nguyen (2014), “Do Credit Market Shocks A¤ect the Real Economy? Quasi-Experimental Evidence from the Great Recession and ‘Norma EconomicTimes, NBER Working Paper 20704. [26] Guiso, L., F. Schivardi, and L. Pistaferri (2013), “Credit within the Firm“, Review of Economic Studies 80, 211-247. [27] Hochfellner, D., J. Montes, M. Schmalz, and D. Sosyura (2015), “Winners and Losers of Financial Crises: Evidence from Individuals and Firms“, University of Michigan, unpublished manuscript. [28] Iacus, S. M., G. King, and G. Porro (2012), “Causal Inference without Balance Checking: Coarsened Exact Matching”, Political Analysis 20, 1-24. [29] Iacus, S. M., G. King, and G. Porro (2011), “Multivariate Matching Methods that Are Monotonic Imbalance Bounding”, Journal of the American Statistical Association 106, 345-361. [30] Illueca, M., L. Norden, and G. F. Udell (2013), “Liberalization and Risk-Taking: Evidence from Government-Controlled Banks”, Review of Finance 18, 1217-1257. 43 [31] Imbens, G. W., and D. B. Rubin (2015), Causal Inference in Statistics and the Social Sciences, Cambridge, Cambridge University Press. [32] Imbens, G. W. and J. M. Wooldridge (2009), “Recent Developments in the Econometrics of Program Evaluation”, Journal of Economic Literature 47, 5-86. [33] International Monetary Fund (2012), “Spain: The Reform of Spanish Savings Banks. Technical Note”, Monetary and Capital Markets Department, Washington, DC. [34] Jiménez, G., A. Mian, J.-L. Peydro and J. Saurina (2014), “The Real E¤ects of the Bank Lending Channel”, available at http://ssrn.com/abstract=1674828. [35] Jiménez, G., S. Ongena, J.-L. Peydró, and J. Saurina (2012), “Credit Supply and Monetary Policy: Identifying the Bank Balance-Sheet Channel with Loan Applications”, American Economic Review 102, 2301-2326. [36] Jiménez, G., S. Ongena, J.-L. Peydró, and J. Saurina (2016), “Macroprudential Policy, Countercyclical Bank Capital Bu¤ers and Credit Supply: Evidence from the Spanish Dynamic Provisioning Experiments”, Journal of Political Economy, forthcoming. [37] Khwaja, A. I. and A. Mian (2008), “Tracing the Impact of Bank Liquidity Shocks: Evidence from an Emerging Market”, American Economic Review 98, 1413-1442. [38] Martínez-Pagés, J. (2009), “Encuesta sobre Préstamos Bancarios en España”, Boletín Económico, Banco de España, January, 67-76. [39] Mian, A. and A. Su… (2014), “What Explains the 2007-2009 Drop in Employment?”, Econometrica 82, 2197-2223. [40] Oster, E. (2017), “Unobservable Selection and Coe¢ cient Stability: Theory and Evidence”, Journal of Business Economics and Statistics, forthcoming. [41] Paravisini, D., V. Rappoport, P. Schnabl, and D. Wolfenzon (2015), “Dissecting the E¤ect of Credit Supply on Trade: Evidence from Matched Credit-Export Data”, Review of Economic Studies 82, 333-359. [42] Petersen, M. A. and R. G. Rajan (1994), “The Bene…ts of Lending Relationships: Evidence from Small Business Data”, The Journal of Finance 49, 3-37. [43] Petersen, M. A. and R. G. Rajan (2002), “Does Distance Still Matter? The Information Revolution in Small Business Lending”, Journal of Finance 57, 2533-2570. [44] Petrosky-Nadeau, N. and E. Wasmer (2013), “The Cyclical