COLLATERAL CHANNEL AND SMALL BUSINESS LENDING

COLLATERAL CHANNEL AND SMALL BUSINESS LENDING Ari Hyytinen * and Ilkka Ylhäinen This version: May 15, 2015 Abstract We use a large panel of privately-held small businesses, matched to detailed regional and zoning data, to study the relation between regional house prices and the use of bank loans. Our econometric identification strategy exploits regional zoning differences. We find that controlling for firm-specific unobserved heterogeneity and endogeneity of regional housing prices, housing price increases facilitate the use of bank loans. There is no effect on broader (non-collateralizable) forms of debt and the effect is weaker after the financial crisis. The firms dividend payout does not vary with housing prices. These findings support the existence of a collateral channel. Keywords: small business finance, house prices, collateral channel JEL codes: G21, G30 * Jyväskylä University School of Business and Economics. Address: P.O. Box 35, FI-40014 University of Jyväskylä, FINLAND. Email: ari.t.hyytinen@jyu.fi. Jyväskylä University School of Business and Economics. Address: P.O. Box 35, FI-40014 University of Jyväskylä, FINLAND. Email: ilkka.ylhainen@gmail.com. Acknowledgements: This study uses firm-level data from Suomen Asiakastieto Oy. We would like to thank Etlatieto ltd for making this research possible by providing access to the data and Mika Pajarinen for help with the construction of the data. We thank Elias Einiö, Tuukka Saarimaa, Vesa Vihriälä, Thomas Åstebro, William Kerr, Diana Bonfim, and participants at the EEA s Annual Congress (Toulouse, 2014), the ECB Workshop on SMEs Access to Finance (Frankfurt, 2014), the Annual Meeting of the Finnish Economic Association (Kuopio, 2014), and VATT seminar (Helsinki, 2014) for useful comments. Ylhäinen thanks the OP-Pohjola Group Research Foundation and the Foundation for Economic Education for financial support

2 1 INTRODUCTION Is the collateral channel important for small businesses? Many policymakers and researchers alike seem to think so, because firms debt capacity is often found to be directly linked to their ability to pledge collateral. 1 Especially young and smaller firms with nearly non-existent prior credit history suffer from a limited access to bank credit, if they do not own enough collateralizable assets (see, e.g., Kashyap, Stein, and Wilcox, 1993, Gertler and Gilchrist, 1994; and Berger and Udell 1998). 2 It is therefore not surprising that for startups and small firms, housing and real estate owned by the founder-entrepreneurs appears to be an important source of pledgeable collateral (Black et al. 1996, Robb and Robinson 2014). And yet, what is less well understood is how important variation in the value of such housing assets is for the ability of small businesses to borrow 1 The distribution of collateralizable assets across the corporate and household sectors and the valuation of such assets over time can therefore have important implications to the real economy. For example, shocks to collateral values, and their interaction with credit limits, provide the micro foundation for the theories of credit-driven business cycles (Bernanke and Gertler 1989; Kiyotaki and Moore 1997). 2 Economic theory suggests that the reason for this tight link between collateral and use of debt is that collateral pledging is a means to alleviate both ex ante informational asymmetries and ex post contracting frictions between the lenders and borrowers (see, e.g., Stiglitz and Weiss 1981, Bester 1985, Boot, Thakor, and Udell 1991, Townsend 1979 as examples of the early theoretical contributions). The link may also reflect risk-management considerations: Firms can sustain higher leverage when collateral values are high (e.g., Rampini and Viswanathan 2010).

3 from banks. We address this question by using a large panel of privately held Finnish small businesses, matched to regional and zoning data, to study the relation between regional housing prices and the use of bank loans by the small businesses. Our analysis contributes to the emerging literature which suggests that shocks to the prices of real estate owned by firms and their owners may change firms access to credit and thereby their opportunities to enter a market and/or expand. In one strand of this literature, the focus has been on how changes in the values of collateral and collateralizable assets affect established, publicly listed firms. Gan (2007) uses a difference-in-differences approach and finds that publicly traded firms that suffered from greater collateral losses due to the burst of the Japanese land market bubble in the 1990s reduced investment and were less likely to sustain banking relationships. Furthermore, conditional on not discontinuing their banking relationships, these firms were able to raise less bank credit. Chaney, Sraer and Thesmar (2012) find that improvements in the collateral values of real estate increase the investments of publicly listed corporations in the US. The result also holds when the possibility that the prices of real estates are associated with local demand shocks is controlled for by using local housing elasticities as the instrument. Cvijanovic (2014) uses data on the actual real estate holdings of US listed firms measured in 1993 and land supply elasticities at the Metropolitan Statistical Areas (interacted with a measure of aggregate real estate prices) as an instrument for the potentially endogenous

4 local real estate prices. 3 She finds that an increase in the value of real estate increases firms leverage, reduces their costs of borrowing and lessens covenants. 4 In another strand of the literature, researchers have used either aggregated or regional data to study the link between the value of the real estate assets and firm behavior and entrepreneurial activity. In a seminal UK study, Black et al. (1996) provided time-series evidence that improvements in the value of housing equity are positively associated with the formation of new businesses. Balasubramanyan and Coulson (2013) used business starts in the Metropolitan Statistical Areas to document that the local US housing prices are robustly associated with small (but not with larger) business starts, as measured by employment. Recently, Adelino et al. (2014) have used county-level data from the US and classifications by establishment size and industry to explore the effects of shocks to home values on employment in small firms. They find that in areas 3 Lin (2014) provides closely related evidence, using partly overlapping sample of US listed firms from 2002 to 2011. 4 There is also a strand of the literature, which investigates the importance of liquidation values and redeployability of assets for firms financing choices (e.g., Benmelech, Garmaise and Moskowitz 2005 and Benmelech and Bergman 2009 for interesting US evidence). Recently, Norden and van Kampen (2013) have documented that easily redeployable assets, such as property, plant and equipment, are a key part of the mechanism through which the collateral channel manifests itself. The findings of Giambona, Golec, and Schwienbacher (2014) suggest that real estate assets have larger effect on leverage than other tangible assets, especially if the firms are financially constrained.

5 where the housing prices increase, the employment of small businesses grows. They also find that there is no corresponding effect on employment at the larger firms. Firm-level analyses of the effects of shocks to the value of the real estate and housing assets owned by small businesses, startups and their ownermanagers are more scant. Robb and Robinson (2014) use Kauffman Firm Survey to document that the (surprisingly extensive) reliance of US startups on formal bank lending as a source of liquidity is closely linked to the ability (and willingness) of the founder-entrepreneurs to pledge their personal assets as collateral for the loans. 5 Robb and Robinson find, in particular, that startups located in the areas that have greater elasticity of housing supply carry more bank debt. They argue that this finding is consistent with the collateral channel, because the greater elasticity makes housing prices less volatile and thus housing assets more pledgeable. Schmalz, Sraer and Thesmar (2014) study firm- and individual level data from France and compare entrepreneurial entry and subsequent expansion of homeowners with those of renters within the same geographical region. Using this identification strategy, they find that owning a house in an area where house prices have risen increases the probability of a homeowner becoming an entrepreneur relative to renters. Their findings also suggest that entrepreneurs create larger firms and have better survival prospects when the 5 The availability of debt financing appears to have a causal impact on the survival and subsequent growth of entrepreneurial firms; see e.g. Fracassi, Garmaise, Kogan, Natividad (2014).

6 increase in the value of their houses is greater. Corradin and Popov (2015) find that increases in the housing equity raise the share of individuals transiting into self-employment in the US, suggesting that housing wealth alleviates credit constraints. This study extends and complements the previous literature in five ways: First, while insightful, the prior analyses by Robb and Robinson (2014) and Schmalz et al. (2014) do not directly evaluate the effect of regional housing prices on the use of bank loans by privately held small businesses. 6 And yet, the housing-asset driven collateral channel ought to be especially important for such firms borrowing, because bank debt is often secured and because the smaller firms typically cannot easily tap other sources of external finance, such as public debt or bond markets. Second, we analyze the link between local housing prices and the use of bank debt in a European-style bank-centered financial system (see, e.g., Hyytinen, Kuosa and Takalo 2003; Korkeamäki, Rainio and Takalo 2013). This empir- 6 Robb and Robinson (2014) consider startups and they do not link their use of bank debt to housing prices directly. The findings of Schmalz et al. (2014) refer mostly to non-financial outcomes. They document that housing capital gains are associated with greater debt levels. However, it is not entirely clear to what extent their debt measure also refers to noncollateralizable and/or non-bank borrowing. Corradin and Popov (2015) document a positive correlation between entrepreneurship and changes in mortgage debt. Kleiner (2014) provides evidence from the UK that increases in the small firms real estate holding values increase their investment, employment, and total debt, but does not directly address whether this is due to greater use of bank debt (despite providing numerous other checks).

7 ical setup is novel (and differs from e.g. that of Robb and Robinson 2014) and provides us with new and complementary insights on an important mechanism through which the collateral channel may be at work. Third, a key econometric challenge that also the prior work has faced is that regional (demand) shocks may simultaneously affect the use of bank loans and housing prices. This source of endogeneity makes the identification of the collateral channel problematic. It is also possible that better regional availability of bank loans inflates housing prices. To address these concerns, we have matched the small business and local housing price data with municipal-level zoning data. 7 Inspired by the recent literature that instruments local housing prices with housing elasticities (see, e.g., Chaney et al. 2012; Cvijanovic 2014; see also Saiz 2010), we use a novel proxy for the availability of developable land for construction and use it to instrument the housing price changes. Our primary instrument is the bureaucratically and politically determined share of planned area in the municipalities. The validity of this instrument rests on the assumption that after controlling for fixed regional and firm-specific effects, the differences between the municipalities in this share should be unrelated to the contemporaneous use of bank loans by the small businesses. Our empirical analysis also exploits politically-driven municipal mergers as an additional source of exogenous variation in the supply of land available for construction. 7 The micro-level panel data is originally compiled by Asiakastieto ltd from various official registers. Asiakastieto is a commercial provider of credit information and firm-level financial statement data.

8 Fourth, a unique feature of the firm-level data available to us is that, besides allowing us to control for unobserved firm-specific heterogeneity, it includes a time-varying credit score ( rating ) for each firm in the data. We use this credit score as a further control, since it allows us isolate the effects of the housing price shocks on the use of bank debt from the shocks to the creditworthiness of the small businesses. As far as we are aware, the prior work has not considered the possibility that regional shocks change both the creditworthiness of firms located a region and the region s housing prices. Finally, the recent studies by Jensen, Leth-Petersen, and Nanda (2014) and Pekkala-Kerr, Kerr, and Nanda (2014) argue that the positive relationship between house prices and entrepreneurship could be due to a wealth effect instead of the collateral channel. This could be the case, if e.g. entrepreneurs risk aversion decreases when the value of their (housing) assets increases. After carefully controlling for the wealth effects using a Danish mortgage market reform, Jensen et al. (2014) find that the effect of relaxing credit constraints on entry is quite small and that the enhanced availability of credit may increase the entry of relatively low quality firms. Building on the recent findings by Bliss, Cheng and Denies (2014) on the relation between financial constraints and dividend payout ratios, we also address this issue, providing thus further empirical evidence that bear on the debate about the wealth effect hypothesis. The main result of this paper is that if housing prices increase in a region, it facilitates the use of bank loans by the privately held small businesses located in the region. Our baseline dynamic estimations for the smallest micro firms (that have less than five employees) imply that in the long-term, a hundred eu-

9 ro per square meter increase in regional house prices increases the use of bank debt of such firms by about 1.61 percentage points. This is not a negligible effect, because the ratio of bank debt to total assets is for them a bit more than ten percent. Five further findings of ours allow us to argue that this effect is due to the collateral channel. First, the effect appears to be somewhat more pronounced among the smallest micro firms. These firms are indeed the ones for whom collateralizable housing assets should be most relevant. Second, the effect of the housing prices appears to be a bit larger for firms with more tangible assets. This suggests that at least a part of the effect is related to the real estate and housing assets owned by the firm. 8 Third, we find that there is an effect of housing prices on the use of bank debt, but not on the total use of debt. This finding provides additional support for the existence of a collateral channel, because total debt contains non-collateralizable and/or non-bank debt items, such as trade credit. The effect on the total debt should therefore be weaker or nonexistent altogether. Fourth, our estimates suggest that the effect of housing pric- 8 Our empirical setup is purposefully agnostic about the actual ownership of the real estate and/or housing assets. We have two reasons for this modeling choice. First, a priori, it is not clear whether it is the real estate owned by a small business (the corporation) or by its owner-entrepreneurs that matters for the debt capacity of the firm. The prior literature suggests that both may matter. Second, in small firms, even if they are incorporated, the line between the corporate and personal real estate assets easily gets blurred. For example, it is not uncommon that an entrepreneur lives in a house that her firm owns, or that the firm s office is located at the home of its owner(s).

