Loan Types and the Bank Lending Channel

Using credit-registry data for Spain and Peru, we document that four main types of commercial credit — asset-based loans, cash flow loans, trade finance and leasing — are easily identifiable and represent the bulk of corporate credit. We show that credit growth dynamics and bank lending channels vary across these loan types. Moreover, aggregate credit supply shocks previously identified in the literature appear to be driven by individual loan types. The effects of monetary policy and the effects of the financial crisis propagating through banks’ balance sheets are primarily driven by cash flow loans, whereas asset-based credit appears to be largely insensitive to these types of effects.

1 2 A significant fraction of commercial credit consists of cash flow loans. Lian and Ma (2018) estimate that as much as 80% of syndicated credit in the U.S. is cash-flow based. The difference between asset-based loans and cash flow loans can again be understood from the perspective of the collateral used to secure the credit. As already mentioned, in the case of asset-based loans, the borrower pledges specific physical assets to secure the loan. In the case of cash flow loans, the lender has a senior claim on all unencumbered assets of the company; that is, the lender has a first claim on all proceeds from asset liquidations (excluding assets that were already pledged). 2 Overall, the collateral in cash flow loans differs from asset-based loans and leases in several dimensions: it is oftentimes less durable, has lower liquidation value due to its less standardized nature, and has lower pledgeability due to uncertainty about its value (e.g., intellectual property or retailer inventory) or lack of title (e.g., computers and office furniture). Indeed, a typical credit agreement for a cash flow loan does not have a comprehensive list of what represents collateral in the transaction. It is its senior secured position in the capital structure-and not the claim over specific assets-that allows it to have recovery in case of a default. 3 As a result, in the credit assessment of the cash flow loan, the emphasis is not on the value of collateral (as in the case of an asset-based loan), but on the borrower's ability to pay the interest and amortization (hence, the label "cash flow" loan).
The last significant category of commercial credit-trade finance-backs business-to-business (B2B) transactions. 4 This type of credit is backed by a bilateral contract, such as a contract for delivery of goods. Amiti and Weinstein (2011) 3 provide a detailed insight into the working of trade finance. 5 The loan in this case is backed by the goods that are being transacted, so the collateral is well identified, valued (and insured), and the title of the good is in transfer (so it is not yet part of what will become collateral of a cash flow loan). Trade finance agreements are probably closest to cash flow loans, but the pledgeability of the collateral in this case is higher. To be fair, there are other unique features of trade finance agreements, as it involves multiple counterparties and the credit risk in this case is no longer simply that of the borrower.
There are also other distinctions in terms of processes and sources of capital among the four lending categories. However, we want to highlight that in practice one often observes that borrowers have multiple loans outstanding of different loan type, and that the collateral for these different loan types can be partitioned without generating conflict among creditors. Importantly, what emerges from the earlier discussion is that different loan types carry different credit risks and involve different practices for mitigating negative shocks. Different loan types would also be differentially affected by fluctuations in the value of collateral.
As we will show-precisely because of its ubiquity-accounting for loan type using representative credit registry data appears to be straightforward. Moreover, accounting for the loan type is pivotal for several reasons: First, failing to account for loan type leads to a mismeasurement of the credit channel effect. The methodology in Khwaja and Mian (2008), henceforth KM,-the workhorse of the literature focused on transmission of financial shocks to the real economy-performs a within-firm cross-sectional comparison of lenders' behavior.
This approach relies on the assumption that firm-specific changes in credit demand are constant across lenders, hence the estimated differences can be attributed to differences in credit supply. But, to the degree that there is a correlation between investment opportunities and credit type, and if different lenders provide different loan types to a given borrower, the identifying assumption that the borrower's credit 4 demand is fixed across lenders is violated. 6 (See Appendix 1 for a more detailed description of the biases that arise when using the KM methodology without accounting for loan heterogeneity.) Second, the caveat of the KM methodology, which relied on borrowers with multiple lenders at a given point in time, is that it discards a significant fraction of the borrowers in the economy (the single lender borrowers). 7 Thus, the generalizability of the conclusions emerging from such identification, as well as the policy implications, are dependent on the broader distribution of credit. Indeed, we show that a substantial number of borrowers in most economies rely on a single loan type. Historically, this number was 60% in Spain and 42% in Peru. Moreover, the type of loans used by these firms tends to be extremely persistent. Note that just because a borrower has a relationship with one lender does not mean that it relies on one loan type. However, we find that the overlap, at a given point in time, between the sample of borrowers that rely on one lender and the sample of borrowers that rely on one loan is substantial: about 79% of single-lender borrowers are also single-loan type for Spain; this number is 69% for Peru.