Volatility of Labor Markets under Frictional Financial Markets”, American Economic Journal: Macroeconomics 5, 1-31. [45] Popov, A. A. and J. Rocholl (2016) “Do Credit Shocks A¤ect Labor Demand? Evidence for Employment and Wages during the Financial Crisis”, Journal of Financial Intermediation (forthcoming) 44 [46] Ramos, R. and E. Moral-Benito (2015), “Agglomeration Matters for Trade”, Bank of Spain Research Department, unpublished manuscript. [47] Stiglitz, J. E. and A. Weiss, “Credit Rationing in Markets with Imperfect Information”, American Economic Review 71, 393-410. [48] Santos, T. (2014), “Antes del Diluvio: The Spanish Banking System in the First Decade of the Euro”, Columbia Business School, mimeo. [49] Siemer, M. (2014), “Firm entry and Employment Dynamics in the Great Recession”, Federal Reserve Board, Finance and Economics Discussion Series, Working Paper 2014-56. [50] Wasmer, E. and P. Weil (2004), “The Macroeconomics of Labor and Credit Market Frictions”, American Economic Review, 94, 844-963. [51] Wooldridge, J. (2010), Econometric Analysis of Cross Section and Panel Data, 2nd ed., Cambridge, MA, MIT Press. 45 Table 1. Descriptive statistics of healthy and weak banks (2006) Healthy banks Weak banks Mean Mean St. dev. Mean St. dev. t test ln(Assets) 13.7 2.1 16.4 1.0 7.1 Own Funds/Assets 8.4 9.0 5.2 1.2 -2.0 Liquidity/Assets 23.7 22.4 11.5 4.5 -3.1 Return on Assets 1.0 1.7 0.9 0.3 -0.5 Non-Performing Loans 1.5 6.3 0.7 0.6 -0.7 Non-Performing Loans (2012) 8.6 12.7 22.0 6.0 3.5 Loans to REI/Loans to NFF 36.8 22.3 67.9 8.1 7.9 Securitized Loans/Assets 14.9 10.5 18.5 6.3 1.6 Normalized di¤erence 1.14 -0.35 -0.54 -0.09 -0.13 0.95 1.31 0.30 Notes. There are 206 healthy and 33 weak banks. Non-performing Loans is a ratio on the value of all loans. Securitized Loans/Assets is computed only for banks that securitize. NFF denotes non-…nancial …rms. Except for assets, variables are ratios in percentages. The last column shows the t ratio of the test for the di¤erence of the means and the last column the normalized di¤erence test of Imbens and Wooldridge (2009). See de…nitions in Appendix 2. Source: Own computations on bank balance sheet data from the Bank of Spain. 46 47 P75 0.0 0.0 13.0 33.3 17.0 1.7 0.0 51.1 17.4 10.2 49.8 100.0 37.1 100.0 100.0 2.0 0.0 0.0 100.00 0.0 Mean 22.8 68.5 20.3 22.7 11.7 3.6 13.1 24.9 8.6 5.2 48.5 45.7 29.5 73.7 72.2 3.0 0.6 2.4 68.9 26.2 Treated St. Dev. P25 17.1 9.7 30.3 40.9 207.2 2.0 26.0 0.0 8.7 6.0 27.5 0.3 33.7 0.0 18.5 10.0 11.8 1.1 9.0 2.0 23.5 29.4 37.1 4.3 36.3 0.0 35.6 47.4 44.8 0.0 2.7 1.0 7.4 0.0 15.2 0.0 46.3 0.0 44.0 0.0 P50 17.3 73.3 5.0 14.5 10.0 0.6 0.0 20.8 4.2 4.4 47.4 44.9 7.4 100.0 100.0 2.0 0.0 0.0 100.0 0.0 P75 30.9 100.0 13.0 36.6 16.0 1.7 0.0 35.8 11.2 8.0 66.2 80.7 59.0 100.0 100.0 4.0 0.0 0.0 100.0 100.0 Mean t test 427.3 403.9 -2.7 16.0 -17.1 -4.0 0.3 -74.3 -47.3 -23.8 120.8 -13.2 39.7 -42.3 12.5 103.7 6.8 12.7 52.7 17.6 Norm. di¤. 