10 es on the use of bank loans is slightly weaker after the onset of the financial crisis. This, too, is consistent with the collateral channel being at work: Because the crisis originated from the US housing markets and highlighted the uncertainty in how imperfectly housing assets protect lenders from the credit losses, it is possible that Northern-European banks have, as a result, become less willing to accept housing or real estate assets as collateral for their small business lending. Finally, we cannot find a positive relation between dividend payout and increased housing prices. Such a relation should be present if there is a wealth effect on consumption that calls for a greater payout ratio. Albeit indirect, this evidence casts doubt on the view that a wealth effect is driving our findings. The rest of the paper is organized as follows: Section 2 describes the data. Section 3 introduces the empirical methods. Section 4 presents our baseline results and a set of robustness tests. Section 5 concludes. 2 ENVIRONMENT AND DATA 2.1 Institutional environment Finland has a bank-centered financial system (see, e.g., Hyytinen et al. 2003; Korkeamäki et al. 2013). The banking sector is highly concentrated and dominated by three large banking groups. While a large part of the small and medium sized firms do not use external financing to a significant extent, the firms that have external financing needs

11 rely on bank borrowing a lot. For example, in a survey conducted in 2004 (i.e., at the beginning of our sample period), around 60-70% of those small and medium sized firms that had raised external financing, indicated that banks were their primary source of external financing (SME-barometer 2004). The most important reasons why debt financing is typically needed are the financing of real estate or equipment investment and working capital. Only a small fraction of the small and medium sized firms use financing that requires the use of covenants. The use of collateral to secure borrowing appears to be quite widespread: The available survey evidence indicates that in 2004, around 40% of the firms used debt financing that had been backed by collateral that the entrepreneurs and/or their relatives had pledged (SME-barometer 2004). Because a nonnegligible part of the debt in the firms balance sheet comes from the sources that are typically not backed by collateral (e.g. from other firms in the form of trade credit, or from the government-backed lenders in the form of subsidized debt or guarantees), the share of borrowing from the banking sector that is backed by collateral is likely to be quite a bit higher than 40%. A bit more than ten percent of the Finnish firms complained at the same time that the lack of collateral was a financing obstacle. Housing prices increased relatively steadily in 2004-2008, which is the period our baseline analysis covers. The increase was greater in the greater Helsinki capital area, which is the most notable urban area in Finland, than in the rest of the country, which is heterogeneous but which can, on average, be regarded more rural. The positive trends in both types of regions were disrupted

12 in 2008, when the financial crisis hit the world. It can be argued that the housing prices are relatively volatile in Finland, because the supply is relatively sticky and because the housing market is sensitive to interest rate changes (due to e.g. many mortgages being tied to a variable references rate). More generally, because the overall performance of Finland s export-oriented economy is sensitive to global economic shocks, also the developments in the housing markets are likely to mirror such shocks. 2.2 Data sources The data used in this study is compiled from three components: firm-level panel data, regional house price data and municipal-level zoning data. Our firm-level data is an unbalanced panel of Finnish firms, compiled initially by Asiakastieto ltd, a provider of firm and credit score data in Finland. The data we use in our baseline analysis cover years 2004-2008, but we also make use of longer sample periods in the robustness tests. The data include financial statements and associated supplementary data, such as the address (location) of the firms. The financial statement data originate from the Finnish Trade Register, an official register of Finnish firms. The dataset also contains a commercial credit score (and the associated letter grade credit ratings ) for each firm, as produced by Asiakastieto. The regional house price data is measured at the zip code -level and are obtained from the Statfin database of Statistics Finland. These data refer to prices (in euros) per square meter of old privately financed condominiums. These annual statistics are compiled from the asset transfer tax data of the Finnish Tax

13 Administration. The house prices data cover the average prices of the completed transactions and the number of transactions in each zip code area. In our baseline analysis, we use price data at the municipal level, which we calculate as a weighted average using the five-digit zip code data. We use the number of transactions in the five-digit zip code areas as the weight. The municipal-level zoning data are obtained from the OIVA databases of the government s environmental management services, operated by the Ministry of the Environment, the Finnish Environmental Institute, the Centres for Economic Development, Transport and the Environment (ELY Centres), and the Regional State Administrative Agencies. The zoning data used in the econometric analysis refer to the percentage of the planned area of the total area of municipalities. 9 The OIVA database also allows us to identify the mergers of the municipalities that have taken place during our sample period. 2.3 Construction of the sample The original firm-level data include all firms for which Asiakastieto has been able to collect financial statements data from the official registers. Using this as the starting point, the estimation sample is constructed as follows: First, we only include non-farm and non-financial corporations (limited liability firms). We then remove firms that belong to mining, utilities, construction, real estate, gov- 9 The zoning data are based on the information available in the OIVA database as of September 2013.

14 ernmental, or non-profit industries. 10 Second, firm-year observations for which total assets are negative are dropped from the sample. Third, the housing prices are matched to the firm-level data using the zip codes of firms addresses and year. 11 Fourth, the zoning data are matched to the financial statement data based on the municipality codes and year. 12 This matching means that if a firm does not have matching information on the zip code or on the municipality, it is dropped from the sample. Our baseline analysis focuses on years 2004-2008 and thus refers to a period before the financial crisis. To focus on the smaller end of the firm size distribution, the estimation sample is further restricted to firms having fewer than 50 10 The following industries were dropped from the sample based on the Standard Industry Classification (SIC) 2008, and if not available, SIC 2002: Agriculture, forestry and fishing; Mining and quarrying; Electricity, gas, steam and air conditioning supply; Water supply; sewerage, waste management and remediation activities; Construction; Financial and insurance activities; Real estate activities; Activities of membership organizations; Public administration and defense; Compulsory social security; Activities of extraterritorial organizations and bodies; Industry unknown. 11 Because the matching of the house prices data is made at the zip code level, firms having their addresses in a box number location (i.e., zip codes ending with one or in nonstandard way) are removed from the estimation sample. We also remove firm observations with missing zip code level housing prices (due to too few transactions in the area). This restriction has the added benefit that the municipal- and zip code level analyses use identical samples and are therefore directly comparable. 12 The municipality divisions are based on the situation that prevailed in 2013.

15 employees. Furthermore, the panel is restricted in such a way that firms having less than three time series observations or gaps in the firm-specific time series are removed from the sample. This sample restriction is imposed, because the econometric models used in this study require using lagged variables. 13 A data limitation that we have to deal with is that we do not have data on the real estate or housing assets owned by the small businesses or by their owner-entrepreneurs. It is not uncommon that an entrepreneur lives in a house that her firm owns, or that the firm s office is located at the home of its owner(s). Indeed, a manual inspection of the data suggested that a non-negligible number of firms in the sample are located in residential areas. Both kinds of assets can matter for the debt capacity of the small businesses. We take this into account in three ways: First, we try to err on the conservative side and measure the housing prices mainly at the more aggregated municipal level. This is appropriate, if entrepreneurs homes are located nearby but not too far from where their firms 13 Three additional observations about the construction of the estimation sample are in order: First, as we will explain below, our empirical analysis divides the firms into three size classes. Missing observations for a firm are allowed, if its size classification changes during the sample period and if there remains a sufficient number of observations for a firm to be included in the estimation. Second, we also use a longer sample period of 2004-2011 in the analysis. Since the employment figures used for the size classification do not cover all firms in year 2011, the missing values are replaced with the values from the previous period. Third, we utilize the same sample restriction criteria for both samples (i.e., we remove firms with gaps in the longer sample also from the baseline sample). We do this to make two samples more comparable.

16 are. 14 We also use the five-digit zip code level price data in robustness tests. The use of this more disaggregated data would be appropriate, if the relevant real estate assets are located at, or very nearby, the firms address. 15 Second, we explore whether our findings change when we focus on the small businesses with more tangible assets. If they become stronger, it suggests that the real estate assets owned by the firm may matter more. Third, our main results rely on instrumental variable methods, which reduce the concern that measurement error biases our results. 2.4 Variable definitions Our primary dependent variable is Bank debt. It refers to the ratio of outstanding debt from financial institutions to total assets at period t. To manage outliers, Bank debt values above one are set to one. Most of these instances occur when a firm has negative equity. We use as an alternative outcome variable Total debt. It is equal to the ratio of total debt to total assets. As we explain in greater detail below, this alternative outcome variable is used to demonstrate that the effects we uncover are related to the collateral channel. Our key explanatory variable is House prices. In our baseline analysis, it refers to the average municipal level prices of old (i.e., previously used and 14 We have also repeated the analysis at the two-digit zip code level, which implies that the areas can cover multiple municipalities; see the robustness tests for details. 15 Note that while some five-digit zip code areas could in principle belong to more than one municipality, we have allocated each five-digit zip area to its main municipality using the zip code data provided by the Finnish postal services.

17 owned) condominiums in each region and is measured in thousands of euros per square meter. We also report robustness checks where we use both more and less disaggregated measures of the housing prices. Our instruments are defined as follows: Planned area measures the ratio of the planned area (i.e., town plan zone) to the total area in a given municipality. We allow the instrument have a nonlinear impact on the house prices by also using Planned area squared (= 100! (Planned area)^2). Our empirical analysis also exploits municipal mergers as an additional source of exogenous variation in the supply of land available for construction. Municipal merger is a dummy taking a value of one in the year when the municipal merges and during all periods thereafter, and is zero before the merger. We discuss the relevance and validity of these instruments in greater detail below. We use the following control variables: Firm age is measured by ln(1+age) where age is calculated by subtracting the year of birth from the current year. 16 Firm size is measured by ln(total assets), where the total assets are measured at t-1. Credit score refers to the commercial credit score of each firm at t-1, as computed by Asiakastieto, with lower values indicating better creditworthiness. We scale the raw credit score for the estimations by dividing it by 100. 17 16 The year of birth is defined as the year the firm was registered in the Finnish Trade Register. In the rare instances when the firm age turned out to be negative, the observation was dropped. 17 In the raw data, the score obtains values from 3 to 99. The credit scores for 2006 are not available in the data because of a change in the dataset. Therefore, values from the previous

18 We keep the list of firm-level control variables short for two reasons. On the one hand, we try to avoid the problem of bad controls by not adding variables that, for instance, anticipated use of bank loans might affect. Second, the credit score is a summary measure, which captures many typically used controls, such as the (lagged) equity ratio. 2.5 Descriptive statistics The descriptive statistics for firms with fewer than 50 employees are provided in Table 1. The table covers the sample period 2004-2008 and includes in total 26 097 firms. 18 The table shows that the mean percentage of bank debt in the balance sheet is 13.3%. The numbers imply that the mean age of the firms in the sample is 12 years and the average size of the balance sheet is about 194 000 euros. The average credit score is 25, which indicates a credit rating A+ (i.e., satisfactory + ). The average municipal- and zip code level house prices are over 1850 and 1950 euros per square meter, respectively. The percentage of the planned area of the total area of the municipality is on average 16.1%. While not reported in the table, the SIC 2002 industry composition of the estimation sample is as follows: Manufacturing 16.6%, Wholesale and retail trade; repair of motor vehicles, motorcycles and personal and household goods 31.2%, Hotels period are used for that year. If such a value was missing, the following year s credit score was used instead. 18 Table A1 in the Appendix reports the descriptive statistics for the longer sample period of 2004-2011. This larger sample includes 40 849 firms.

19 and restaurants 5.1%, Transport, storage and communication 9.2%, (Real estate), renting and business activities 27.7%, Education 1.3%, Health and social work 5.2%, Other community, social and personal service activities 3.7%. Table 1: Descriptive statistics for 2004-2008 Variable mean sd min p50 max NT Bank debt 0.133 0.219 0.000 0.000 1.000 91036 ln(1+age) 2.554 0.680 0.693 2.639 4.727 91036 ln(total assets) 12.177 1.572 5.011 12.139 21.529 91036 Credit score 0.253 0.184 0.030 0.220 0.990 91036 House prices* 1.952 0.982 0.431 1.701 5.254 91036 House prices** 1.853 0.782 0.451 1.654 3.346 91036 Planned area 0.161 0.148 0.000 0.096 0.993 91036 Planned area squared 4.788 7.787 0.000 0.922 98.605 91036 Municipal merger 0.087 0.281 0.000 0.000 1.000 91036 Notes: The table reports descriptive statistics for the sample of firms smaller than 50 employees for 2004-2008. The statistics include the mean, standard deviation, minimum, median, and maximum, respectively. Bank debt is a ratio of loans from financial institutions to total assets. Ln(1+Age) is a natural logarithm of firm age since incorporation (plus one) in years. Ln(Total assets) is a natural logarithm of the total assets in euros at t-1. Credit score measures the observed creditworthiness of firms at the interval 3-99 (scaled by dividing by 100), where lower values imply higher creditworthiness. House prices measure the average regional prices of previously owned condominiums at the zip code (*) and municipal (**) level in thousand euros per square meter, respectively. Planned area measures the share of planned area of the total area in the municipality. Planned area squared is defined as (Planned area*100)^2/100. Municipal merger is a dummy taking a value equal to one in and all the periods after the municipal merger. NT is the number of firm-year observations. Table 2 reports the correlation matrix for the variables used in the estimations. It shows, firstly, that the house prices are negatively correlated with the bank debt. Secondly, the share of planned area in the municipality is positively correlated with regional house prices. A potential explanation for these correlations is that in regions that are doing well in economic terms (e.g., in larger cities and urban areas, with more economic activity), house prices are higher. It is plausible that in such regions, small businesses can rely more on internal financing and have better access to external financing other than bank debt and that the share of (already) planned and thus developed area are greater. These charac-

20 teristics of the raw data suggest that it is important to carefully control for permanent firm- and region-specific heterogeneity and to allow for potential nonlinearities in the relation between the share of planned area and regional house prices in the econometric analysis. Table 2: Correlation matrix for 2004-2008 Bank debt Bank debt 1.0000 ln(1+age) ln(1+age) -0.0694* 1.0000 ln(total assets) ln(total assets) 0.0641* 0.2085* 1.0000 Credit score Credit score 0.2741* -0.2594* -0.2816* 1.0000 House prices House prices -0.1636* -0.0047-0.0340* -0.0017 1.0000 Planned area Planned area -0.1299* -0.0009-0.0185* -0.0009 0.6069* 1.0000 Planned area squared Planned area squared -0.0866* 0.0008-0.0103* -0.0037 0.3799* 0.9023* 1.0000 Municipal merger Municipal merger 0.0404* 0.0023 0.0238* -0.0132* -0.1983* -0.2682* -0.1846* 1.0000 Notes: The table reports pairwise correlations between the variables for 2004-2008. Bank debt is a ratio of loans from financial institutions to total assets. Ln(1+Age) is a natural logarithm of firm age since incorporation (plus one) in years. Ln(Total assets) is a natural logarithm of the total assets in euros at t-1. Credit score measures the observed creditworthiness of firms at the interval 3-99 (scaled by dividing by 100), where lower values imply higher creditworthiness. House prices measure the average regional prices of previously owned condominiums at the municipal level in thousand euros per square meter. Planned area measures the share of planned area of the total area in the municipality. Municipal merger is a dummy taking a value equal to one in and all the periods after the municipal merger. The correlations significant at the 5 percent level or better are marked with an asterisk (*). 3 EMPIRICAL METHODOLOGY 3.1 Econometric model and methods of estimation We build our analysis on the following econometric model:!!" =!"!,!!! +!!!"! +!!!,! +!! +!! +!!", (1)

21 where!!" is the firms use of debt,!!" is a vector of firm-level control variables,!!,! is the measure of municipal- or zip code level regional house prices,!! is a firm-specific fixed effect absorbing also the fixed regional characteristics,!! refers to year dummies that control for the overall macroeconomic conditions, and!!" is an error term. We estimate this model in three ways: First, we impose restriction δ = 0 (no dynamics) and estimate (1) using the standard fixed effects (FE) within estimator. Second, we continue to impose δ = 0 and use the first-differenced instrumental variables (FD-IV) estimator. This estimator assumes a weak exogeneity of the regressors and allows us to account for the endogeneity of the house prices. Weak exogeneity allows feedback effects between the dependent and independent variables. Third, we allow for dynamics (δ 0) by using the Arellano-Bond GMM estimator. 19 3.2 Endogeneity of house prices and instruments If there are unobservable time-varying regional factors that affect both the firms use of bank loans and regional house prices, house prices in (1) can be endoge- 19 The Arellano-Bond estimator used in this paper refers, bar a few exceptions in the robustness tests, to the two-step GMM estimator with robust Windmeijer-corrected standard errors. The estimator requires that the first-differenced residuals do not have second-order serial correlation. This assumption can be tested by the Arellano-Bond serial correlation test for the first-differenced residuals. The Sargan test for overidentification restrictions can be used to test the validity of the moment restrictions, assuming that the errors are independent and identically distributed. In the robustness section, we also report some one-step GMM estimates and robust Hansen tests.