Finally, there is a large body of literature that shows that bank credit shocks affect economic activity. The focus of more recent literature has not been on whether such connections exist, but on the measurement of the sensitivity of real output to bank funding shocks, and the economic mechanisms underlying this connection. Thus, recognizing that these sensitivities vary by the type of credit is an important advance 6 We know that large lenders can provide any loan type. This is not to say that you could easily switch the type of credit across existing lenders. For example, if a given borrower has a cash flow loan with bank "A" and an asset-based loan with bank "B", then it is likely to go back to bank "A" when it needs to increase its cash flow loan. As we said, different loan types are about different credit risk, the type of screening and monitoring is likely to be specific to a given loan type. So information frictions in lender switchingfor a given loan type-are likely. 7 Using monthly observations, Ioannidou, Ongena and Peydró (2015) find that 46% of borrowers in Bolivia have only one lender at a given point in time. Bolton et al. (2016) show that this number for Italy is 60%, and, according to Morais et al. (2019), 79% of Mexican firms tend to have one lender. Quarterly data used in Khwaja and Mian (2008) for Pakistan, Iyer et al. (2014) for Portugal, and Baskaya et al. (2017) for Turkey show that the fraction of borrowers with one lending relationship is 90%, 25%, and 54%, respectively. 5 in both of these directions. Specifically, because the differences between the types of commercial credit are rooted in the nature of collateral, it would likely affect young and old firms differentially, as well as firms in different industries. Although this implication of our findings is outside of the scope of this paper, it could lead to a better understanding of the sources of financial constraints in the cross-section of firms, and over a firm's lifecycle.
Our empirical analysis is based on the credit registry data from Spain and Peru.
We start by showing that in both countries the bulk of bank commercial credit can be grouped into four main types: asset-based loans, cash flow loans, trade finance agreements, and leases. The first two types of loans are the most common type of credit in both countries, both in terms of number and volume of loans. For instance, in 2004, asset-based loans accounted for 39.1% and cash flow loans accounted for 48.2% of the volume of total commercial credit by banks in Spain. For Peru, these figures are 43.5% and 35.8%, respectively. The average size of asset-based loans is much larger than that of the other forms of credit, averaging about 1.0 million euros in the case of Spain and 6.4 million soles in the case of Peru.
Thus, the first contribution of our paper is to give a comprehensive insight into the prevalence of different loan types in representative economies, and provide a transparent methodology for identifying loan-type in a credit registry data. The second contribution is to show that what we know about the magnitude of the credit supply effects is driven by specific loan-types. To do so, we reexamine several notable studies, zooming in on individual loan types.
Applying the methodology proposed by Amiti and Weinstein (2018), we find that aggregate credit supply shocks constructed using their methodology differ substantially across loan types. In particular, in the Spanish data, we find that aggregate credit supply shocks are most strongly correlated with supply shocks in asset-based loans, while the correlation is weaker for cash flow loans, trade finance, and leasing.
We re-estimate the baseline specifications in Jiménez, Ongena, Peydró and Saurina (2012), Paravisini, Rappoport, Schnabel and Wolfenzon (2015) and 6 Bentolila, Jansen and Jiménez (2018) separately for different loan types. 8 Jiménez et al. (2012) uses Spanish data to assess how variation in bank capital interacts with changes in monetary policy rates to influence credit growth. Consistent with a risktaking channel of monetary policy, they find that lower interest rates spur loan growth especially at lowly capitalized banks. Paravisini et al. (2015) and Bentolila et al. (2018) build on the KM methodology using shocks to banks' liquidity following the 2007-2008 global financial crisis, in Peru and Spain, respectively. Using variation in bank exposure they quantify relative effects of the credit channel. We find that the results in all of these papers are sensitive to the type of loan considered. Moreover, they appear to be driven mainly by cash flow loans, while asset-based loans exhibit a different pattern in the estimates. Gertler (1989, 1995) argue that monetary policy affects the external finance premium of firms by altering the agency costs associated with asymmetric information between borrowers and lenders about the quality of firm investments. Easing monetary policy increases cash flows and collateral value, thus leading to a reduction in agency costs, which, in turn, makes it easier for the firm to borrow. While this theory does not have clear predictions for different loan types, arguably, the liquidation value of cash flow loans is more sensitive to changes in agency costs, in which case monetary policy should affect cash flow loans more so than loans based on hard collateral. Similarly, if in the financial crisis the rise in agency costs was dominating the impact on collateral, it would explain why we find that cash flow loans appear to be driving contraction in credit supply during the financial crisis.
In this paper, we build on and contribute to a number of strands of literature.
First, we contribute to the literature on how the supply of credit is influenced by monetary policy and financial shocks including Kashyap, Stein and Wilcox (1993), Bernanke and Gertler (1995), Kashyap and Stein (2000), Ivashina and Scharfstein (2010) and the set of papers that trace the impact of credit market disruptions on real 7 outcomes including Kashyap, Lamont and Stein (1994), and Chodorow-Reich (2013).
Naturally, our paper contributes to the large body of empirical studies that build on the empirical approach in Khwaja and Mian (2008) and use loan-level data to measure effects of credit shocks and their transmission. As already mentioned, we will specifically replicate Jiménez et al. (2012), Paravisini et al. (2015) and Bentolila et al. (2018). Perhaps closest to our research are recent papers by Paravisini, Rappoport and Schnabl (2017), which considers lender specialization, and Jiménez, Mian, Peydró and Saurina (2019), which incorporates firm-level general equilibrium adjustments. Both studies refine the KM methodology. We contribute to this literature by considering the importance of loan types. We show that this differentiation is grounded in actual lending practices of banks and that credit growth dynamics and bank lending channels vary across loan types.