1.31 1.55 -0.01 -0.06 -0.07 -0.02 0.00 -0.31 -0.20 -0.10 0.50 -0.05 0.16 -0.17 0.05 0.37 0.03 0.05 0.22 0.07 Notes. Observations: 149,458 …rms; 106,128 control and 43,330 treated …rms. WB denotes weak banks. Variables are ratios in percentages unless otherwise indicated. The twelfth column shows the t ratio on the test for the di¤erence of the means and the last column the normalized di§erence test of Imbens and Wooldridge (2009). See de…nitions in Appendix 2. Loans with WB/Assets Share of Loans with WB Employment (employees) Temporary Employment Age (years) Size (million euros) Exporter Own Funds/Assets Liquidity/Assets Return on Assets Bank Debt Short-Term Bank Debt (< 1 yr) Long-Term Bank Debt (> 5 yrs) Non-Collateralized Bank Debt Credit Line (has one) Banking Relationships (no.) Current Loan Defaults Past Loan Defaults Past Loan Applications All Loan Applications Accepted Mean 0.3 8.5 25.3 20.4 12.6 5.8 13.0 34.4 12.4 6.7 30.7 48.8 21.5 81.9 69.0 1.9 0.3 1.4 54.2 22.0 Control St. Dev. P25 P50 0.9 0.0 0.0 24.1 0.0 0.0 365.6 2.0 6.0 25.4 0.0 11.1 9.8 6.0 11.0 118.2 0.3 0.6 33.7 0.0 0.0 23.8 14.2 30.3 15.1 1.9 7.0 11.4 1.8 4.9 26.7 6.8 24.8 41.5 0.0 46.7 35.3 0.0 0.0 33.4 82.7 100.0 46.3 0.0 100.0 1.5 1.0 1.0 5.6 0.0 0.0 11.9 0.0 0.0 49.8 0.0 100.00 41.4 0.0 0.0 Table 2. Descriptive statistics of control and treated …rms (2006) Table 3. Credit rationing at the …rm-bank level Dependent variable: 4 log (1 + Creditijb ) W Bb (1) All …rms -0.232 (0.088) (2) Multibank -0.256 (0.094) (3) Fixed e¤ects -0.255 (0.008) I (Credit Lineib ) R2 No. obs. No. …rms (5) Positive credit -0.079 (0.034) (6) Real estate -0.180 (0.096) 0.074 (0.015) -0.106 (0.039) I (Credit Lineib ) W B b Firm …xed e¤ects Firm controls Bank …xed e¤ects Bank controls Firm-bank controls Ind. Prov. f.e. Several banks Balance-sheet data (4) Interactions no yes no yes yes yes no yes no yes no yes yes yes yes yes yes – no yes yes – yes yes yes – yes yes yes – yes yes yes – no yes yes – yes yes yes – no yes yes – yes yes 0.060 304,089 139,685 0.059 236,691 72,287 0.407 236,691 72,287 0.452 236,691 72,287 0.394 126,863 42,630 0.406 236,691 72,287 Notes. OLS estimates for 2010. Bank controls: log of total assets, leverage ratio, liquidity ratio, rate of return on assets and provisions normalized by net interest income. Firm-bank controls: length of …rm-bank relationship in months and past defaults. Firm control variables: see Table 2. “yes/no/–”indicates whether the corresponding set of variables is either included, not included or redundant. Robust standard errors corrected for multiclustering at the …rm and bank level appear between parentheses. Signi…cance levels: * p<0.10, ** p<0.05, *** p<0.01. 48 Table 4. Credit rationing at the …rm level Dependent variable: 4 log (1 + Creditij ) (1) All …rms -0.053 (0.015) W Bi Firm controls Industry Municipality …xed e¤ects Multiple banking relationships Balance-sheet data R2 (2) Multibank -0.031 (0.011) (3) Real estate -0.039 (0.017) yes yes yes yes yes yes no yes yes yes no yes 0.215 149,458 No. obs. 0.246 74,045 0.215 149,458 Notes. OLS estimates for 2010. Control variables: see Table 2. Robust standard errors corrected for multiclustering at the industry, municipality, and main bank levels appear between parentheses. Signi…cance levels: * p <0.10, ** p <0.05, *** p <0.01. Table 5. The employment e¤ect of weak-bank attachment. Di¤erence in Di¤erences Dependent variable: 4 log (1 + nij ) (1) W Bi Firm controls (1) Firm controls (2) Municipality …xed e¤ects Industry …xed e¤ects Industry Municipal. f.e. Main bank …xed e¤ects R2 No. obs. -0.074 (0.013) (2) -0.076 (0.010) no no yes yes no no no no – – yes no 0.046 149,458 0.150 149,458 (3) Sign. ctrls -0.035 (0.006) (4) Baseline -0.028 (0.006) yes no – – yes no yes yes – – yes no yes yes – – yes yes yes yes – – yes no 0.177 149,458 0.179 149,458 0.203 112,933 0.155 149,458 (5) -0.028 (0.007) (6) Placebo 0.006 (0.007) Notes. OLS estimates for 2010, except for column (6), which is dated in 2006. Control variables: see Table 2. “yes/no/–” indicates whether the corresponding set of variables is either included, not included or redundant. In column (3) only performance-related …rm control variables are included. Robust standard errors corrected for multiclustering at the industry, municipality, and main bank levels appear between parentheses. Signi…cance levels: * p<0.10, ** p<0.05, *** p<0.01. 49 Table 6. The employment e¤ect of weak-bank attachment. Instrumental Variables Dependent variable: 4 log (1 + nij ) (1) All …rms Instrumented variable 4 (2) Multi-bank …rms log (1 + Creditijk ) 0.519 (0.179) 0.797 (0.294) -0.053 (0.015) -0.031 yes yes yes yes -0.028 -0.025 First stage W Bi Firm controls Industry Municipality …xed e¤ects Overall e¤ect ( ) F test / p value 13.1/0.00 149,458 No. obs. (0.011) 7.65/0.00 74,045 Notes. Instrumental variable estimates for 2010. Control variables: see Table 2. Robust standard errors corrected for multiclustering at the industry, municipality, and main bank levels appear between parentheses. Signi…cance levels: * p<0.10, ** p<0.05, *** p<0.01. 50 51 0.177 149,458 yes yes 0.177 149,458 yes yes -0.030 (0.008) (2) Median 0.177 149,458 yes yes (3) Third quartile -0.033 (0.008) 0.181 133,122 yes yes -0.014 (0.004) (4) Survivors 0.183 149,458 yes yes (5) Alternat. measure -0.034 (0.004) 0.200 16,199 yes yes (6) Tradable goods -0.058 (0.023) 0.177 149,458 yes yes (7) Loans to REI -0.030 (0.008) 0.130 145,322 yes yes (8) 2007 ex-ante -0.019 (0.006) 0.188 71,703 yes yes -0.015 (0.006) (9) 2002 Notes. OLS estimates for 2010. Control variables: see Table 2. Robust standard errors corrected for multiclustering at the industry, municipality, and main bank levels appear between parentheses. Signi…cance level: p<0.10, ** p<0.05, *** p<0.01. No. obs. R2 Firm controls Industry Municipality f.e. W Bi (1) WB Intensity -0.092 (0.020) Table 7. The employment e¤ect of weak-bank attachment. Di¤erence in Di¤erences Dependent variable: 4 log (1 + nij ) Table 8. The employment e¤ect of weak-bank attachment. Panel estimates Dependent variable: 4 log (1 + nijt ) 2008 W Bi 2009 W Bi 2010 W Bi -0.012 (0.004) -0.020 (0.004) -0.027 (0.006) Firm controls Firm …xed e¤ects Industry Municipality Year …xed e¤ects R2 yes yes yes 0.789 563,189 No. obs. Notes. OLS estimates for 2007-2010. Control variables: see Table 2. Robust standard errors corrected for multiclustering at the industry, municipality, and main bank levels appear between parentheses. Signi…cance level:* p<0.10, ** p<0.05, *** p<0.01. Table 9. The employment e¤ect of weak-bank attachment. Matching Dependent