22 nous. Examples of such factors are demand shocks and increased regional availability of bank loans. They may both inflate house prices and increase the use of bank debt by the small businesses. This endogeneity problem can be solved with an instrumental variable that is correlated with the house prices but uncorrelated with the dependent variable, besides its effect via the house prices. The previous literature suggests that restrictions in zoning and the availability of developable land are highly correlated with real estate prices (see, e.g., Saiz 2010). Based on this logic, the earlier studies (Chaney et al. 2012 and Adelino et al. 2014) on the house-price driven collateral channel have used a measure for the inelasticity of land supply, sometimes interacted with another variable, as an instrument for house prices. In areas where land supply is relatively inelastic, shocks to the housing demand translate into higher house prices. The main instrument used in this paper, Planned area, refers, as we explained above, to the percentage of the planned area of the total area in the municipality of a firm s location. This percentage is a bureaucratically and politically determined quantity: The decision to change the planned area is an outcome of a relatively complex bureaucratic and political zoning process that is subject to a number of laws and regulations. The ultimate decision-maker is the municipal council, the members of which are elected in every fourth year. The Finnish municipalities use a proportional election system and multiple parties participate in the local municipal elections. The multi-party councils make the decisions about the new areas that are subjected to planning, but are constrained by what the prior councils have decided in the past. The regional differences in zoning between the different municipalities thus mirror, besides the

23 local economic conditions, the political strength and views of the various local political actors in the current and past councils. The validity of Planned area as an instrument rests on the assumption that after controlling for fixed regional and firm-specific effects, the differences between the municipalities in this share should be unrelated to the use of bank loans by the small businesses. In particular, even if an increase in the planned area mirrored a local demand shock, the increase is likely to only happen after a notable time lag due to the time consuming and uncertain bureaucratic nature of the zoning process. The available anecdotes and estimates indicate that a median town plan zoning process takes over 10 months from the official initiation before becoming effective (see, e.g., Rinkinen 2007). What this means is that a local demand shock in period t, which may boost the demand for small business loans in period t, is likely to show up as an increase in the planned area in period t+1 at earliest. Most likely, the planned area increases even with a greater delay, as the cited figure assumes that there are no complaints about the proposed additional zoning. Such complaints could easily double or triple the lag (because of possible additional rounds of political bargaining and/or court processes). A further reason for why the lag between the perceived demand shock and the increase in the planned area is likely to be greater in practice is that before a new piece of zoning can be officially proposed, there is a considerable

24 amount of preliminary planning that precedes the official initiation of the zoning process. 20 The previous literature suggests that Planned area ought to be a relevant instrument (see, e.g., Saiz 2010, Chaney et al. 2012 and Adelino et al. 2014): In the areas where the share of (already) planned area is low to begin with, an increase in the share of planned area increases land supply on the market and therefore lowers house prices. When the local potential supply of land is elastic, an increase in the housing demand ought to translate into more construction rather than in higher land and housing prices. There is, however, a potential nonlinearity between regional house prices and planned area, as there are a couple of reasons why the relation between house prices and planned area may turn positive after a threshold: First, in the areas where the share of (already) planned area is high, new developable land available in prime or easily usable locations may be scant. This suggests that in such areas land supply is increasingly inelastic. This means that at some point, the quantity is increasingly hard to adjust and thus that the shocks to the housing demand translate into higher house prices. Second, in the areas where the share of (already) planned area is high to begin with, additional zoning may disproportionately affect the desirability of the areas that have been planned 20 A second point worth noting is that lobbying pressure is not likely to be a great concern for us, because the political bargaining power (e.g., for more construction land) of a very small business is limited and because as a group, they face a coordination problem and may even have conflicting interests.

25 and developed earlier. If additional zoning improves the overall infrastructure of an area, this may increase the value of existing, already developed land. To allow for this type of nonlinearities in the simplest possible way, we add the second-order polynomial of the planned area as the instrument. 21 In short, Planned area is a relevant and valid instrument for two reasons. On the one hand, there is a degree of uncertainty in the zoning process. Once the final decision is made, the zoning decision affects housing supply and thus prices. On the other hand, due to the bureaucratic lags, the final decision is not correlated with the contemporaneous (unobservable) determinants of the use of bank debt. We also exploit municipal mergers, as captured by Municipal merger, as an additional instrument. The decision by two or more municipalities to merge is an outcome of a relatively complex political process that is subject to a number of procedural rules and strategic political behaviour (on the political determinants of these mergers, see Hyytinen, Saarimaa and Tukiainen 2014). Municipal mergers increase the amount of land available for construction, as they allow for better coordination and zoning and planning across the former municipal borders. Holding other things constant, they should therefore increase supply elasticity. 21 We display the estimated relation between the house prices and planned area in the Appendix.

26 3.3 Calculation of standard errors For our baseline analyses, we cluster the standard errors at the firm level. We use this level of clustering for three main reasons. First, the firm-level clustering adjusts standard errors for (autocorrelated) firm-specific shocks in the use of bank loans. This choice is consistent with e.g. the findings from the literature on zero-leverage firms, which emphasize the firm-specific persistence and dynamics of leverage choices (see, e.g., Strebulaev and Yang 2013). Second, the firmlevel clustering results in more balanced cluster sizes than the regional-level clustering: Indeed, the municipal-level clustering does not seem to be a particularly viable option for us, because it would result in clearly unbalanced cluster sizes. This would violate the assumption of equal cluster sizes (see, e.g., MacKinnon and Webb 2014) and could decrease the effective number of clusters. This may lead to incorrect inference on standard errors (see, e.g., Cameron and Miller 2015). Third, our main qualitative conclusions do not depend on the chosen level of clustering: To illustrate this, we report a set of robustness tests that use both municipal- and zip code level clustering in the robustness section. 4 RESULTS 4.1 Baseline results Our baseline results refer to period 2004-2008 and use regional house prices that are measured at the municipal level. We report the baseline results for the FE, FD-IV and the Arellano-Bond methods.

27 We start with the FE (within) method, reported in Table 3. Columns (1)- (3) use three different subsamples, determined on the basis of firm size. The three subsamples overlap and consist of firms having less than five, ten, and fifty employees, respectively. The coefficient of house prices is positive, but statistically somewhat imprecisely measured. This finding is consistent with the view that higher housing values are associated with greater use of bank debt by small businesses. 22 The FE results also suggest that relative to the size of their balance sheet, younger and larger firms use more bank debt. The credit score is insignificant. However, it is worth pointing out that these estimations do not correct for the potential endogeneity of housing prices and require that the other explanatory variables are strongly exogenous. Table 3: FE (within) estimates for 2004-2008 (1) (2) (3) Size class <5 employees <10 employees <50 employees Dependent variable Bank debt Bank debt Bank debt House prices 0.016 * 0.010 0.012 * (0.009) (0.007) (0.007) ln(1+age) -0.037 *** -0.039 *** -0.040 *** (0.010) (0.008) (0.007) 22 We want to make two further observations about these baseline results. First, we would not get a positive coefficient for the house prices if we did not control for the unobserved permanent firm-specific heterogeneity: Unreported OLS regressions reveal that the coefficient is negative and significant, in line with the raw correlations, when no controls are used. The coefficient remains negative but turns insignificant when municipal-level regional fixed effects and two-digit industry fixed effects are included in the regression. Second, we acknowledge that the distribution of the dependent variable has a probability mass at zero. We return to this feature of the data in the robustness tests.

28 ln(total assets) 0.028 *** 0.028 *** 0.027 *** (0.003) (0.002) (0.002) Credit score 0.004 0.002 0.004 (0.008) (0.006) (0.005) NT 55549 73002 91036 rho 0.804 0.804 0.805 r2 0.010 0.010 0.009 Notes: The table reports the fixed effects (within) estimates of the effects of regional house prices on small firms use of bank loans over the 2004-2008 period. The results are reported separately for the size classes of firms having less than five, ten, and 50 workers, respectively. The dependent variable Bank debt is defined as a ratio of bank loans to total assets at t. The independent variables are defined as follows: House prices measure the average prices of previously owned condominiums at the municipal level measured in thousand euros per square meter at t. Ln(1+Age) is a natural logarithm of firm age since incorporation (plus one) in years at t. Ln(Total assets) is a natural logarithm of total assets in euros at t-1. Credit score measures the observed creditworthiness of firms at the interval 3-99 (divided by 100) at t-1, where lower values imply higher creditworthiness. All the models include year dummies. NT is the number of firm-year observations. Rho measures the intra-class error correlation. R2 stands for r-squared. Standard errors in parentheses: * p < 0.10, ** p < 0.05, *** p < 0.01. The standard errors are clustered at the firm level. The baseline results for the FD-IV method are shown in Table 4. These estimations allow house prices to be endogenous and only require weak (rather than strong) exogeneity of the other explanatory variables. The first stage regressions, reported at the bottom of Table 4, indicate that the zoning instrument and its square are correlated with regional house prices and have statistically highly significant first-stage coefficients. The municipal merger instrument is highly significant at the 1% level, too. The magnitudes of the first stage estimates reported imply that if the share of planned area increases from the sample mean by ten percentage points, housing prices decrease by 30-40 euros per square meter (i.e., by 1.5-2.2%). The merger decreases the prices by a bit more than 40 euros per square meter (i.e., by 2.3-2.4%). These negative relations are in line with the view that less sticky housing supply results in lower house prices. We also find that the instruments do not appear to be weak, nor is their validity rejected by the overidentification test: The Kleibergen-Paap Wald F statistics for weak identification of the first-stage regression vary be-

29 tween 926.916 and 1922.062 and thus greatly exceed the Stock-Yogo critical values. Hansen s J statistics do not reject the null hypothesis that the overidentification restrictions are valid. The results in Table 4 indicate that higher regional house prices increase the use of bank loans by small businesses. The coefficient is a bit larger for the smaller firms (columns (1) and (2)). The coefficient of the house prices is significant at the 10% level in column (1) and at the 5% level in columns (2) and (3). We also find that younger and larger firms and firms with a better credit score use more bank debt. Table 4: FD-IV estimates for 2004-2008 (1) (2) (3) Size class <5 employees <10 employees <50 employees Dependent variable Bank debt Bank debt Bank debt House prices 0.112 * 0.131 ** 0.102 ** (0.068) (0.058) (0.048) ln(1+age) -0.028 *** -0.030 *** -0.033 *** (0.010) (0.009) (0.007) ln(total assets) 0.011 *** 0.010 *** 0.008 *** (0.003) (0.002) (0.002) Credit score -0.011-0.012 ** -0.013 ** (0.007) (0.006) (0.005) NT 39336 51799 64939 Hansen s J statistics 2.722 3.282 1.686 [0.2564] [0.1938] [0.4303] Kleibergen-Paap Wald statistics 926.916 1378.302 1922.062 First-stage House prices House prices House prices Planned area -1.725 *** -1.589 *** -1.546 *** (instrument) (0.092) (0.076) (0.067) Planned area squared 0.031 *** 0.030 *** 0.030 *** (instrument) (0.001) (0.001) (0.001) Municipal merger -0.044 *** -0.042 *** -0.043 *** (instrument) (0.002) (0.002) (0.002) Notes: The table reports the first-differenced IV model (GMM) estimates on the effects of regional house prices on small firms use of bank loans over the 2004-2008 period. The results are reported for the size classes of firms having less than five, ten, and 50 workers, respectively. The dependent variable Bank debt is defined as a ratio of bank loans to total assets at t. The independent variables are defined as follows: House prices measure the average prices of previously owned condominiums at the municipal level measured in thou-

sand euros per square meter at t. The instruments include Planned area, its square, and Municipal merger dummy taking a value of one in and after the period of merger. Ln(1+Age) is a natural logarithm of firm age since incorporation (plus one) in years at t. Ln(Total assets) is a natural logarithm of total assets in euros at t-1. Credit score measures the observed creditworthiness of firms at the interval 3-99 (divided by 100) at t-1, where lower values imply higher creditworthiness. All the models include year dummies. NT is the number of firmyear observations. The first-stage results for the instruments are reported omitting the rest of the output from the table. Hansen s J statistics for overidentifying restrictions [p-values in brackets] and the Kleibergen-Paap Wald F statistics for weak identification of the first-stage regression are provided. Standard errors in parentheses: * p < 0.10, ** p < 0.05, *** p < 0.01. The standard errors are clustered at the firm level. Our baseline results for the dynamic version of model (1), estimated by the Arellano-Bond method, are provided in Table 5. For the implementation of this method, we assume that firm size is pre-determined and that credit score is endogenous. These assumptions are conservative, as they allow for feedback effects with the dependent variable and take into account the fact that equity ratio is a determinant of the credit score. The conventional Arellano-Bond instruments consist of the lagged levels of the endogenous and pre-determined variables and the first-differenced exogenous variables. The share of planned area, its square, and municipal merger dummy are included (in differenced form) as additional instruments among the conventional Arellano-Bond instruments. Table 5 provides us with four main findings. First, the lagged dependent variable is positive and highly significant at the 1% level. This finding indicates that there is persistence in the use of bank loans. Second, the house prices variable is positive and significant at the better than the 5% level in each column. Third, the coefficient estimate of the house prices variable is largest among the smallest micro firms that have less than five employees. For these firms, the long-run effect of the house prices on Bank debt is 0.161. This means that in the long-term, a hundred euro per square meter increase in regional house prices increases the use of bank debt of the smallest micro firms by about 1.61 percent- 30