Finally, there is the emerging literature focused on quantifying the aggregate effects of credit shocks on real outcomes, such as investment and output. We specifically build on Amiti and Weinstein (2018) and show that accounting for loan types is also relevant in evaluating the drivers of aggregate effects.
The paper proceeds as follows. Section 1 discusses the data from the Spanish and Peruvian credit registry. Section 2 shows patterns of use of different types of commercial loans by borrowers. In Section 3, we present results of estimating effects of credit supply accounting for loan heterogeneity. Section 4 concludes.

Data
In the analysis, we use credit registry data from two countries: Spain and Peru.
Credit registries (or credit bureaus) are depositories of loan level information typically collected and maintained by the central bank for purposes of monitoring and regulation. They are also regularly used by local lenders to verify the credit history of a prospective borrower. To the extent that the type of credit is key information for assessing credit risk, information on the type of credit should be recorded in a credit registry and in any other major loan-level database. That is 8 indeed the case in Spanish and Peruvian credit registries. 9 We elaborate on this below.
Typically, a credit registry tracks loan stock, that is, outstanding credit amount with monthly or quarterly frequency. One cannot observe individual loans in such data, but instead one observes lending relationships for a given borrower at a given point in time. Empirical work building on Khwaja and Mian (2008) generally constrains the analysis to observations in periods when the borrower has more than one lending relationship outstanding. As discussed earlier, this substantially limits the sample of borrowers used in the estimation. In our data, restricting the sample to firms with multiple lending relationships drops the number of unique borrowers by 50 percentage points for Spain and 39 percentage points for Peru. Naturally, once one accounts for loan type, this further restricts the sample of unique borrowers covered in the analysis. This is because, in the KM approach, borrowers that had one lender per loan type would be part of the analysis (as long as there is more than one lender/loan type). Once we account for the loan type, these borrowers drop from the sample. Given this constraint, in what follows we will consider quarterly observations as a less restrictive time-unit of observation.

The Spanish CIR Dataset
The Central Credit Register (Central de Información de Riesgos or CIR in Spanish) is maintained by the Banco de España in its role as primary banking supervisory agency, and contains detailed monthly information on all outstanding loans over 6,000 euros to non-financial firms granted by all banks operating in Spain since 1995. Given the low reporting threshold, virtually all firms with outstanding bank debt will appear in the CIR. We also use a dataset on loan applications, or, more precisely, bank requests for firm information, which are interpreted as loan 9 applications. 10 By matching the monthly records on loan applications with the stock of credit, we infer whether the loan materialized. If not, either the bank denied it or the firm obtained funding elsewhere. This loan granting information is available from February 2002 onwards and is the same data used by Jiménez et al. (2012). Because several of the results will build on Jiménez et al. (2012), we restrict the Spanish sample to the 2002-2010 period. However, the descriptive statistics by loan type generalize to a longer sample.
In the Spanish data, we identify four main loan types based on two first-order variables: Clase or loan-risk class and Garantia or collateral. 11 From the indicator of type of risk, we consider the following three categories: 12 trade finance or Crédito comercial (clase A) in Spanish, commercial and industrial (C&I) loans or Crédito financiero (clase B), and leasing or Operaciones de arrendamiento financiero (clase K). That is, leasing and trade finance in the Spanish data can be identified using solely Clase variable. To separate between asset-based and cash-flow based loans, we turn to information on collateral. Spanish credit registry focuses on non-personal collateral (Garantia real) that includes assets like real estate, naval mortgages, securities, deposits, and merchandise (i.e., hard collateral). As mentioned earlier, it is the senior secured status (and contractual restriction on asset sales) that imply that the loan has collateral. The definition of collateralization in Spanish and Peruvian credit registries only concerns whether the loan is collateralized by hard assets. In the Spanish data, we know whether the value of the loan is collateralized by hard assets at (i) 100%, (ii) (50%; 100%); or (iii) is not collateralized. However, 98% of the loans in the data either have 100% real collateral or no real collateral. We categorize C&I loans "with collateral" as asset-based loans, and as cash flow loans otherwise. 10 Banco de España can fine those banks requesting information without intent of loan origination. 11 For full details, we refer to Circular 3/1995, de 25 de septiembre, a entidades de crédito, sobre la Central de Información de Riesgos (available at: https://www.boe.es/buscar/pdf/1995/BOE-A-1995-22113-consolidado.pdf). 12 We exclude guarantees and other contingent claims that are recorded in the Spanish data: "avales, cauciones y garantías" account for around 10% in number of loans and "riesgo indirecto" for around 6%. All the other types represent between 0.02% and 0.89% and are thus negligible. In terms of volume, these figures are even smaller.
10 Consistent with our categorization of loan types, in Appendix Table A1, we use balance-sheet information taken from Almunia et al. (2018) available for the firms in the Spanish sample, and confirm that firms with a higher share of fixed assets over total assets rely more on asset-based lending. In particular, asset tangibility has a strong economic and statistical power in explaining borrower's reliance on assetbased loans, yet practically no economic power on whether it uses leasing. On the other hand, the correlation between asset tangibility and borrower's share of cash flow lending is negative. These results are robust to inclusion of controls for firm's age, size, leverage, and industry.