variable: 4 log (1 + nijt ) W Bi (1) Propensity score -0.032 (0.009) (2) Exact (4) Exact -0.016 (0.009) W B i Intensity Firm controls Municipality Industry …xed e¤ects (3) Propensity score -0.065 (0.016) -0.052 (0.020) yes yes yes yes yes yes yes yes Overall e¤ect – – -0.016 -0.012 R2 0.228 55,712 No. obs. 0.245 133,816 0.228 55,712 0.245 133,816 Notes. Weighted least squares estimates for 2010. Control variables: see Table 2. Robust standard errors corrected for multiclustering at the industry, municipality, and main bank levels appear between parentheses. Signi…cance level: * p<0.10, ** p<0.05, *** p<0.01. 52 53 P75 2.1 24.2 14.0 33.3 17.0 1.8 0.0 48.3 16.1 9.9 54.6 97.3 43.4 100.0 100.0 3.0 0.0 0.0 100.00 0.0 Mean 8.8 32.3 18.3 22.2 12.0 3.5 13.6 30.6 11.0 6.0 37.4 47.4 24.6 79.4 68.8 2.3 0.4 1.7 59.8 23.7 Treated St. Dev. P25 15.1 0.0 39.8 0.0 273.3 2.0 25.9 0.0 9.1 6.0 52.8 0.2 34.2 0.0 22.5 12.1 13.8 1.6 10.3 1.9 26.8 14.0 39.8 0.0 35.7 0.0 34.0 64.5 46.3 0.0 2.0 1.0 6.2 0.0 13.0 0.0 49.0 0.0 42.5 0.0 P50 0.0 5.0 5.0 14.0 11.0 0.6 0.0 25.9 5.9 4.7 34.6 45.5 0.0 100.0 100.0 2.0 0.0 0.0 100.0 0.0 P75 11.8 67.8 12.0 35.1 16.0 1.6 0.0 45.2 15.1 9.1 57.6 90.6 48.6 100.0 100.0 3.0 0.0 0.0 100.0 100.0 Normalized di¤erences 0.20 0.25 -0.02 0.06 -0.04 -0.02 0.02 -0.06 -0.03 -0.03 0.08 -0.02 0.03 -0.01 -0.03 0.03 0.00 0.00 0.04 0.02 Notes. Observations: 149,458 …rms; 106,128 control and 43,330 treated …rms. WB denotes weak banks. Variables are ratios in percentages unless otherwise indicated. The last column shows the normalized di¤erence test of Imbens and Wooldridge (2009). See de…nitions in Appendix 2. Loans with WB/Assets Share of Loans with WB Employment (employees) Temporary Employment Age (years) Size (million euros) Exporter Own Funds/Assets Liquidity/Assets Return on Assets Bank Debt Short-Term Bank Debt (< 1 yr) Long-Term Bank Debt (> 5 yrs) Non-Collateralized Bank Debt Credit Line (has one) Banking Relationships (no.) Current Loan Defaults Past Loan Defaults Past Loan Applications All Loan Applications Accepted Mean 4.8 19.3 29.5 19.9 12.6 6.9 12.6 32.6 11.7 6.4 34.3 48.4 23.0 79.7 71.0 2.2 0.4 1.7 57.2 22.7 Control St. Dev. P25 P50 11.9 0.0 0.0 34.2 0.0 0.0 375.3 2.0 6.0 25.3 0.0 10.0 9.9 6.0 11.0 132.9 0.3 0.6 33.1 0.0 0.0 23.2 13.3 28.2 14.8 1.6 6.1 11.3 1.7 4.7 27,2 10.0 30.4 40.9 0.0 46.7 35.7 0.0 0.0 34.5 68.2 100.0 45.4 0.0 100.0 2.0 1.0 2.0 6.2 0.0 0.0 12.9 0.0 0.0 49.5 0.0 100.00 41.9 0.0 0.0 Table 10. Descriptive statistics of …rms below (control) and above (treated) the median of weak bank density ratio in 1988 (2006) Table 11. The employment e¤ect of weak-bank attachment. Exogenous variation Dependent variable: 4 log (1 + nijt ) Instrumented variable (1) (2) W Bi W Bi Intensity -0.076 (0.036) -0.320 (0.157) 0.445 (0.084) 0.105 (0.025) First stage W eak bank density i Firm controls Municipality Industry …xed e¤ects Industry …xed e¤ects Coast …xed e¤ects Overall e¤ect F test / p value No. obs. yes no yes yes yes no yes yes -0.076 -0.073 17.8/0.00 149,458 28.3/0.00 149,458 Notes. Instrumental variable estimates for 2010. Control variables: see Table 2. “yes/no” indicates whether the corresponding set of variables is either included, not included. Robust standard errors corrected for multiclustering at the industry, municipality, and main bank levels appear between parentheses. Signi…cance levels: * p<0.10, ** p<0.05, *** p<0.01. 