31 age points. This is not a small effect, given that the mean of bank debt for this size class is 12.4%. Finally, the credit score, which is now instrumented, switches sign relative to the FD-IV estimates. It is statistically significant only for firms with less than ten employees. Once the endogeneity of this variable is better controlled, it thus seems that firms with a worse credit score rely more on bank debt. 23 Table 5: Arellano-Bond estimates for 2004-2008 (1) (2) (3) Size class <5 employees <10 employees <50 employees Dependent variable Bank debt Bank debt Bank debt Bank debt (t-1) 0.649 *** 0.656 *** 0.654 *** (0.027) (0.024) (0.024) House prices 0.057 ** 0.038 ** 0.044 *** (0.022) (0.019) (0.016) ln(1+age) -0.009-0.017-0.018 (0.013) (0.013) (0.015) ln(total assets) 0.014 0.021 0.014 (0.015) (0.016) (0.021) Credit score 0.042 ** 0.034 ** 0.017 (0.019) (0.016) (0.014) NT 39336 51799 64939 Long-run effects 0.161 ** 0.110 ** 0.128 *** (0.064) (0.056) (0.048) Arellano-Bond test 0.3769 0.5955 1.2751 [0.7063] [0.5515] [0.2023] Sargan test 16.5529 19.3476 18.5939 [0.4850] [0.3090] [0.3523] Notes: The table reports the dynamic Arellano-Bond GMM estimates on the effects of regional house prices on small firms use of bank loans over the 2004-2008 period. The results are reported for the size classes of firms having less than five, ten, and 50 workers, respectively. The dependent variable Bank debt is a ratio of bank loans to total assets at t. The independent variables are defined as follows: Bank debt (t-1) is the lagged dependent variable measured at t-1. House prices measure the average regional prices of previously owned condominiums at the municipal level in thousand euros per square meter at t. Ln(1+Age) is a natural logarithm of firm age since incorporation (plus one) in years at t. Ln(Total assets) is 23 The Arellano-Bond test for zero second-order autocorrelation in the first differenced residuals and the Sargan test for the overidentification restrictions do not reject the null in any of the specifications.

a natural logarithm of total assets in euros at t-1. Credit score measures the observed creditworthiness of firms at the interval 3-99 (divided by 100) at t-1, where lower values imply higher creditworthiness. Planned area, its square, and Municipal merger dummy taking a value of one in and after the period of merger, are included as additional instruments. All the models include year dummies. House prices and Credit score are defined as endogenous and ln(total assets) is defined as predetermined. NT is the number of firm-year observations. The long-run effects of house prices are computed as follows: _b[house prices(t)]/(1-_b[bank debt (t-1)]). The Arellano-Bond test for the second-order autocorrelation in the first differences residuals and the Sargan test for the overidentification restrictions are reported [pvalues in brackets]. The estimations use a two-step estimator with robust Windmeijercorrected standard errors. The Sargan test is based on conventional standard errors. Standard errors in parentheses: * p < 0.10, ** p < 0.05, *** p < 0.01. 4.2 Is there a collateral channel at work? 32 Our baseline results suggest that if housing values increase, small businesses rely more on bank debt. Moreover, this effect appears to be more pronounced among the smallest firms, though the evidence is somewhat mixed in this regard. In this subsection, we take a closer look at the possible reasons for these findings. To this end, we ask four questions: First, is the effect greater for firms with more tangible assets? Second, is the effect smaller after the onset of the financial crisis? Third, is there an effect on total debt? Fourth, could a wealth effect be driving our findings? As we explain below, it is more likely that there is a collateral channel at work if we answer affirmatively to the first and second questions and negatively to the third and fourth questions. 4.2.1 Is the effect greater for firms with more tangible assets? So far, we have been agnostic about whether the real estate or housing assets are owned by the small businesses or personally by their owner-entrepreneurs. If the former is more important, we should find larger effects for firms with more tangible assets. However, if the personal housing assets of the owner-

33 entrepreneurs are an important source of collateral and serve as a substitute for corporate assets, we may find smaller effects for firms with more tangible assets. To explore the empirical relevance of these channels, we drop from the sample the firm-year observations for which the ratio of tangible assets to total assets is less than 2.2%. Roughly speaking, such firms belong to the lowest quintile of the distribution of tangible assets in the baseline sample (and to the lowest quartile in the longer sample, 2004-2011). Because we need at minimum three adjacent observations per firm for our estimations, the sample restriction removes roughly one third of the firm-year observations from the baseline sample. The FD-IV results are shown in Table 6. The coefficients of the house prices are positive and significant at the 5% level or better. Compared to our baseline FD-IV estimates, the size of the coefficients increases quite a bit. We report the corresponding Arellano-Bond estimates in Table 7. The long-run coefficients are 0.152, 0.110, and 0.115 in columns (1)-(3), so that the effects sizes are comparable to what we found in the baseline Arellano-Bond estimations. The long-run coefficients are significant at the 5%, 10%, and 5% significance level, respectively. These findings support the view that housing prices are associated with the use of bank loans, but the evidence is a bit mixed as to whether the effect is greater for firms with more tangible assets. Table 6: FD-IV estimates for 2004-2008, firms with tangibles assets (1) (2) (3) Size class <5 employees <10 employees <50 employees Dependent variable Bank debt Bank debt Bank debt

34 House prices 0.207 ** 0.207 *** 0.152 ** (0.086) (0.072) (0.060) ln(1+age) -0.041 *** -0.044 *** -0.044 *** (0.012) (0.011) (0.009) ln(total assets) 0.009 *** 0.008 *** 0.008 *** (0.003) (0.003) (0.002) Credit score -0.007-0.009-0.010 * (0.008) (0.007) (0.006) NT 26541 36446 46866 Hansen s J statistics 1.838 3.018 1.106 [0.3988] [0.2211] [0.5751] Kleibergen-Paap Wald statistics 690.059 1050.240 1467.644 First-stage House prices House prices House prices Planned area -1.378 *** -1.268 *** -1.223 *** (instrument) (0.109) (0.088) (0.077) Planned area squared 0.028 *** 0.028 *** 0.028 *** (instrument) (0.001) (0.001) (0.001) Municipal merger -0.038 *** -0.036 *** -0.036 *** (instrument) (0.003) (0.002) (0.002) Notes: The table reports the first-differenced IV model (GMM) estimates on the effects of regional house prices on small firms use of bank loans over the 2004-2008 period. The subsample focuses on tangible firms defined as those above the lowest quintile in tangibility. The results are reported for the size classes of firms having less than five, ten, and 50 workers, respectively. The dependent variable Bank debt is defined as a ratio of bank loans to total assets at t. The independent variables are defined as follows: House prices measure the average prices of previously owned condominiums at the municipal level measured in thousand euros per square meter at t. The instruments include Planned area, its square, and Municipal merger dummy taking a value of one in and after the period of merger. Ln(1+Age) is a natural logarithm of firm age since incorporation (plus one) in years at t. Ln(Total assets) is a natural logarithm of total assets in euros at t-1. Credit score measures the observed creditworthiness of firms at the interval 3-99 (divided by 100) at t-1, where lower values imply higher creditworthiness. All the models include year dummies. NT is the number of firmyear observations. The first-stage results for the instruments are reported omitting the rest of the output from the table. Hansen s J statistics for overidentifying restrictions [p-values in brackets] and the Kleibergen-Paap Wald F statistics for weak identification of the first-stage regression are provided. Standard errors in parentheses: * p < 0.10, ** p < 0.05, *** p < 0.01. The standard errors are clustered at the firm level. Table 7: Arellano-Bond estimates for 2004-2008, firms with tangibles assets (1) (2) (3) Size class <5 employees <10 employees <50 employees Dependent variable Bank debt Bank debt Bank debt Bank debt (t-1) 0.621 *** 0.630 *** 0.636 *** (0.033) (0.030) (0.036) House prices 0.058 ** 0.041 * 0.042 ** (0.029) (0.024) (0.020) ln(1+age) -0.024-0.031 * -0.034 (0.017) (0.017) (0.022) ln(total assets) 0.036 0.040 0.039 (0.023) (0.026) (0.039) Credit score 0.022 0.020 0.002 (0.024) (0.020) (0.017)

35 NT 26541 36446 46866 Long-run effects 0.152 ** 0.110 * 0.115 ** (0.076) (0.065) (0.057) Arellano-Bond test -0.0899-0.2446 0.4602 [0.9284] [0.8067] [0.6454] Sargan test 26.9418 26.0473 24.8968 [0.0589] [0.0736] [0.0970] Notes: The table reports the dynamic Arellano-Bond GMM estimates on the effects of regional house prices on small firms use of bank loans over the 2004-2008 period. The subsample focuses on tangible firms defined as those above the lowest quintile in tangibility. The results are reported for the size classes of firms having less than five, ten, and 50 workers, respectively. The dependent variable Bank debt is a ratio of bank loans to total assets at t. The independent variables are defined as follows: Bank debt (t-1) is the lagged dependent variable measured at t-1. House prices measure the average regional prices of previously owned condominiums at the municipal level in thousand euros per square meter at t. Ln(1+Age) is a natural logarithm of firm age since incorporation (plus one) in years at t. Ln(Total assets) is a natural logarithm of total assets in euros at t-1. Credit score measures the observed creditworthiness of firms at the interval 3-99 (divided by 100) at t-1, where lower values imply higher creditworthiness. Planned area in the municipality at t, its square, and Municipal merger dummy taking a value of one in and after the period of merger, are included as additional instruments. All the models include year dummies. House prices and Credit score are defined as endogenous and ln(total assets) is defined as predetermined. NT is the number of firm-year observations. The long-run effects of house prices are computed as follows: _b[house prices(t)]/(1-_b[bank debt (t-1)]). The Arellano-Bond test for the secondorder autocorrelation in the first differences residuals and the Sargan test for the overidentification restrictions are reported [p-values in brackets]. The estimations use a two-step estimator with robust Windmeijer-corrected standard errors. The Sargan test is based on conventional standard errors. Standard errors in parentheses: * p < 0.10, ** p < 0.05, *** p < 0.01. 4.2.2 Is the effect smaller after the onset of the financial crisis? The financial crisis originated from the US housing markets and highlighted the risks in using real estate and housing assets as collateral. The prior work supports the view that the financial crisis has changed the lending behavior of banks (see, e.g., Ivashina and Scharfstein 2010) and that it has had an effect on the collateral channel (see, e.g., Norden and van Kampen 2013). It is thus possible that Northern-European banks have become less willing to accept housing or real estate assets as collateral for their small business lending. If they have, there may be a structural break in the relation between house prices and the use of bank loans by small firms. Such a break would be consistent with the view

36 that our results are related to a collateral channel that works through housing and real estate assets and their collateral value in particular. To explore whether the financial crisis has changed how the house prices affect small business lending, we extend the sample to cover 2004-2011. 24 To allow for a break, we add to model (1) an interaction term between the house prices and a dummy for the post-2009 period. 25 We include as additional instruments the interactions of the instruments (Planned area, and its square) with the post-crisis dummy; see Figure A2 in the Appendix on how the first-stage works. The FD-IV estimates based on the longer sample are provided in Table 8. 26 The results show that House prices still obtains a positive and significant coefficient albeit only weakly so in column (1). We also find that the interaction term between house prices and post-2009 dummy is negative and significant. This shows that the link between regional house prices and small firms use of bank loans is weaker in the post-crisis period. This finding suggests that there has been a structural change in the relation between house prices and small business lending in the aftermath of the financial crisis. The null hypothesis that the 24 Table A1 in the Appendix gives the descriptive statistics and the correlation matrix for this longer sample. 25 The crisis dummy is dropped from the model because of its collinearity with year effects. The direct effect (of the crisis) is captured by the year dummies. 26 The first-stage regressions indicate that the instruments are highly significant at the 1% level. The Kleibergen-Paap Wald F statistics vary between 1472.536 and 1820.886. These findings indicate that weak identification ought not be a concern for us.

37 coefficients of House prices and the interaction term are jointly zero is rejected at better than the 10%, 5%, and 1% level in columns (1)-(3), respectively. This finding is consistent with the view that banks have become more cautious about the valuation of housing assets in the post-crisis period and that they are therefore less willing to lend against housing-related collateral. Table 8: FD-IV estimates for 2004-2011 (1) (2) (3) Size class <5 employees <10 employees <50 employees Dependent variable Bank debt Bank debt Bank debt House prices 0.077 * 0.080 ** 0.088 *** (0.040) (0.033) (0.029) Post-crisis*house prices -0.004 ** -0.004 *** -0.005 *** (0.002) (0.002) (0.001) ln(1+age) -0.044 *** -0.045 *** -0.045 *** (0.006) (0.006) (0.005) ln(total assets) 0.007 *** 0.007 *** 0.006 *** (0.002) (0.001) (0.001) Credit score -0.019 *** -0.020 *** -0.018 *** (0.005) (0.004) (0.004) NT 93823 120211 147458 Hansen s J statistics 0.087 1.002 1.338 [0.9933] [0.8007] [0.7202] Kleibergen-Paap Wald statistics 1472.536 1569.472 1820.886 First-stage House prices House prices House prices Planned area -0.959 *** -1.079 *** -1.155 *** (instrument) (0.096) (0.080) (0.068) Planned area squared 0.027 *** 0.029 *** 0.031 *** (instrument) (0.001) (0.001) (0.001) Municipal merger -0.009 *** -0.010 *** -0.010 *** (instrument) (0.001) (0.001) (0.001) Post-crisis! planned 0.337 *** 0.333 *** 0.330 *** area (instrument) (0.007) (0.006) (0.006) Post-crisis! planned -0.006 *** -0.006 *** -0.006 *** area squared (instrument) (0.0001) (0.0001) (0.0001) First-stage Post-crisis! house prices Post-crisis! house prices Post-crisis! house prices Planned area -4.568 *** -4.542 *** -4.685 *** (instrument) (0.453) (0.390) (0.351) Planned area squared -0.045 *** -0.040 *** -0.034 *** (instrument) (0.007) (0.006) (0.006)