The Peruvian CIR Dataset
Similar to the Spanish data, the Central Credit Register of Peru is maintained by the bank supervisory agency, which in this case is the Superintendencia de Banca y Seguros (SBS). The data available to us covers a period between January 2001 and April 2018. 13 The sample is constraint to credit to firms with sales above 20 million soles in sales (about $6 at the current exchange rate). We further focus on lending by banks, finance companies and cajas (savings banks). 14 Microcredit institutions are excluded from the sample.
We assign loans into four basic types using the variable called Cuenta, which describes the type of loan as well as the collateral used in this transaction. We assign as asset-based credit the following loan types: garantía hipotecaria, garantías preferidas, garantías preferidas autoliquidables, garantías preferidas de muy rápida realización, and otras garantías. That is, we treat loans with the following collateral as asset-based loans: real estate, other collateral with stand-alone title, deposits and other liquid financial securities, and other collateral. The values of Cuenta that are assigned to cash flow-based lending include: líneas de crédito (revolving lines) and 13 Peruvian data has a structural break in 2010:Q2, at which point the Credit Registry applied a new size filter for corporate loans. This should not affect the estimates of the supply shocks; however, this compositional shift in the data will be apparent in the descriptive statistics. 14 Several countries have community banks and savings banks. Their origin and stated main objective is different than that of commercial banks; however, they often act as traditional lenders.
préstamos (loans). Créditos-comercio exterior (loans for international trade) are assigned to the trade finance category. Finally, leasing is identified as arrendamiento financiero. The types of credit that do not fall into one of these categories represent, on aggregate, about 10% of all loans and 20% of total commercial loan volume. For Peru, we only capture international trade, whereas for Spain we capture all types of credit to firms. As such, the behavior of trade financing is not directly comparable across the two countries.

Use of Different Types of Secured Bank Credit
In Table 1, we report descriptive statistics by loan type for 2004, 2008, and 2012. Panel A corresponds to the Spanish sample, and Panel B presents the same statistics for Peru. In line with evidence from the U.S. syndicated loan market presented by Lian and Ma (2018), cash flow loans are an essential form of bank credit representing about half of all commercial credit in Spain and about a third of all commercial credit in Peru by loan volume or loan number. However, the universe of bank credit shows that asset-based credit is also pivotal, and perhaps not surprisingly even more so for an emerging economy. According to Djankov et al. (2008), debt enforcement tends to be weaker in emerging markets, making it more difficult to recover value in case of a default, rendering collateralized debt more attractive.
Asset-based loans correspond to about 40% of lending volume in Spain, but only about a quarter of all outstanding loans; whereas in Peru, asset-based loans in the past ten years had represented about 50% of the lending volume and close to 40% of all loans. 13 Turning to the evolution over time, Figure 1 illustrates that there is heterogeneity in the patterns across loan types suggesting that the Spanish credit boom was mostly driven by asset-based lending. The average size of asset-based loans doubled during the boom, but it was then adjusted. Also, the increase in total credit was fourfold during the boom for asset-based loans while it was three-and two-fold for leasing and cash-flow lending, respectively. Similarly, the rise in asset-based lending seems to be at the heart for the credit expansion experienced in Peru. 15 Consistent with the weaker creditor protection that we expect for an emerging market, leasing (often a direct alternative to asset-based loans) plays a substantial role as well.  Figure 1 only includes borrowers that appear in the credit registry after the reclassification. This softens the structural break in the data, but it can still be seen. The focus on the growth of leasing and asset-based loans however is unaffected by the methodology for constructing this graph, and is clearly there even in the raw sample.   borrowers with one loan type at the year end, even if in the past they had loans of other types.) The first matrix looks at the probability that 1 year later (at the end of the next year) the borrower has a given loan type (indicated in columns). The sample is conditional on the borrower remaining in the credit registry sample at the end of the period. That is, borrowers that leave the sample are not counted. To summarize the results, we first take an average across borrowers within a year, and then report the average across years. The borrower can migrate to more than one type of credit. As a result, each row can add up to more than 100%. That said, if loan types are irrelevant, and the assignment is random, the benchmark would be 25%.
Taking Peru as an example, we see that 94.9% of borrowers that have assetbased loans will have an asset-based loan next year, and only 0.6% of them will completely substitute to a different loan type. However, we can see that about 16 These differences in the use of one or more loan types do not appear related to borrowing from one or more banks. Figure A1 shows that the distribution of loan type is broadly similar for borrowers with multiple lenders and borrowers with single lenders. 17 "Over the past decade, Peru has been one of Latin America's fastest-growing economies, with an average growth rate 5.9 percent in a context of low inflation (averaging 2.9 percent)." Source: https://www.worldbank.org/en/country/peru "Peru At-A-Glance" accessed May 31, 2019. As mentioned earlier, 2010 in the Peruvian data is also characterized by a reporting switch.