54 Table 12. The employment e¤ect of weak-bank attachment. Triple Di¤erences Dependent variable: 4 log (1 + nijt ) W Bi Rejected applicationi Rejected applicationi W B i Past Defaultsi Past Defaultsi W B i Short-term debti Short-term debti W B i Bank debti Bank debti W B i Own funds ratioi Own funds ratioi W B i Liquidity ratioi Liquidity ratioi W B i Single banki Single banki W B i log(Total Assetsi ) log(Total Assetsi ) W B i log(1+Agei ) log(1+Agei ) W B i I(Exporteri ) I(Exporteri ) W B i I(Temporary employeesi ) I(Temporary employeesi ) W B i I(Servicesi ) W B i Industry Municipality …xed e¤ects Firm controls R2 -0.019 -0.066 -0.029 -0.209 -0.041 -0.089 -0.036 -0.096 -0.081 0.061 0.134 0.118 0.050 0.012 0.019 0.009 0.003 -0.054 0.027 0.176 0.062 -0.112 -0.027 0.019 (0.007) (0.008) (0.012) (0.029) (0.020) (0.013) (0.011) (0.017) (0.022) (0.026) (0.027) (0.022) (0.061) (0.005) (0.015) (0.005) (0.005) (0.006) (0.008) (0.011) (0.020) (0.010) (0.014) (0.017) yes yes 0.177 149,458 No. obs. Notes. OLS estimates for 2007-2010. Control variables: see Table 2. Robust standard errors corrected for multiclustering at the industry, municipality, and main bank levels appear between parentheses. Signi…cance level: * p<0.10, ** p<0.05, *** p<0.01. 55 Table 13. The employment e¤ect of weak-bank attachment. Margins of adjustment (1) (2) Dependent variable (n =n ) log (W age billijk ) 4 temp;ijk ijk 4 W Bi -0.005 -0.016 (0.002) (0.006) Firm controls Municipality Industry Year …xed e¤ects yes yes R2 yes yes 0.174 122,725 No. obs. 0.205 87,451 Notes. OLS estimates for 2007-2010. Control variables: see Table 2. Robust standard errors corrected for multiclustering at the industry, municipality, and main bank levels appear between parentheses. Signi…cance level: * p<0.10, ** p<0.05, *** p<0.01. Table 14. E¤ect of weak-bank attachment on …rm exit Dependent variable: Probability of exit (1) (2) W Bi 0.011 (0.004) W B i Intensity 0.059 (0.014) Firm controls Municipality Industry …xed e¤ects R2 No. obs. yes yes yes yes 0.173 150,442 0.173 150,442 Notes. OLS estimates for 2010. Control variables: see Table 2. “yes/no” indicates whether the corresponding set of variables is either included or not. Robust standard errors corrected for multiclustering at the industry, municipality, and main bank levels appear between parentheses. Signi…cance level: * p<0.10, ** p<0.05, *** p<0.01. 56 Figure 1: New credit to non-…nancial …rms by bank type (12-month backward moving average, 2007:10=100) . 57 Figure 2: Acceptance rates of loan applications by non-current clients, by bank type, 2002:4-2012:6 (%) 58 Figure 3: Average annual interest rate for new loans to non-…nancial …rms by bank type and the policy rate, 2003:1-2012:6 (%) 59 Figure 4a. E¤ect of weak-bank attachment on credit at the local level, 2002-2010 (pp) Figure 4b. E¤ect of weak-bank attachment at the …rm level, 2002-2010 (pp) 60 Figure 5. Employment e¤ect of weak-bank attachment, 2002-2010 (pp) 61