38 Municipal merger -0.040 *** -0.036 *** -0.032 *** (instrument) (0.004) (0.004) (0.003) Post-crisis! planned 8.121 *** 8.391 *** 8.631 *** area (instrument) (0.228) (0.237) (0.241) Post-crisis! planned -0.095 *** -0.102 *** -0.108 *** area squared (instrument) (0.005) (0.006) (0.006) Notes: The table reports the first-differenced IV model (GMM) estimates on the effects of regional house prices on the use of bank loans over the 2004-2011 period. The results are reported separately for the size classes of firms having less than five, ten, and 50 workers, respectively. The dependent variable Bank debt is a ratio of bank loans to total assets at t. The endogenous variables are defined as follows: House prices measure the average prices of previously owned condominiums at the municipal level in thousand euros per square meter at t. Post-crisis! house prices is an interaction between a dummy for the 2009-2011 period and house prices. The instruments include Planned area, Planned area squared, Municipal merger dummy, Post-crisis! planned area, and Post-crisis! planned area squared. Ln(1+Age) is a natural logarithm of firm age since incorporation (plus one) in years at t. Ln(Total assets) is a natural logarithm of total assets in euros at t-1. Credit score measures the observed creditworthiness of firms at the interval 3-99 (divided by 100) at t-1, where lower values imply higher creditworthiness. All the models include year dummies. NT is the number of firmyear observations. The first-stage results for the instruments are reported omitting the rest of the output from the table. Hansen s J statistics for overidentifying restrictions [p-values in brackets] and the Kleibergen-Paap Wald F statistics for weak identification of the first-stage regressions are provided. Standard errors in parentheses: * p < 0.10, ** p < 0.05, *** p < 0.01. The standard errors are clustered at the firm level. Table 9 reports the dynamic Arellano-Bond GMM estimates for the longer sample. These analyses include a second lag of the dependent variable, because it turned out to be necessary for us to be able eliminate serial correlation in the first-differenced residuals. The instruments are the same as before, except that the interactions of Planned area and its square with the post-crisis dummy are also used. As the table shows, the house prices continue to have a positive and significant effect on the bank loans. 27 The interaction term is negative and highly 27 The Arellano-Bond test for the second-order autocorrelation in the first-differenced residuals does not reject the null of zero autocorrelation. The Sargan test for overidentification restrictions, on the other hand, rejects the null hypothesis for this model and sample. This rejection may be driven by the assumption that the errors ought to be i.i.d.

significant throughout. The table also shows that the long-run effects are smaller when the longer sample covering the financial crisis is used. 39 Table 9: Arellano-Bond estimates for 2004-2011 (1) (2) (3) Size class <5 employees <10 employees <50 employees Dependent variable Bank debt Bank debt Bank debt Bank debt (t-1) 0.705 *** 0.706 *** 0.685 *** (0.022) (0.019) (0.018) Bank debt (t-2) 0.028 *** 0.028 *** 0.029 *** (0.009) (0.008) (0.007) House prices 0.023 *** 0.017 ** 0.022 *** (0.008) (0.007) (0.006) Post-crisis*house prices -0.005 *** -0.004 *** -0.004 *** (0.002) (0.001) (0.001) ln(1+age) -0.003 0.001-0.010 (0.010) (0.009) (0.010) ln(total assets) 0.008 0.008 0.021 (0.011) (0.011) (0.014) Credit score 0.018 0.019-0.001 (0.016) (0.014) (0.012) NT 66236 85780 106609 Long-run effects pre-crisis 0.087 *** 0.066 ** 0.078 *** (0.032) (0.028) (0.023) Long-run effects post-crisis 0.070 ** 0.052 ** 0.064 *** (0.028) (0.024) (0.020) Arellano-Bond test 0.8516 0.5042 0.3662 [0.3945] [0.6141] [0.7142] Sargan test 93.9442 94.2158 115.8255 [0.0246] [0.0236] [0.0004] Notes: The table reports the dynamic Arellano-Bond GMM estimates on the effects of regional house prices on small firms use of bank loans over the 2004-2011 period. The results are reported for the size classes of firms having less than five, ten, and 50 workers, respectively. The dependent variable Bank debt is a ratio of bank loans to total assets at t. The independent variables are defined as follows: Bank debt (t-1) and Bank debt (t-2) are the lagged dependent variables measured at t-1 and t-2, respectively. House prices measure the average regional prices of previously owned condominiums at the municipal level in thousand euros per square meter at t. Post-crisis! house prices is an interaction between house prices and a dummy for the 2009-2011 period. Ln(1+Age) is a natural logarithm of firm age since incorporation (plus one) in years at t. Ln(Total assets) is a natural logarithm of total assets in euros at t-1. Credit score measures the observed creditworthiness of firms at the interval 3-99 (divided by 100) at t-1, where lower values imply higher creditworthiness. Planned area, Planned area squared, Municipal merger, Post-crisis! planned area, and Post-crisis! planned area squared are included as additional instruments. House prices, Post-crisis! house prices and Credit score are defined as endogenous and ln(total assets) is defined as predetermined. All the models include year dummies. NT is the number of firm-year observations. The longrun effects of house prices for the pre- and post-crisis periods are computed as follows: _b[house prices(t)]/( 1-_b[Bank debt (t-1) -_b[bank debt (t-2)]) and (_b[house prices(t)]+_b[postcrisis! house prices])/(1-_b[bank debt (t-1) -_b[bank debt (t-2)]), respectively. The Arellano-Bond test for the second-order autocorrelation in the first differences residuals and the Sargan test for the overidentification restrictions are reported [p-values in brackets]. All the estimations use a two-step estimator with robust Windmeijer-corrected standard errors. The Sar-

gan test is based on conventional standard errors. Standard errors in parentheses: * p < 0.10, ** p < 0.05, *** p < 0.01. 4.2.3 Is there an effect on total debt? 40 If our main findings are driven by a collateral channel, the results should be weaker for debt items that include non-collateralizable debt. To explore whether this is the case, we use Total debt as an alternative outcome variable. It is defined as the ratio of total debt to total assets. It includes, for example, trade credit and various forms of non-bank borrowing. The mean of this variable is 0.511. 28 Tables A2 and A3 in the Appendix report the results when model (1) is estimated using the FD-IV and Arellano-Bond methods, using the main 2004-2008 sample and house prices at the municipal level. The results show that Total debt is not significantly associated with house prices. These findings indicate that it is the use of bank loans, which are typically secured, that is most affected by regional house prices. 4.2.4 Is there a wealth effect at work? Pekkala-Kerr et al. (2014) and Jensen et al. (2014) argue that if there is a positive relation between housing prices and entrepreneurship, it could be related to a wealth effect rather than to the collateral channel. They argue, for example, that (potential) entrepreneurs risk aversion may decrease when the value of their housing assets increases. The entrepreneurs willingness to expand and take 28 The values below zero and above one are set to zero and one, respectively.

41 risks could also be greater, if they have smaller mortgage commitments (e.g., Bracke, Hilber and Silva 2013). In our case, similar mechanisms could be at work, if the owner-managers of the smaller firms become less risk-averse when housing prices increase. We test the wealth effect hypothesis by building on the recent work of Bliss et al. (2014) on the effects of financial constraints on the dividend payout ratios. They argue and provide evidence that in the presence of financial constraints, firms resort to dividend payout reductions as a substitute source of capital. We argue that in the presence of a genuine collateral channel effect, the owner-managers of small firms do not increase the dividend payout when housing prices increase. The logic of this argument is that if a small business is financially constrained and does not therefore pay (large) dividends to start with, enhanced availability of external finance ought to increase the dividend payout if and only if the firm s (marginal) investments are no longer financially constrained after the additional external finance becomes available. It seems plausible to argue that if there are financial constraints that (still) bind, the firm would merely increase its borrowing against the unleashed new housing collateral after increases in the value of housing assets and would not increase the dividend payout. 29 In contrast, in the case of a wealth effect, we could expect firms to increase the dividend payout ratio, or to adopt a policy to pay a posi- 29 Our argument builds on the internal and external financing being substitutes, quite like in the standard pecking-order theory. One could also argue that in some cases, they are complementary; see, e.g., Almeida and Campello (2010).

42 tive amount of dividends. One motivation for such a change in the payout policy is that it would allow the entrepreneurs to finance a wealth-induced increase in the desired consumption. This mechanism would be consistent with a standard wealth effect on consumption (e.g., Campbell and Cocco 2007). To implement these tests, we use two alternative dividend-based measures. The first of them is a binary indicator for positive dividends (and is zero, if no dividends are paid). 30 The second measure is the ratio of dividends to total assets, measured at period t (and winsorized at the 99 th percentile). We then replicate the regressions reported in Tables 3-5, using the two dividend measures as the dependent variable. Overall, we do not find a significant, positive relation between the dividend payout and house prices based on either of the two measures. 31 If anything, the estimates seem to indicate a weak negative relation instead. We 30 The dividend payments are proxied for each firm by the difference in the retained earnings between t-1 and t. This proxy can be obtained for a smaller sample, covering close to 60% of the baseline sample. To check that non-random selection into this subsample does not drive our wealth effect findings, we repeat the estimations by replacing the dividend observations that are missing at t with observations from t-1, and if not available, from t+1. This adjusted sample is more comparable to the baseline sample and when we use it, the results for the wealth effects are similar to what we obtain using the smaller sample. 31 There are some econometric challenges when we implement these estimations. For example, the Arellano-Bond specification tests suggest that in some of the specifications, there are traces of second-order autocorrelation in the first-differenced residuals. A degree of caution is therefore warranted when interpreting these findings.

43 acknowledge that this evidence is indirect, but together with our other findings, it nevertheless casts doubt on the view that a wealth effect is driving our findings. 4.2.5 Summary This subsection has addressed four questions: First, is the effect greater for firms with more tangible assets? Second, is the effect smaller after the onset of the financial crisis? Third, is there an effect on total debt? Fourth, is there a wealth effect at work? We answer weakly affirmatively to the first question and more affirmatively to the second question, and negatively to the third and fourth questions. We argue that taken together, these findings suggest that there is a collateral channel at work. 4.3 Additional robustness tests In this subsection, we explore the robustness of our baseline findings in a number of ways. 4.3.1 Definition of the dependent variable In our baseline sample, the mean of Bank debt is 13.3%, and its median, when calculated over all firm-year observations, is zero. These numbers raise the question of how commonly small firms borrow from banks and, especially, whether our results are sensitive to the way how we define our dependent variable.

44 To check how prevalent the use of bank debt is, we use a binary indicator that is equal to one if Bank debt > 0 and that is zero otherwise. The mean of this indicator in our baseline estimation sample is 43%. Moreover, it turns out that about 55% of the firms in our baseline sample resort to bank borrowing at least once during our sample period (i.e., for them, Bank debt > 0 at least during one year). These patterns suggest that access to bank debt is relevant for the majority of the small firms in the sample. To check how sensitive our results are to the definition of the dependent variable, we use a binary indicator, 1(Bank debt > 0), that is equal to one if Bank debt > 0 and that is zero otherwise as the dependent variable (in place of Bank debt). When we repeat the regressions of Table 3, 4 and 5, we find that in the FE estimations, the coefficients of house prices are all positive and significant at the better than the 10% level in specifications (1)-(2), and positive and insignificant in specification (3). In the Arellano-Bond GMM estimations, the coefficients are still positive and significant. 32 In the FD-IV regressions, the coefficients are however imprecisely measured and turn out to be insignificant. 33 32 The Arellano-Bond test, however, rejects the specification. 33 Given that Bank debt has a probability mass at zero, estimating a Tobit model is an obvious alternative modeling choice. Using an IV-Tobit model with pooled data and the same instruments as in our baseline analysis results in a negative and significant coefficient for the house prices. This finding is consistent with the raw correlations in the data. In light of our analysis, the finding is expected, because the IV-Tobit analysis does not allow us to control for the unobserved permanent firm-specific heterogeneity.

45 It is not entirely unexpected that when we use a coarser measure of bank borrowing, our estimates become less accurate and stable. The reason is that we have firm fixed effects in each model: If there is a lot of persistence at the extensive margin of the leverage decision (as the prior literature suggests; see, e.g., Strebulaev and Yang 2013), a large part of the within-firm variation may be related to changes in the level of bank borrowing at the intensive margin. Such variation is swiped out when the binary indicator, 1(Bank debt > 0), is used as the dependent variable. 4.3.2 Measurement of housing prices We have re-estimated all the specifications using an alternative, more aggregated measure for house prices, based on two-digit zip code level house prices. We construct the alternative measure as a weighted average of the five-digit zip code level data, using the number of transactions as the weight. This measure divides the country into 99 regions (instead of the 226 municipalities observed in the sample). Overall, the results are somewhat stronger (i.e., house price coefficients are larger and the statistical significance improves in some specifications) when this more aggregated measure is used. One potential explanation for the finding is that the measure implicitly allows the relevant housing assets to be located in a wider regional area (i.e., further from the firm). As a complementary robustness test, we measure house prices at the fivedigit zip code level. The use of this more disaggregated measure assumes that the relevant real estate and housing assets are located in the five-digit zip code area where the firm is registered. This would be a reasonable assumption if, for

46 example, the firms typically operate at the entrepreneurs home addresses and if those assets matter for small business lending. For completeness, we use the five-digit prices in three ways: First, we consider the baseline 2004-2008 sample. When the five-digit zip code level and this shorter sample are used, the coefficient of the house prices remains positive. Unlike in the baseline estimations, the FE estimates are all significant at the better than the 5% level. The FD-IV estimates remain significant and are thus in line with the baseline results. However, the effects are statistically less precisely measured, especially for the smallest firms, in the Arellano-Bond specifications. The second way we use the five-digit prices is that we consider the longer sample that cover 2004-2011 (and thus include the post crisis dummy, interacted with the house prices). We report the FD-IV and Arellano-Bond GMM results for this longer sample in Table A4 and A5 of the Appendix. The FD-IV estimates suggest that the coefficient of the house prices is positive and significant at 10%, 5% and 1% level in columns (1)-(3), respectively. The Arellano-Bond GMM estimates largely echo these findings and are significant at better than the 5% level in each specification. Third, we have also rerun the models that exclude the firms with few tangible assets using the more disaggregated, five-digit zip code -level data. Again, the results are largely in line with our baseline findings. We found large and highly significant estimates in the FD-IV specifications. However, the house price coefficients are smaller and statistically less precisely measured in the Arellano-Bond specifications, when compared to the corresponding results that we obtained using the municipal-level prices.