17% of borrowers that exclusively rely on asset-backed loan expand into cashflow based loans. The typical maturity of a loan in Peru is about one year, so this result is unlikely to be mechanical. However, we also report results for a 3-year window (bottom panel). Overall, even at the longer horizon, the loan type appears to be very persistent for each of the loan types. There is also little mobility of the loan type during the financial crisis (middle panel). Similarly, we see very little loan migration in the Spanish sample, indicating that loan type choices are persistent.

Credit Supply Shocks and Loan Heterogeneity
In what follows, we incorporate loan types into the existing studies of the shocks to credit supply that use micro-data. Specifically, we will replicate core results in three prominent papers that isolate the effects of credit supply: Amiti and Weinstein (2018) Conveniently, Jiménez et al. (2012) look at the Spanish markets, and Paravisini et al. (2015) look at the Peruvian market, which is the data that we have, and it will help us confirm that our results are not driven by sample differences. However, a more profound reason for selecting these three studies is that they use related but different methodologies: they deviate in the strength of the assumptions required for identification of credit supply and, consequently, the generality of their insight. Amiti and Weinstein (2018) present an empirical methodology for constructing time-series of aggregate supply and demand shocks. As the authors themselves emphasize, their shortcoming is that shocks to supply are identified from the data using a stringent structure for the behavior of supply and demand.
(We present a detailed discussion of the identifying assumptions in the Appendix.) On the other side of the spectrum are studies that use a direct exogenous shock that impacts the balance sheet of the banks: an anticipated nuclear test in Pakistan that led to the collapse of the dollar deposit funding Electronic copy available at: https://ssrn.com/abstract=3543557 (Khwaja and Mian, 2008), or shortages in foreign funding to Peruvian banks following the Lehman collapse (Paravisini et al., 2015). These studies rely on differential exposure of banks to such shocks; they "cleanly" identify marginal effects in the cross-section, but do not provide a path for measuring an aggregate impact.
We want to stress that it is not our intention to opine on the usefulness or validity of these different methodological approaches. We take these studies as a state of empirical literature on the credit channel. Our core insight is thatregardless of the approach-the credit channel operates differentially for separate loan types.

Credit Supply Shocks in the Time-Series
Amiti and Weinstein (2018), henceforth AW, develop an empirical methodology for constructing aggregate supply and demand shocks using matched bank-firm loan data (the AW application is based on a sample of around 150 banks and 1,600 listed firms in Japan from 1990 to 2010). As the authors point out, many studies have shown that bank shocks matter for loan supply; however, that tells us little about how important bank loan supply in determining aggregate variables such as investments and employment is. The AW methodology builds on the following specification: where ∆ refers to loan growth by firm f from bank b in time t measured as log changes, refers to the "bank lending channel," and refers to the "firm borrowing channel." One approach to identify both channels is to empirically estimate (1) in a regression framework that is saturated with firm-time and banktime fixed effects. AW show that this procedure is inefficient because it ignores general equilibrium considerations. For instance, a firm cannot borrow more without at least one bank willing to lend more (and vice versa). The core of AW's methodological insight is that one can account for general equilibrium conditions at the aggregate level by imposing that total credit growth is recovered by summing up the sequences of the estimated fixed effects so that the R-squared of the regression is equal to one by construction.
Using annual data from the Spanish and Peruvian credit registry, we estimate equation (1) with the AW methodology and recover a sequence of bank-year and firm-year fixed effects that can be interpret as bank credit supply shocks and firm demand shocks. These shocks are labelled as "All loans (ALL)" because they are based on total credit encompassing all loan types at the bank-firm-year level as in the original AW approach. We next construct four alternative samples in which each bank-firm-year observation refers to credit exposure from a particular loan category. The AW methodology is then applied to each subsample and four different supply shocks are estimated for each bank in the sample: asset-based loans (ABL), cash flow loans (CF), trade finance loans (TF), and leasing loans (LEA) shocks.  shock identified from, for instance, cash flow loans, which definitely indicates that loan heterogeneity matters for the identification of bank supply shocks. Note that we focus here on bank shocks, but the patterns are similar in the case of the estimated firm shocks shown in Figure A2 in the Appendix. Table 4 presents correlations of AW credit supply shocks across different loan types. While the evidence in Figure 3 is already suggestive, the patterns in Table   - 4 are even more revealing of the importance of loan heterogeneity for identifying bank credit supply shocks. In particular, Table 4 shows that in all the six cases the estimated correlations are basically zero for both countries, which clearly points to loan-specific credit supply shocks at the bank level.  Finally, in unreported results, we also estimated equation (1) considering two alternative sets of fixed effects. In the Spanish data, if we regress credit growth on firm-quarter fixed effects, we explain 28 percent of the total variation in credit growth (R-squared = 0.28). However, if we instead include (firm × loan type × quarter) fixed effects, the variation explained increases to 39 percent (R-squared = 0.39). The corresponding numbers for Peru are 0.23 and 0.42. This shows that loan type heterogeneity matters to explain differences in credit growth across banks within each firm.
Overall, in line with the evidence in Figure 1, the AW methodology reveals stark differences in behavior of bank credit shocks for different loan types over our sample period.