47 In sum, the results from these additional analyses suggest that using the more aggregated (two-digit) and less aggregated (five-digit) house prices largely confirm our baseline findings. Perhaps the most important qualification to this is that measuring the housing prices at the five-digit zip code level is not the most appropriate measurement level, especially when analyzing the smallest firms for which the private housing assets are likely to be the most relevant. It hence seems the real estate and housing prices nearby, but not solely at firms office address, are more important for the smallest firms. Consistent with this view, the somewhat more robust estimates for the larger small businesses suggest that the five-digit zip code level prices may also capture a collateral channel effect that is related to the other real estate assets (say, office buildings) of the firms, located in the office addresses or very nearby. 4.3.3 Clustering of standard errors So far, we have clustered the standard errors at the firm level to control for the autocorrelated firm-specific shocks related to the potential persistence of leverage choices and use of bank loans. A commonly applied approach in the literature is to cluster the standard errors at the level of the regressor that is of key interest (see, e.g., Angrist and Pischke 2009; Cameron and Miller 2015). This approach allows us to adjust standard errors for any potential regional shocks that might affect the firms located in the same regional area. In our case, the use of regional clustering is complicated by the unbalanced sizes of clusters that

48 would violate the assumption of equal cluster sizes. 34 Nevertheless, we report a set of robustness tests utilizing regional-level clustering in Tables A6-A8 of the Appendix. Like in our baseline analyses, we analyze housing prices measured at the municipal level and focus on the firms smaller than 50 employees over the 2004-2008 period. We also report for comparison the results from the estimations that use the five-digit zip code level housing prices. We apply the municipal- and zip code level clustering to both of them, acknowledging the unbalanced nature of cluster sizes at the municipal level. 35 The FE estimates in Table A6 are positive and weakly significant when we measure housing prices at the municipal level and use either municipal- or zip code level clustering. On the other hand, the estimates are highly significant (t the 1% level) when we use the disaggregated (five-digit) zip code level housing prices, regardless of the level of regional clustering. The FD-IV estimates are reported in Table A7. They show that the significance of the housing price coefficients vary somewhat with the level of cluster- 34 The cluster sizes vary between 0% and 20% of the observations in the sample in the case of municipal-level clustering (226 clusters), warranting caution due to their unbalanced nature. The cluster sizes vary between 0% and 3% of the sample in the case of zip code level clustering (834 clusters). This latter cluster choice hence results in more balanced clusters. 35 In this robustness test, we utilize 2SLS and one-step GMM estimation in the FD-IV and Arellano-Bond specifications, respectively, to provide a better comparability between the specifications. This approach avoids the small differences observed in the coefficient estimates when using different levels of regional clustering in the two-step GMM estimations.

49 ing. They are, however, significant at better than the 10% level in all cases. Moreover, the first-stage estimates suggest that the instruments are less precisely measured when the municipal-level clustering is used, whereas the zip code level clustering provides more precise first-stage estimates. This is consistent with the latter being less affected by the problem of few effective clusters. It is comforting to report that the Kleibergen-Paap Wald statistics greatly exceed the critical values in each specification, alleviating thus the concern of our findings being driven by weak instruments. The Arellano-Bond estimates are shown in Table A8. They show that the short- and long-run housing price estimates are significant at the 1% level regardless of the level of clustering. In sum, our main qualitative conclusions do not depend on the chosen level of clustering. 4.3.4 Nonlinearity in credit score We can also check the robustness of the results for potential nonlinearity in the relationship between the bank debt and credit score measures. To do so, we include both squared and cubed terms of the credit score in addition to the linear effect into the FE, FD-IV, and Arellano-Bond models. The results remain very similar to the baseline estimates (shown in Tables 3-5), with the Arellano-Bond estimates suggesting somewhat larger long-run effects. The additional polynomial terms are not individually significant. This insignificance is not surprising, because the correlation between the credit score and its square and cube is high, around 0.94 and 0.83, respectively.

50 4.3.5 Investment opportunities As a further robustness test, we explore the sensitivity of our results to using realized sales growth and municipal-level controls as alternative proxies for local investment opportunities. The concern that motivates this analysis is that our instruments (Planned area and Municipal merger) may mirror such opportunities, despite us conditioning on the fixed effects and observable characteristics. If they do, they might be correlated with the use of bank debt, on and above their effect on the housing prices. Such correlation would mean that the instruments do not satisfy the exclusion restriction. To implement this robustness test, we proceed as follows: First, we reestimate the FE, FD-IV and Arellano-Bond models of Tables 3-5 using the realized growth of sales (in logs) from t-2 to t-1 as an additional control variable. We utilize pre-2004 data when constructing the new variable in order to avoid losing the first cross-section of our baseline sample. However, the need for the additional data means that we lose some 4000-6000 observations in the FD-IV and Arellano-Bond estimations compared to the baseline estimation sample. 36 Overall, the results remain intact when the realized sales growth is controlled 36 Three additional details are worth mentioning: First, to remove outliers, we winsorize the realized sales growth variable at the 1 st and 99 th percentiles. Second, we treat the sales growth as a pre-determined variable in the Arellano-Bond models. Third, the measurement period of the additional regional controls is dominated by the data availability. While we aim to use lagged data on regional variables, we allow some of them to be measured at the end of period t or at the beginning of t+1.

51 for. The coefficient of house prices remains positive in each specification and the size of coefficients remains similar to the baseline estimates. The results are significant at the better than the 5% level in each specification, with the exception of one FE specification (column (2)) and one FD-IV specification (column (1)), where the coefficients are weakly significant and insignificant, respectively. Second, we control for the local regional demand shocks using the municipal-level regional control variables. We use the median of taxable income (including tax-free dividends and interest) at t, the share of start-up firms of the firm population at t-1, and the unemployment rate at the beginning of the next period (i.e., end of January). The sample sizes decrease by a couple of thousand observations in these specifications compared to the baseline estimations reported in Tables 3-5. However, our main results are robust to the inclusion of these municipal-level controls. If anything, the size of the house price coefficients increases and the statistical significance also improves. In the FE model, the coefficients of house prices are significant at the better than the 10% level and vary in range [0.013, 0.018]. In the FD-IV model, the estimates are significant at the 10%, 1% and 5% level in columns (1)-(3), respectively, and vary in range [0.119, 0.155]. In the Arellano-Bond estimations, the coefficients of house prices are significant at the 1% level and vary in range [0.049, 0.068]. These estimates suggest long-run coefficients in range [0.137, 0.187]. These long-run coefficients are somewhat larger than those implied by the baseline estimates.

52 4.4 Real effects Finally, we briefly analyze whether higher housing prices enhance the performance of firms. They should, if the marginal project, which can be financed when the housing prices increase, is of good quality. We analyze the effects of regional house prices using three alternative measures: net investments at t, scaled by tangible assets at t-1, the natural logarithm of the number of employees at t, and labor productivity, measured as the natural logarithm of (nominal) value added divided by employment at t. We estimate the FE, FD-IV, and Arellano-Bond models using these measures as dependent variable and the same controls as previously. In the case of labor productivity, we supplement the controls with a crude measure for capital intensity, defined as the natural logarithm of tangible assets divided by employment measured at t-1. 37 We would like to remind the reader that unlike the previous literature, we control for the creditworthiness of the firms using the commercial credit score in each of these regressions. 37 Several further details on the implementation of these estimations are worth mentioning: First, we winsorize the investment measure at the 1 st and 95 th percentiles and the labor productivity measure at the 1 st and 99 th percentiles. Second, we assume that the capital intensity is a pre-determined variable in the Arellano-Bond specifications. Third, the firms with zero tangible assets at period t-1 are removed from the sample when the capital intensity is included. Finally, the Arellano-Bond specification tests suggest that some caution is warranted in some of these specifications due to traces of second-order autocorrelation in the first-differenced residuals.

53 We do not find a statistically significant relationship between contemporaneous investments and housing prices, while the relationship between employment and housing prices is (weakly) significant only in the FD-IV model. Interestingly, there is some (admittedly weak) evidence of a negative relationship between the labor productivity and housing prices. Furthermore, this negative productivity effect is more pronounced among the smallest micro firms, for whom the coefficient is largest in absolute terms and highly significant in the Arellano-Bond models (and weakly significant in the FE and FD-IV models). Taken together, these findings suggest that the marginal borrowers, who benefit from increasing housing values in the form of increased collateral values, tend to perform no better than the others. This result is in line with what Pekkala-Kerr et al. (2014) report for the US entrepreneurs and with the results of Andersen and Meisner Nielsen (2012) and Jensen et al. (2014) for Danish entrepreneurs. All these papers find that when financial constraints are relaxed, the quality of the marginal entrepreneurial projects that can, as a result, be implemented, is low. Our finding is also consistent with the predictions of Manove, Padilla and Pagano (2001), who show theoretically that the use of collateral may reduce banks screening effort and lead them to finance lower quality projects. 5 CONCLUSIONS This paper has analyzed the importance of collateral channel for the borrowing of privately held small businesses using a large firm-level panel dataset from Finland over 2004-2011. We have focused on estimating the effect of regional

54 house prices on the use of bank loans by such small businesses. We identify these effects by exploiting bureaucratically- and politically-driven differences in municipal zoning and municipal mergers. The main result of this paper is that if housing prices increase, it facilitates the use of bank loans by the privately held small businesses located in the region. Our baseline (dynamic) estimations for the smallest micro firms suggest that in the long-term, a hundred euro per square meter increase in regional house prices increases the use of bank debt by about 1.61 percentage points. This is not a small effect, considering that the mean of bank debt is a bit more than ten percent. This positive effect is consistent with the economic theories, which postulate that collateral values are import for the borrowing capacity of opaque small businesses. Five further findings of ours allow us to argue that the effect that we document is due to a collateral channel: i) The effect appears to be a bit more pronounced among the smallest micro firms; ii) The effect of the housing prices is not smaller for firms with more tangible assets; iii) There is an effect of housing prices on the use of bank debt, but not on the total use of debt; iv) The effect of housing prices on the use of bank loans is somewhat weaker after the onset of the financial crisis; and v) There is no detectable effect on the dividend payout. These findings support the existence of a collateral channel, because the smallest firms and/or the firms with tangible assets are indeed the ones for whom housing assets and housing price increases should be most relevant. More importantly, if there is a collateral channel at work, the effect on the total debt should be weaker, because such a measure of total debt contains non-

55 collateralizable and/or non-bank debt items. We can also argue that if there is a collateral channel at work, it may well have become weaker after the onset of the financial crisis, as the prior work also suggests (e.g., Ivashina and Scharfstein 2010 and Norden and van Kampen 2013). Finally, if our findings were driven by a wealth effect, we should also expect to see a wealth effect on consumption that ought to, at least in part, be financed by increased dividend payouts. There is cumulating evidence that shocks to the prices of housing assets and real estate owned by firms (and their owners) change the firms access to bank credit and thereby their opportunities to invest and expand. It is unclear whether this dependence is a piece of good or bad news: It is good news, if it enhances resource allocation in the business sector. It is bad news, if it leads to housing-boom (and bust) driven borrowing and investment cycles where the marginal quality of projects that get implemented vary in an undesirable fashion. These questions are clearly policy relevant and warrant further research, as highly pro-cyclical corporate lending is a potential source of misallocation that may erode the prospects for long-term economic growth and increase the risk of credit-driven financial crises.

56 REFERENCES Adelino, M., Schoar, A. and Severino, F. 2014. House prices, collateral and selfemployment. Journal of Financial Economics, forthcoming. Almeida, H., Campello, M. 2010. Financing frictions and the substitution between internal and external funds. Journal of Financial and Qualitative Analysis 45, 589-622. Andersen, S. and Meisner Nielsen, K. 2012. Ability or finances as constraints on entrepreneurship? Evidence from survival rates in a natural experiment. Review of Financial Studies 25 (12), 3684-3710. Angrist, J.D. and Pischke, J.-S. 2009. Mostly harmless econometrics: An empiricist s companion. Princeton University Press. Balasubramanyan, L. and Coulson, E. 2013. Do house prices impact business starts? Journal of Housing Economics 22, 36-44. Benmelech, E., Garmaise, M.J. and Moskowitz, T.J. 2005. Do liquidation values affect financial contracts? Evidence from commercial loan contracts. Quarterly Journal of Economics 120 (3), 1121-1154. Benmelech, E. and Bergman, N.K. 2009. Collateral pricing. Journal of Financial Economics 91, 339-360. Berger, A.N. and Udell, G.F. 1998. The economics of small business finance: The roles of private equity and debt markets in the financial growth cycle. Journal of Banking and Finance 22, 613-673. Bernanke, B. and Gertler, M. 1989. Agency costs, net worth, and business fluctuations. American Economic Review 79 (1), 14-31.

57 Bester, H. 1985. Screening vs. rationing in credit markets with imperfect information. American Economic Review 75 (4), 850-855. Black, J., de Meza, D. and Jeffreys, D. 1996. House prices, the supply of collateral and the enterprise economy. Economic Journal 106 (434), 50-75. Bliss, B.A., Cheng, Y. and Denis, D.J. 2014. Corporate payout, cash retention, and the supply of credit: Evidence from the 2008-2009 credit crisis. Journal of Financial Economics, forthcoming. Boot, A.W., Thakor, A.V. and Udell, G.F. 1991. Secured lending and default risk: Equilibrium analysis, policy implications and empirical results. The Economic Journal 101 (406) 458-472. Bracke, P., Hilber, C. and Silva O. 2013. Homeownership and entrepreneurship: The role of commitment and mortgage debt. IZA Discussion Paper No. 7417. Cameron, A.C. and Miller, D.L. 2015. Practioner s guide to cluster-robust inference. Journal of Human Resources, forthcoming. Campbell, J.Y. and Cocco, J. 2007. How do house prices affect consumption? Evidence from micro data. Journal of Monetary Economics 54 (3), 591-621. Chaney, T., Sraer, D. and Thesmar, D. 2012. The collateral channel: How real estate shocks affect corporate investment. American Economic Review 102 (6), 2381-2409. Corradin, S. and Popov, A. 2015. House prices, home equity, and entrepreneurship. Review of Financial Studies, forthcoming. Cvijanovic, D. 2014. Real estate prices and firm capital structure. Review of Financial Studies, forthcoming.