Bank Lending Channel of Monetary Policy
In this section, we re-estimate the baseline specifications in Jiménez et al. (2012), which uses Spanish data. This study gets closer to the Khwaja and Mian (2008) setting, in that it relies on within-borrower differential response of lenders to monetary policy. One potential critique of this study, however, is that it does not have a clean exogenous credit supply shock, but instead looks at the ECB's changes in interest rates for the euro area.
As before, our idea is to incorporate loan-type into the analysis and to see if credit supply substantially differs across core loan types. Our identification is thus based on differences across banks within the same firm and loan type pair.
Intuitively, we compare the same firm which has cash flow loans from two (or more) banks that are differentially exposed to changes in monetary policy.
The complete replication exercise follows the following econometric model: where f, b, and t refer to firm, bank, and quarter, respectively. l refers to the loan type, and corresponds to (firm × loan type × quarter) fixed effects. Even though a large fraction of the borrowers relies on one loan type, many borrowers use more than one lender by loan type, which is what ultimately allows us to do the replication exercise. We consider two alternative dependent variables, namely, credit growth for the intensive margin and issuance of new loans for the extensive margin. Credit growth is based on annual log differences winsorized at -100% and +200%, and the new loan dummy takes value 1 when a bank-firm-loan type triplet first appears in the sample and zero otherwise. ∆IR t is the annual change in a 3-month Spanish interbank interest rate. ∆GDP t is annual growth of real GDP.
CAP bt−1 and LIQ bt−1 refer to the capital and liquidity ratios at the bank level.
With a minor exception, controls are as in Jiménez et al. (2012). Banks characteristics include log total assets, doubtful assets ratio, return on assets, capital ratio, and liquidity ratio. Firm controls include: (i) ratio of equity over total assets, (ii) ratio of the current assets over total assets, (iii) the log of the total assets of the firm (in 2008 euros), (iv) the log of one plus the firm's age in years, Estimates are reported in Table 5. In order to analyze the role of loan heterogeneity, we cannot use loan application data because we do not observe the loan type of the rejected applications (i.e., the zeros). Instead, in columns (1) through (4) Overall, the explanatory power increases, but the magnitude of the effects seems relatively unaltered when including loan type fixed effects. Note however that the estimates in Table 5 can be interpreted as a weighted average of the different effects by loan type with weights given by the number of observations.
In Table 6, we report the estimates by loan type. All specifications include firmquarter fixed effects. The strongest specifications, the ones that can be interpreted as identification of credit supply, are the ones corresponding to interactions with bank capitalization and liquidity (Table 6, Panel B).
The estimates in Table 6 indicate that cash flow loans are at the root of the overall results in Jiménez et al. (2012). Indeed, cash flow loans represent around 53% of the total number of loans (about 11 million out of 21 million observations). Interestingly, asset-based loans, which were central to the Spanish credit boom, present a different pattern in the estimates.  (1) through (4) replicate results in Jiménez et al. (2012), and are directly comparable to the results reported in their paper with the exception that we use credit register data. All regressions include the same control variables as in Jimenez et al. (2012). The variables of interest are: the annual change of Spanish 3-month interbank interest rates (ΔIR); the annual change of Spanish GDP in real terms (ΔGDP); the ratio of bank's equity over total assets (CAP); and the ratio of bank's liquid assets over the total assets (LIQ). In columns (5) through (8)    This table reports the estimated coefficients from re-running specifications (1) through (4) in Table 5 by loan type. For example, regressions (2.a) though (2.d) that appear in Panel B correspond to Table 5, specification (2). The four columns correspond to four key loan types. As in Table 5, the variables of interest are: the annual change of Spanish 3-month interbank interest rates (ΔIR); the annual change of Spanish GDP in real terms (ΔGDP); the ratio of bank's equity over total assets (CAP); and the ratio of bank's liquid assets over the total assets (LIQ). All regressions include the same controls as in Table 5. For each column, Panel A and B have the same number of observations and bank-quarter clusters; these are reported at the end of Panel B. Standard errors are reported in parentheses.

The Global Financial Crisis and the Bank Lending Channel
As mentioned earlier, the "gold standard" of measurement of credit supply in the cross-section requires a credible exogenous shock to banks balance sheet to instrument for the supply of credit. In that sense, some might argue that earlier results that built on Amiti and Weinstein (2018)  (iv) by construction, we only look at the universe of exporters; and (v) we look at the volume of exports measured in kilograms.
We focus on two key results in Paravisini et al. (2015). The central insight of their paper can be already seen in the cross-sectional analysis that they report in their Table 3. The explanatory variables are the share of lenders' assets funded with foreign debt, and an indicator variable of whether that share exceeds 10% (which is the sample mean for commercial banks for that time period.) The results are reported in Table 7. The estimates in column (1) closely match the results in Paravisini et al. (2015). Overall, this result shows that a banks' foreign funding was negatively correlated with its change in supply of credit following the financial crisis. This insight is the core building block of their study.  Table 3 in Paravisini et al. (2015). Columns (2)-(5) re-estimate this equation for different loan types. FD b is the share of foreign funding of bank b. D(FD b >10%) is a dummy that signals whether foreign funding of bank b is above 10%, the mean among the commercial banks (as in Paravisini et al. 2015). Standard errors are clustered at the bank level and are reported in parentheses. As before, our goal is to illustrate that this aggregate result is limited to some, though not all loan types. The rest of Table 7 reports the same regression, but estimated within a sample of individual loan types. Once again, what seems to emerge is that cash-flow appears to be behind the core underlying mechanism.
Asset-backed loans display a sharp difference compared to other types of credit; the statistical significance of the aggregate result comes from the loan types other than asset-based loans. Overall, the structure of this empirical model substantially reduces the number of observations for individual loan types, which likely also impacts the statistical significance of the results. 18 Table 8 instead uses a firm-product-destination level panel of export volumes to estimate intensive margin elasticity of exports to credit shocks. Credit supply is instrumented using banks reliance of foreign funding. Paravisini et al. (2015) emphasize that their empirical strategy relies on accounting for shocks to export demand and input cost which they achieve by looking at variation within productdestination. As with Table 7, our estimates are higher than in the original paper, but reasonably close. Separation of the result by loan-type indicates that the aggregate result does not generalize to leasing or international trade, although in this result, contraction in asset-backed loans does lead to substantial contraction in exports. As before, cash-flow loans continue to play a significant role in economic terms.  (1) and (2) replicate the results in Table 5 in Paravisini et al. (2015). Columns (3)-(6) re-estimate this equation for different loan types. Columns (2)-(6) are estimated using instrumental variables. In the IV regressions, the change in (log of) credit of firm i, ∆lnC i , is instrumented with the measure of lender's exposure to foreign funding which is a weighted average of D(FD b >10%), a dummy that signals whether foreign funding of bank b is above 10%, the mean among the commercial banks (as in Paravisini et al. 2015). Standard errors are clustered at the product-destination level and are reported in parentheses. As a final insight, we look at another study that is also focused on the financial crisis as a source of a shock to banks liquidity, but uses Spanish data, that is Bentolila et al. (2018). This paper exploits differences across banks that were bailed out by the Spanish government ("weak banks") versus the rest ("healthy banks"). Following Khwaja and Mian (2008), Bentolila et al. (2018) consider a first-stage regression at the bank-firm level showing that weak banks curtailed lending relative to the other banks during the global financial crisis. We re-estimate the first-stage regression in Bentolila et al. (2018) at the bank-firmloan type level. More formally: where, as before, f, b, and l refer to firm, bank, and type of loan, respectively. The dependent variable is credit growth between 2006:Q4 and 2010:Q4. The unit of observation is bank-firm-loan type level. Weak bank, a dummy identifying bailed out banks, is the explanatory variable of interest. Z is the log of (one plus) the length of the bank-firm relationship, measured in months. Bank is a vector of bank controls, namely, log total assets, doubtful assets ratio, return on assets, capital ratio, and liquidity ratio. In the first specification of equation (3) we include firm fixed effects, as in Bentolila et al. (2018). All other specifications include firm-loan type fixed effects. The results are reported in Table 9. Column (1) corroborates the finding in Bentolila et al. (2018) that, for a given firm, weak banks reduced credit supply vis-a-vis healthy banks. Column (2) shows that this result is similar when including firm-loan type fixed effects. However, results by individual loan type indicate that cash-flow based loans are at the root of this finding. Estimated effects are not significant for asset-based loans and trade finance. Also, the number of cash flow loans is larger than that of the other loan types together.

Conclusions
Practitioners commonly refer to four distinct loan types: asset-based loans, cash flow loans, trade financing, and leasing. At the heart of this distinction is the speed and size of recovery in default. Some of these types of credit had been directly or indirectly studied in the literature; however, this distinction has been overlooked by the literature focusing on the conditions of bank credit supply. Yet, as we show, such distinction is important as a large fraction of companies in any economy relies on a single loan type, and these loan types tend to be very persistent. Moreover, given that the quality of measurement of supply effects is central to several of the studies using narrow fixed effect identification, accounting for the loan type is an important identifying assumption.  Khwaja and Mian (2008): where ∆ refers to loan growth by firm f from bank b in time t, denotes a bank-specific shock (e.g., a liquidity shock due to nuclear tests in the case of Khwaja and Mian, 2008), and refers to a firm-specific demand shock. The expectation of the error term is assumed to be zero: [ ] = 0. This type of empirical specification has been used to disentangle the firm-borrowing channel (demand shock ) from the bank lending channel (supply shock ). The inclusion of time-varying firm fixed effects implies that identification is based on variation in credit across banks with the same firm, keeping firm credit demand constant across banks.
The crucial assumption is that firms' credit demand is the same across all banks. This assumption may be violated if firms' credit demand is bank specific.
Such would be the case if different lenders are providing different types of credit.
For example, a firm pursuing an acquisition of another company could get a cashflow based loan from bank "A", and in parallel, it could get an asset-based loan to finance an equipment purchase from bank "B". Now, imagine a given firm experiences a demand shock leading to an increase in its demand for credit of the second type. If this were the case, the demand shock would apply only to the asset-based loan (bank "B") instead of overall credit (from both banks "A" and "B").