58 Fracassi, C., Garmaise, M.J., Kogan, S., and Natividad, G. 2014. Business microloans for U.S. subprime borrowers. Journal of Financial and Quantitative Analysis, forthcoming. Gan, J. 2007. Collateral, debt capacity, and corporate investment: Evidence from a natural experiment. Journal of Financial Economics 85, 709-734. Gertler, M. and Gilchrist, S. 1994. Monetary policy, business cycles, and the behavior of small manufacturing firms. Quarterly Journal of Economics 109 (2), 309-340. Giambona, E., Golec, J. and Schwienbacher, A. 2014. Debt capacity of real estate collateral. Real Estate Economics 42 (3), 578 605. Hurst, E. and Lusardi., A. 2004. Liquidity constraints, household wealth, and entrepreneurship. Journal of Political Economy 112 (2), 319-347. Hyytinen, A. Kuosa, I. and Takalo T. 2003. Law or finance? Evidence from Finland. European Journal of Law and Economics 16, 59-89 Hyytinen, A., Saarimaa T. and Tukiainen, J. 2014. Electoral vulnerability and size of local governments: Evidence from voting on municipal mergers. Journal of Public Economics, forthcoming. Ivashina, V. and Scharfstein, D. 2010. Bank lending during the financial crisis of 2008. Journal of Financial Economics 97 (3), 319-338. Jensen, T.L., Leth-Petersen, S. and Nanda, R. 2014. Housing collateral, credit constraints and entrepreneurship Evidence from a mortgage reform. Working paper 15-050, Harvard Business School.

59 Kashyap, A.K., Stein, J.C. and Wilcox, D.W. 1993. Monetary policy and credit conditions: Evidence from the composition of external finance. American Economic Review 83 (1), 78-98. Kiyotaki, N. and Moore, J. 1997. Credit cycles. Journal of Political Economy 105 (21), 211-248. Kleiner, K. 2013. How did the 2007-09 credit crisis become the Great Recession? Budget constraints, collateral, and small firm employment. Manuscript. Korkeamäki, T., Rainio, E. and Takalo T., 2013. Reforming corporate law in an emerging market: The case of Finland in the 1970s. Economics of Transition 21 (3), 509-551. Lin, L. 2014. Collateral and the choice between bank debt and public debt. Working paper. MacKinnon, J.G. and Webb, M.D. 2014. Wild bootstrap inference for wildly different cluster sizes. Manuscript. Manove, M., Padilla, A.J. and Pagano, M. 2001. Collateral versus project screening: A model of lazy banks. The RAND Journal of Economics 32 (4), 726-744. Norden, L. and van Kampen, S. 2013. Corporate leverage and the collateral channel. Journal of Banking and Finance 37 (12), 5062-5072. Pekkala-Kerr, S., Kerr, W.R. and Nanda, R. 2014. House money and entrepreneurship. Manuscript. Rampini, A.A. and Viswanathan, S. 2010. Collateral, risk management, and the distribution of debt capacity. Journal of Finance 65 (6), 2293-2322.

60 Rinkinen, K. 2007. Asemakaavoituksen kestoon voi vaikuttaa. Suomen Kuntaliitto (The Association of Finnish Local and Regional Authorities). Robb, A.M. and Robinson, D.T. 2014. The capital structure decisions of new firms. Review of Financial Studies 27 (1), 153-179. Saiz, A. 2010. The geographic determinants of housing supply. Quarterly Journal of Economics, 1253-1296. Schmalz, M.C., Sraer, D.A. and Thesmar, D. 2014. Housing collateral and entrepreneurship. Manuscript. SME-barometer 2004. (In Finnish: Pk-yritysbarometri, kevät 2004) of Finnvera Oyj and Suomen Yrittäjät ry. Stiglitz, J. and Weiss, A. 1981. Credit rationing in the markets with imperfect information. American Economic Review 71, 393-410. Strebulaev, I.A. and Yang, B. 2013. The mystery of zero-leverage firms. Journal of Financial Economics 109 (1), 1-23. Townsend, R.M. 1979. Optimal contracts and competitive markets with costly state verification. Journal of Economic Theory 21, 265-393.

61 APPENDIX In this appendix, we report a set of auxiliary results and analyses. Figure A1 shows the estimated relation between house prices and planned area over the 2004-2008 period. Figure A2 displays the same for the 2004-2011 period. Table A1 displays descriptive statistics and correlation matrix for 2004-2011. Table A2 reports the FD-IV estimates for 2004-2008, using Total debt as the dependent variable. Table A3 reports the same, but using Arellano-Bond GMM. Table A4 reports the FD-IV estimates for 2004-2011, using zip code level house prices. Table A5 reports the Arellano-Bond GMM estimates for 2004-2011, using the zip code level house prices. Finally, Tables A6-A8 report various alternative estimates using the regional clustering of standard errors and municipal- or zip code level housing prices focusing on the main sample period 2004-2008. Figure A1 displays the model fit from a fixed effects (within) specification where the house prices are regressed on the linear and squared Planned area, Municipal merger and time dummies. The fixed effects control, for instance, for permanent differences between urban and rural regions. The figure shows a u- shaped pattern in the relation between house prices and planned area: After a cut-off point, the relation switches sign and becomes positive. 40 We stress that 40 Two additional observations are in order: First, the municipalities located on the positively sloped part of the curve consist of municipalities located in or close to more urban regional hubs. Half of these municipalities are located in the Uusimaa region near Helsinki, which is Finland s capital. Second, there are eight municipalities in which the share of the

62 having this kind of nonlinearity is not required for the identification of the effect of house prices on small business borrowing. Indeed, Figure A2 shows that in the longer sample that also includes the years of the financial crisis, there is a negative relation between regional house prices and planned area once the permanent regional heterogeneity is controlled for. The results shown in Tables A1-A8 extend our baseline analysis, reported in the main text. They illustrate the mechanism at work and the robustness of our baseline findings. planned area is greater than 30% (Imatra, Järvenpää, Kauniainen, Kerava, Lahti, Raisio, Turku, and Vantaa). One of these municipalities (Kauniainen) is a clear outlier: Its share of planned area is 99.3%.

63 House prices, euros per sq meter 1600 1800 2000 2200 House prices and zoning 0 10 20 30 40 50 Share of planned area, % Figure A1: House prices and planned area over the 2004-2008 period Notes: The figure shows the predicted relationship between regional house prices and the share of planned area in municipalities over the 2004-2008 period. The plot is based on the fixed effects (within) regression of municipal-level House prices on Planned area and its square, Municipal merger dummy, and time dummies. The standard errors are adjusted for firm-level clustering.

64 House prices, euros per sq meter 1700 1800 1900 2000 2100 House prices and zoning 0 10 20 30 40 50 Share of planned area, % Figure A2: House prices and planned area over the 2004-2011 period Notes: The figure shows the predicted relationship between regional house prices and the share of planned area in municipalities over the 2004-2011 period. The plot is based on the fixed effects (within) regression of municipal-level House prices on Planned area, Planned area squared, Municipal merger dummy, Post-crisis! planned area, Post-crisis! planned area squared (where Post-crisis is a dummy for the post-2009 period), and time dummies. The standard errors are adjusted for firm-level clustering.

65 Table A1: Descriptive statistics and correlation matrix for 2004-2011 Panel A: Descriptive statistics variable mean sd min p50 max NT Bank debt 0.137 0.230 0.000 0.000 1.000 188307 ln(1+age) 2.550 0.697 0.693 2.708 4.754 188307 ln(total assets) 12.122 1.595 5.011 12.084 21.529 188307 Credit score 0.263 0.190 0.030 0.230 0.990 188307 House prices* 2.142 1.130 0.431 1.846 6.646 188307 House prices** 2.028 0.896 0.451 1.750 3.973 188307 Planned area 0.166 0.151 0.000 0.102 0.993 188307 Planned area squared 5.041 8.080 0.000 1.040 98.605 188307 Municipal merger 0.183 0.387 0.000 0.000 1.000 188307 Notes: The panel reports descriptive statistics for the sample of firms smaller than 50 employees over 2004-2011. The statistics include the mean, standard deviation, minimum, median, and maximum, respectively. Bank debt is a ratio of loans from financial institutions to total assets. Ln(1+Age) is a natural logarithm of firm age since incorporation (plus one) in years. Ln(Total assets) is a natural logarithm of the total assets in euros at t-1. Credit score measures the observed creditworthiness of firms at the interval 3-99 (scaled by dividing by 100), where lower values imply higher creditworthiness. House prices measure the average regional prices of previously owned condominiums at the zip code (*) and municipal (**) level in thousand euros per square meter, respectively. Planned area measures the share of planned area of the total area in the municipality. Planned area squared is defined as (Planned area*100)^2/100. Municipal merger is a dummy taking a value equal to one in and all the periods after the municipal merger. NT is the number of firm-year observations. Panel B: Correlation matrix for 2004-2011 Bank debt ln(1+age) ln(total assets) Credit score House prices Planned area Planned area squared Municipal merger Bank debt 1.0000 ln(1+age) -0.0896* 1.0000 ln(total assets) 0.0619* 0.2108* 1.0000 Credit score 0.2816* -0.2849* -0.3176* 1.0000 House prices -0.1588* -0.0137* -0.0525* 0.0153* 1.0000 Planned area -0.1319* -0.0006-0.0255* 0.0023 0.5767* 1.0000 Planned area squared -0.0897* 0.0033-0.0149* -0.0010 0.3541* 0.9083* 1.0000 Municipal merger 0.0667* -0.0072* 0.0128* 0.0033-0.2763* -0.3909* -0.2810* 1.0000 Notes: The table reports pairwise correlations for 2004-2011. Bank debt is a ratio of loans from financial institutions to total assets. Ln(1+Age) is a natural logarithm of firm age since incorporation (plus one) in years. Ln(Total assets) is a natural logarithm of the total assets in euros at t-1. Credit score measures the observed creditworthiness of firms at the interval 3-99 (scaled by dividing by 100), where lower values imply higher creditworthiness. House prices measure the average regional prices of previously owned condominiums at the municipal level in thousand euros per square meter. Planned area measures the share of planned area of the total area in the municipality. Municipal merger is a dummy taking a value equal to one in and all the periods after the municipal merger. The correlations significant at the 5 percent level or better are marked with an asterisk (*).

66 Table A2: FD-IV estimates for 2004-2008, Total debt as the dependent variable (1) (2) (3) Size class <5 employees <10 employees <50 employees Dependent variable Total debt Total debt Total debt House prices 0.003 0.002 0.003 (0.080) (0.066) (0.054) ln(1+age) -0.016-0.022 ** -0.029 *** (0.011) (0.010) (0.008) ln(total assets) -0.031 *** -0.032 *** -0.031 *** (0.003) (0.003) (0.002) Credit score -0.011-0.011 * -0.009 (0.008) (0.006) (0.005) NT 39336 51799 64939 Hansen s J statistics 2.899 4.800 3.100 [0.2347] [0.0907] [0.2123] Kleibergen-Paap Wald statistics 926.916 1378.302 1922.062 First-stage House prices House prices House prices Planned area -1.725 *** -1.589 *** -1.546 *** (instrument) (0.092) (0.076) (0.067) Planned area squared 0.031 *** 0.030 *** 0.030 *** (instrument) (0.001) (0.001) (0.001) Municipal merger -0.044 *** -0.042 *** -0.043 *** (instrument) (0.002) (0.002) (0.002) Notes: The table reports the first-differenced IV model (GMM) estimates on the effects of regional house prices on the use of total debt over the 2004-2008 period. The results are reported separately for the size classes of firms having less than five, ten, and 50 workers, respectively. The dependent variable Total debt is a ratio of total debt to total assets at t. The endogenous variables are defined as follows: House prices measure the average prices of previously owned condominiums at the municipal level in thousand euros per square meter at t. The instruments include Planned area, Planned area squared, and Municipal merger dummy. Ln(1+Age) is a natural logarithm of firm age since incorporation (plus one) in years at t. Ln(Total assets) is a natural logarithm of total assets in euros at t-1. Credit score measures the observed creditworthiness of firms at the interval 3-99 (divided by 100) at t-1, where lower values imply higher creditworthiness. All the models include year dummies. NT is the number of firm-year observations. The first-stage results for the instruments are reported omitting the rest of the output from the table. Hansen s J statistics for overidentifying restrictions [p-values in brackets] and the Kleibergen-Paap Wald F statistics for weak identification of the first-stage regressions are provided. Standard errors in parentheses: * p < 0.10, ** p < 0.05, *** p < 0.01. The standard errors are clustered at the firm level. Table A3: Arellano-Bond estimates for 2004-2008, Total debt as the dependent variable (1) (2) (3) Size class <5 employees <10 employees <50 employees Dependent variable Total debt Total debt Total debt Total debt (t-1) 0.740 *** 0.747 *** 0.763 *** (0.037) (0.034) (0.041) House prices -0.009-0.008 0.017 (0.030) (0.025) (0.022) ln(1+age) 0.066 *** 0.064 *** 0.091 *** (0.021) (0.020) (0.022)

67 ln(total assets) -0.080 *** -0.083 *** -0.128 *** (0.027) (0.026) (0.032) Credit score 0.054 *** 0.047 *** 0.047 *** (0.020) (0.017) (0.015) NT 39336 51799 64939 Long-run effects -0.034-0.033 0.072 (0.114) (0.098) (0.093) Arellano-Bond test 1.9693 1.9314 1.9228 [0.0489] [0.0534] [0.0545] Sargan test 37.1913 35.9066 30.7109 [0.0032] [0.0047] [0.0217] Notes: The table reports the dynamic Arellano-Bond GMM estimates on the effects of regional house prices on small firms use of total debt over the 2004-2008 period. The results are reported for the size classes of firms having less than five, ten, and 50 workers, respectively. The dependent variable Total debt is a ratio of total debt to total assets at t. The independent variables are defined as follows: Total debt (t-1) is the lagged dependent variable measured at t-1. House prices measure the average regional prices of previously owned condominiums at the municipal level in thousand euros per square meter at t. Ln(1+Age) is a natural logarithm of firm age since incorporation (plus one) in years at t. Ln(Total assets) is a natural logarithm of total assets in euros at t-1. Credit score measures the observed creditworthiness of firms at the interval 3-99 (divided by 100) at t-1, where lower values imply higher creditworthiness. Planned area, Planned area squared, and Municipal merger are included as additional instruments. House prices and Credit score are defined as endogenous and ln(total assets) is defined as predetermined. All the models include year dummies. NT is the number of firm-year observations. The long-run effects of house prices are computed as follows: _b[house prices(t)]/( 1-_b[Total debt (t-1)]). The Arellano-Bond test for the secondorder autocorrelation in the first differences residuals and the Sargan test for the overidentification restrictions are reported [p-values in brackets]. All the estimations use a two-step estimator with robust Windmeijer-corrected standard errors. The Sargan test is based on conventional standard errors. Standard errors in parentheses: * p < 0.10, ** p < 0.05, *** p < 0.01. Table A4: FD-IV estimates for 2004-2011 (zip code level house prices) (1) (2) (3) Size class <5 employees <10 employees <50 employees Dependent variable Bank debt Bank debt Bank debt House prices 0.070 * 0.067 ** 0.069 *** (0.037) (0.029) (0.025) Post-crisis*house prices -0.003 ** -0.003 ** -0.004 *** (0.002) (0.001) (0.001) ln(1+age) -0.045 *** -0.045 *** -0.046 *** (0.006) (0.006) (0.005) ln(total assets) 0.007 *** 0.007 *** 0.006 *** (0.002) (0.001) (0.001) Credit score -0.019 *** -0.020 *** -0.018 *** (0.005) (0.004) (0.004) NT 93823 120211 147458 Hansen s J statistics 0.322 1.528 2.435 [0.9558] [0.6759] [0.4872] Kleibergen-Paap Wald statistics 307.781 380.554 448.094 First-stage House prices House prices House prices Planned area -0.874 *** -1.166 *** -1.359 ***