We can formalize the bias that arises when the true specification includes firm-loan specific shocks by decomposing the firm demand shock into twocomponents: = � + , namely, an overall firm demand shock ( � ) and a firm-loan specific shock ( ), where l denotes the loan type. The true model to be estimated would then be: where ∆ refers to loan growth of loan type l by firm f from bank b in time t.
We can assess the magnitude of the bias by comparing different estimates of equation (A2). In particular, we can first estimate � by including time-varying firm fixed effects � in equation (A2) but without including time-varying firmloan fixed effects : We can then estimate � by including time-varying firm-loan type fixed effects ( ) in the regression. The inclusion of firm-loan type fixed effects implies that identification is based on variation across banks in credit with the same firm and the same type of loan. Since � = , we can obtain the magnitude of the bias: The empirical model in (A1) can be generalized, following Amiti and Weinstein (2018), to: where ∆ refers to loan growth by firm f from bank b in time t, refers to the bank-lending channel (bank-specific supply shock), and refers to the firm borrowing channel (firm-specific demand shock).

APPENDIX 2. ASSUMPTIONS FOR IDENTIFICATION OF CREDIT SUPPLY
In the existing literature, there exist three approaches to identifying credit supply using loan level data. These approaches vary in the strength of the assumptions required for identification of credit supply and in the generality of their application. In this Appendix, we set out the assumptions of each method, starting with the method that is the most general but also requires the strongest set of assumptions.
The first approach is introduced by Amiti and Weinstein (2018) and it estimates the following model of credit growth: where ∆ refers to loan growth by firm f from bank b in time t, refers to a bank-specific shock, and refers to a firm-specific shock. The authors interpret as a (bank-specific) supply shock and as a (firm-specific) demand shock.
The key assumptions underlying the identification in AW are: (i) credit demand is firm-specific, not bank-specific; and (ii) credit supply is bank-specific, not firm-specific. Any shock that leads banks to lend more to some firms than others will result in supply being misinterpreted as demand. Such firm-specific supply of credit by banks could originate from lender specialization, differences in market power, or the value of bank relationships. Similarly, any shift in firm's credit demand from some banks relative to others may lead to credit demand being misinterpreted as supply. Such bank-specific demand by firms could originate from borrower specialization in certain type of credit products, differences in the type of collateral used for credit, or the existence of different loan types more generally.
The approach proposed by Khwaja and Mian (2008) instead estimates the following model of credit growth: The difference is in which denotes a bank-specific shock (typically a liquidity shock of some sort). Focusing on an exogenous bank shock imposes more structure on the model and potentially allows for stronger identification -to the extent that relates to bank-specific shocks -at the loss of generality. The key assumptions underlying KM are: (i) credit demand is firm-specific, not bankspecific; (ii) the time variation in is not correlated with shifts in credit demand; and (iii) the coefficient is constant across banks and firms.
The first assumption is identical in both approaches. Any shift in firm demand for loans from some banks relative to others (e.g. originating from differences in type of credit) may lead to demand being misinterpreted as supply, rendering this assumption invalid. The second and third assumptions are strictly weaker than the first assumption in AW. If there is lender specialization and firm demand is bankspecific, then these two assumptions are not met. This would be the case, for instance, if in response to a demand shock, firms demand more credit from banks that specialize in the type of credit that the firms demand (e.g. export credit). Paravisini et al. (2017) extends the KM approach by estimating: The new element, , denotes a demand shock that varies by firm and bank. Paravisini et al. (2017) interpret as a source of variation related to credit demand that varies with lender specialization. In their study, they look at banks specializing in specific export markets. This approach allows for stronger identification to the extent that relates to bank-specific firm demand.
The key additional assumption in Paravisini et al. (2017) is that the coefficient is constant across banks and firms. This assumption is not met if the relationship between bank-specific firm demand and credit growth is not stable over time. For instance, this could be due to changes over time in the type of firms that banks specialize in or shifts in bank market power that cause disproportionate shifts in the demand for certain type of loans. In the application of Paravisini et al. (2017) this would be the case if the type of banks operating in the firm's export markets suddenly changes (for instance, due to regulation or firm entry). In that case, the term may capture bank-specific supply shocks that will be misinterpreted as demand. However, to the extent that lender specialization is stable over time, the assumption underlying the approach in Paravisini et al. (2017) is likely to be met. This approach, therefore, requires the weakest identifying assumption of the three methods considered. Its application is, however, also more limited given the additional structure imposed on the model and the need to find suitable proxies for .  TABLE A1-ASSET TANGIBILITY AND LOAN TYPE   This table provides support for our loan classification. For Spanish firms, we have financial information from a source that is independent of the credit registry, taken from Almunia et al. (2018), for the period 2002:Q1-2010:Q4. Our focus is on assets' tangibility, measured as PPE/Total assets. Each observation in the sample is firm-quarter. The dependent variable is the share of credit of each loan type: asset-based loans, cash flow loans, trade financing, and leasing, respectively. Controls include: firm age, total assets, leverage ratio, a set of industry-year dummies, and year fixed effects. Standard errors are clustered at the firm level and are reported in parentheses. (1)