68 (instrument) (0.120) (0.104) (0.093) Planned area squared 0.028 *** 0.034 *** 0.037 *** (instrument) (0.002) (0.002) (0.001) Municipal merger -0.018 *** -0.019 *** -0.019 *** (instrument) (0.002) (0.002) (0.001) Post-crisis! planned 0.235 *** 0.234 *** 0.233 *** area (instrument) (0.011) (0.010) (0.009) Post-crisis! planned -0.004 *** -0.004 *** -0.004 *** area squared (instrument) (0.0001) (0.0001) (0.0001) First-stage Post-crisis! house prices Post-crisis! house prices Post-crisis! house prices Planned area -5.424 *** -5.406 *** -5.541 *** (instrument) (0.518) (0.448) (0.407) Planned area squared -0.044 *** -0.039 *** -0.034 *** (instrument) (0.008) (0.007) (0.007) Municipal merger -0.046 *** -0.042 *** -0.038 *** (instrument) (0.006) (0.005) (0.005) Post-crisis! planned 9.193 *** 9.528 *** 9.821 *** area (instrument) (0.258) (0.265 (0.267) Post-crisis! planned -0.109 *** -0.117 *** -0.124 *** area squared (instrument) (0.006) (0.006) (0.006) Notes: The table reports the first-differenced IV model (GMM) estimates on the effects of regional house prices on the use of bank loans over the 2004-2011 period. The results are reported separately for the size classes of firms having less than five, ten, and 50 workers, respectively. The dependent variable Bank debt is a ratio of bank loans to total assets at t. The endogenous variables are defined as follows: House prices measure the average prices of previously owned condominiums at the five-digit zip code level in thousand euros per square meter at t. Post-crisis! house prices is an interaction between a dummy for the 2009-2011 period and house prices. The instruments include Planned area, Planned area squared, Municipal merger dummy, Post-crisis! planned area, and Post-crisis! planned area squared. Ln(1+Age) is a natural logarithm of firm age since incorporation (plus one) in years at t. Ln(Total assets) is a natural logarithm of total assets in euros at t-1. Credit score measures the observed creditworthiness of firms at the interval 3-99 (divided by 100) at t-1, where lower values imply higher creditworthiness. All the models include year dummies. NT is the number of firm-year observations. The first-stage results for the instruments are reported omitting the rest of the output from the table. Hansen s J statistics for overidentifying restrictions [p-values in brackets] and the Kleibergen-Paap Wald F statistics for weak identification of the first-stage regressions are provided. Standard errors in parentheses: * p < 0.10, ** p < 0.05, *** p < 0.01. The standard errors are clustered at the firm level. Table A5: Arellano-Bond estimates for 2004-2011 (zip code level house prices) (1) (2) (3) Size class <5 employees <10 employees <50 employees Dependent variable Bank debt Bank debt Bank debt Bank debt (t-1) 0.709 *** 0.709 *** 0.688 *** (0.022) (0.019) (0.018) Bank debt (t-2) 0.027 *** 0.027 *** 0.029 *** (0.009) (0.008) (0.007) House prices 0.013 ** 0.011 ** 0.014 ***

69 (0.006) (0.005) (0.005) Post-crisis*house prices -0.003 *** -0.003 ** -0.003 *** (0.001) (0.001) (0.001) ln(1+age) -0.003 0.001-0.010 (0.010) (0.009) (0.010) ln(total assets) 0.006 0.006 0.020 (0.011) (0.011) (0.014) Credit score 0.019 0.019-0.001 (0.016) (0.014) (0.012) NT 66236 85780 106609 Long-run effects pre-crisis 0.049 ** 0.040 ** 0.048 *** (0.024) (0.020) (0.016) Long-run effects post-crisis 0.037 * 0.031 * 0.038 *** (0.020) (0.018) (0.014) Arellano-Bond test 0.9157 0.5369 0.3907 [0.3598] [0.5913] [0.6960] Sargan test 94.6389 94.0146 115.9058 [0.0220] [0.0244] [0.0004] Notes: The table reports the dynamic Arellano-Bond GMM estimates on the effects of regional house prices on small firms use of bank loans over the 2004-2011 period. The results are reported for the size classes of firms having less than five, ten, and 50 workers, respectively. The dependent variable Bank debt is a ratio of bank loans to total assets at t. The independent variables are defined as follows: Bank debt (t-1) and Bank debt (t-2) are the lagged dependent variables measured at t-1 and t-2, respectively. House prices measure the average regional prices of previously owned condominiums at the five-digit zip code level in thousand euros per square meter at t. Post-crisis! house prices is an interaction between house prices and a dummy for the 2009-2011 period. Ln(1+Age) is a natural logarithm of firm age since incorporation (plus one) in years at t. Ln(Total assets) is a natural logarithm of total assets in euros at t-1. Credit score measures the observed creditworthiness of firms at the interval 3-99 (divided by 100) at t-1, where lower values imply higher creditworthiness. Planned area, Planned area squared, Municipal merger, Post-crisis! planned area, and Post-crisis! planned area squared are included as additional instruments. House prices, Post-crisis! house prices and Credit score are defined as endogenous and ln(total assets) is defined as predetermined. All the models include year dummies. NT is the number of firm-year observations. The long-run effects of house prices for the pre- and post-crisis periods are computed as follows: _b[house prices(t)]/( 1-_b[Bank debt (t-1) -_b[bank debt (t-2)]) and (_b[house prices(t)]+_b[post-crisis! house prices])/(1-_b[bank debt (t-1) -_b[bank debt (t-2)]), respectively. The Arellano-Bond test for the second-order autocorrelation in the first differences residuals and the Sargan test for the overidentification restrictions are reported [p-values in brackets]. All the estimations use a two-step estimator with robust Windmeijer-corrected standard errors. The Sargan test is based on conventional standard errors. Standard errors in parentheses: * p < 0.10, ** p < 0.05, *** p < 0.01. Table A6: FE (within) estimates for 2004-2008 (regional clustering) (1) (2) (3) (4) Measurement level Municipal Municipal Zip code Zip code Dependent variable Bank debt Bank debt Bank debt Bank debt House prices 0.012 * 0.012 * 0.012 *** 0.012 *** (0.006) (0.006) (0.004) (0.004) ln(1+age) -0.040 *** -0.040 *** -0.040 *** -0.040 *** (0.010) (0.007) (0.010) (0.007) ln(total assets) 0.027 *** 0.027 *** 0.027 *** 0.027 *** (0.004) (0.002) (0.004) (0.002) Credit score 0.004 0.004 0.004 0.004

70 (0.006) (0.006) (0.006) (0.006) NT 91036 91036 91036 91036 rho 0.805 0.805 0.806 0.806 r2 0.009 0.009 0.009 0.009 Clustering of std. err. Municipal Zip code Municipal Zip code Notes: The table reports the fixed effects (within) estimates of the effects of regional house prices on small firms use of bank loans over the 2004-2008 period. The results are reported for the firms having less than 50 workers. The dependent variable Bank debt is defined as a ratio of bank loans to total assets at t. The independent variables are defined as follows: House prices measure the average prices of previously owned condominiums at the municipal or five-digit zip code level (as indicated in the table) measured in thousand euros at t. Ln(1+Age) is a natural logarithm of firm age since incorporation (plus one) in years at t. Ln(Total assets) is a natural logarithm of total assets in euros at t-1. Credit score measures the observed creditworthiness of firms at the interval 3-99 (divided by 100) at t-1, where lower values imply higher creditworthiness. All the models include year dummies. NT is the number of firm-year observations. Rho measures the intra-class error correlation. R2 stands for r-squared. Standard errors in parentheses: * p < 0.10, ** p < 0.05, *** p < 0.01. The standard errors are clustered at the municipal or zip code level as indicated in the table. Table A7: FD-IV estimates for 2004-2008 (regional clustering) (1) (2) (3) (4) Measurement level Municipal Municipal Zip code Zip code Dependent variable Bank debt Bank debt Bank debt Bank debt House prices 0.105 *** 0.105 * 0.083 *** 0.083 * (0.040) (0.057) (0.028) (0.049) ln(1+age) -0.033 *** -0.033 *** -0.033 *** -0.033 *** (0.009) (0.007) (0.009) (0.007) ln(total assets) 0.008 *** 0.008 *** 0.008 *** 0.008 *** (0.003) (0.002) (0.003) (0.002) Credit score -0.013 *** -0.013 ** -0.013 *** -0.013 ** (0.005) (0.005) (0.005) (0.005) NT 64939 64939 64939 64939 Hansen s J statistics 1.231 1.089 0.763 0.567 [0.5403] [0.5801] [0.6830] [0.7532] Kleibergen-Paap Wald statistics 1717.770 1863.340 470.510 477.975 First-stage House prices House prices House prices House prices Planned area -1.546-1.546 *** -1.363-1.363 * (instrument) (1.168) (0.403) (1.082) (0.801) Planned area squared 0.030 *** 0.030 *** 0.034 *** 0.034 *** (instrument) (0.011) (0.005) (0.011) (0.013) Municipal merger -0.043 * -0.043 *** -0.053-0.053 *** (instrument) (0.025) (0.009) (0.035) (0.014) Clustering of std. err. Municipal Zip code Municipal Zip code Notes: The table reports the first-differenced IV model (2SLS) estimates on the effects of regional house prices on small firms use of bank loans over the 2004-2008 period. The results are reported for the firms having less than 50 workers. The dependent variable Bank debt is defined as a ratio of bank loans to total assets at t. The independent variables are defined as follows: House prices measure the average prices of previously owned condominiums at the municipal or five-digit zip code level (as indicated in the table) measured in

thousand euros per square meter at t. The instruments include Planned area, its square, and Municipal merger dummy taking a value of one in and after the period of merger. Ln(1+Age) is a natural logarithm of firm age since incorporation (plus one) in years at t. Ln(Total assets) is a natural logarithm of total assets in euros at t-1. Credit score measures the observed creditworthiness of firms at the interval 3-99 (divided by 100) at t-1, where lower values imply higher creditworthiness. All the models include year dummies. NT is the number of firm-year observations. The first-stage results for the instruments are reported omitting the rest of the output from the table. Hansen s J statistics for overidentifying restrictions [p-values in brackets] and the Kleibergen-Paap Wald F statistics for weak identification of the first-stage regression are provided. Standard errors in parentheses: * p < 0.10, ** p < 0.05, *** p < 0.01. The standard errors are clustered at the municipal or zip code level as indicated in the table. Table A8: Arellano-Bond estimates for 2004-2008 (regional clustering) (1) (2) (3) (4) Measurement level Municipal Municipal Zip code Zip code Dependent variable Bank debt Bank debt Bank debt Bank debt Bank debt (t-1) 0.648 *** 0.648 *** 0.650 *** 0.650 *** (0.024) (0.024) (0.023) (0.024) House prices 0.057 *** 0.057 *** 0.027 *** 0.027 *** (0.017) (0.016) (0.008) (0.009) ln(1+age) -0.017-0.017-0.018-0.018 (0.017) (0.015) (0.016) (0.015) ln(total assets) 0.011 0.011 0.014 0.014 (0.023) (0.020) (0.023) (0.020) Credit score 0.022 * 0.022 0.022 * 0.022 (0.011) (0.015) (0.012) (0.015) NT 64939 64939 64939 64939 Long-run effects 0.161 *** 0.161 *** 0.076 *** 0.076 *** (0.049) (0.048) (0.024) (0.027) Arellano-Bond test 1.24 1.24 1.27 1.27 [0.216] [0.215] [0.203] [0.203] Hansen test 13.83 19.18 15.38 20.71 [0.679] [0.318] [0.568] [0.240] Clustering of std. err. Municipal Zip code Municipal Zip code Notes: The table reports the dynamic Arellano-Bond GMM estimates on the effects of regional house prices on small firms use of bank loans over the 2004-2008 period. The results are reported for the firms having less than 50 workers. The dependent variable Bank debt is a ratio of bank loans to total assets at t. The independent variables are defined as follows: Bank debt (t-1) is the lagged dependent variable measured at t-1. House prices measure the average regional prices of previously owned condominiums at the municipal or five-digit zip code level (as indicated in the table) in thousand euros per square meter at t. Ln(1+Age) is a natural logarithm of firm age since incorporation (plus one) in years at t. Ln(Total assets) is a natural logarithm of total assets in euros at t-1. Credit score measures the observed creditworthiness of firms at the interval 3-99 (divided by 100) at t-1, where lower values imply higher creditworthiness. Planned area, its square, and Municipal merger dummy taking a value of one in and after the period of merger, are included as additional instruments. All the models include year dummies. House prices and Credit score are defined as endogenous and ln(total assets) is defined as predetermined. NT is the number of firm-year observations. The long-run effects of house prices are computed as follows: _b[house prices(t)]/(1-_b[bank debt (t-1)]). The Arellano-Bond test for the second-order autocorrelation in the first differences residuals and the Hansen test for the overidentification restrictions are reported [pvalues in brackets]. The estimations use a one-step GMM estimator with standard errors 71

clustered at the municipal or zip code level as indicated in the table. Standard errors in parentheses: * p < 0.10, ** p < 0.05, *** p < 0.01. 72