How do insured deposits affect bank stability? Evidence from the 2008 Emergency Economic Stabilization Act

Transcription

1 How do insured deposits affect bank stability? Evidence from the 2008 Emergency Economic Stabilization Act Claudia Lambert, Felix Noth and Ulrich Schüwer preliminary version June 4, 2013 Abstract This paper tests whether an increase in the amount of insured deposits causes a bank to become more risky. We use variation introduced by the U.S. Emergency Economic Stabilization Act in October 2008, which increased the deposit insurance coverage from $100,000 to $250,000 per depositor. For some U.S. banks, this event significantly increased the amount of insured deposits. For other U.S. banks, it had only a minor effect. Our analysis shows that an increase in the amount of insured deposits induces the affected banks to become more risky relative to the unaffected banks. In particular, the affected banks increase their investments in risky loans. To our knowledge, this is the first study that provides causal within-country evidence on the effect of insured deposits on bank stability for a large economy. Keywords: financial crisis, deposit insurance, bank regulation JEL Classification: G21, G28 We would like to thank Jan Pieter Krahnen, Karen K. Lewis, Isabel Schnabel, Falko Fecht, Simon Kwan, the participants of the 2013 Financial Intermediation Research Society Conference in Dubrovnik, the 2012 Research Seminar in Riezlern and the 2013 Research Workshop in Financial Economies in Mainz for valuable comments and suggestions. Any remaining errors are, of course, our own. DIW Berlin, German Institute for Economic Research, Mohrenstr. 58, Berlin; Germany Goethe University Frankfurt, Department of Finance, Grüneburgplatz 1, Frankfurt am Main, Germany, claudia.lambert@hof.uni-frankfurt.de, noth@finance.uni-frankfurt.de, schuewer@finance.unifrankfurt.de.

2 1 Introduction The deposit insurance is a corner stone of many banking systems worldwide because it helps to protect small savers and to prevent bank runs. However, it also gives banks incentives for excessive risk-taking because, first, it weakens the market discipline carried out by creditors, and second, mispricing of the deposit insurance premium, which come from the regulators limited ability to assess risks and to charge risk-adjusted premiums, make higher risk-takings attractive for shareholders. Empirical evidence from cross-country studies indicates that deposit insurance tends to increase the likelihood of banking crises (e.g., Demirgüç-Kunt and Detragiache, 2002). 1 Whether the deposit insurance also had an impact on banks risk taking during the recent financial crisis has not been explored yet. An effect of the deposit insurance may be complementary to other important aspects of the financial crisis that have been explored by the literature, such as the effect of bank CEO incentives (Fahlenbrach and Stulz, 2011) and government bailout policies on banks risk taking (Gropp et al., 2011; Dam and Koetter, 2012; Black and Hazelwood, 2012). This paper explores how a bank s amount of insured deposits affects its stability and lending decisions during the recent financial crisis. The main challenge for such an empirical test is an identification problem. The banks with more insured deposits might be the ones that take more risks, or, the more risky banks might be the ones that make more efforts to attract deposits and thus have more insured deposits. To circumvent this problem, we use variation introduced by the U.S. Emergency Economic Stabilization Act in October 2008, which increased the deposit insurance coverage from $100,000 to $250,000 per depositor. This change increased the total sum of insured deposits in the U.S. from roughly $4,800 billion to roughly $5,300 billion. Importantly for our identification strategy, banks were affected differently. For some banks, this event significantly increased the amount of insured deposits ( affected banks ). For other banks, it only had a minor effect ( unaffected banks ). Using the affected banks as the treatment group and the unaffected banks as the control group, we employ a difference-in-difference estimation technique conditional on propensity 1 On the contrary, Gropp and Vesala (2004) find in a cross-country study for 15 European countries that explicit deposit insurance may increase bank stability because of market monitoring through non-deposit creditors. As depicted by Morrison and White (2011, p. 3400), our understanding of the design and consequences of deposit insurance schemes is in its infancy. 2

3 score matching. Through the matching procedure we make sure that both groups of banks are similar before the event and thereby that our results do not reflect systematic differences in certain parameters between both groups. Our empirical analysis shows that an increase in the amount of insured deposits causes the affected banks to become more risky relative to the unaffected banks. This is reflected in estimations for a variety of risk measures, including banks predicted probabilities of default, banks z-scores 2, and loan loss provisions. Further, our analysis shows that affected banks increase their investments in risky commercial real estate loans and consumer loans, which may explain lower asset quality and higher bank risk, relative to unaffected banks. Thereby, the study provides empirical evidence on unintended long-term consequences of an important and controversial part of bank regulation. Figure 1 presents illustrative evidence. The upper left graph shows the number of failures for affected banks and unaffected banks in the period following the introduction of the Emergency Economic Stabilization Act. 3 We observe that bank failures occur more often in the group of affected banks than in the group of unaffected banks. The upper right graph shows predicted probabilities of default that we estimate with a linear probability model. Here, affected banks show on average higher predicted probabilities of default than unaffected banks following the introduction of the act in the fourth quarter of The lower left graph shows banks z-scores, which are relatively lower for affected banks after the event (indicating lower bank stability). Finally, the lower right graph shows banks loan loss provisions over assets, which are relatively higher for affected banks after the event. In summary, the graphs of Figure 1 provide a first indication that affected banks become riskier relative to the unaffected banks following the increase of insured deposits as induced by the Emergency Economic Stabilization Act. 2 We use banks z-scores as a measure for bank risk following Laeven and Levine (2009). The z-score is defined as the sum of a bank s return on assets and its equity to assets ratio, standardized by the volatility of the bank s return on assets. It thus measures a bank s ability to absorb losses by its equity. A lower z-score indicates lower bank stability and accordingly higher bank risk. 3 As we explain later in more detail, the change in insured deposits that became effective in the fourth quarter of 2008 only became visible in the banks balance sheets in the third quarter of Therefore, we can only report bank failures for banks classified as affected or unaffected after the third quarter of

4 Figure 1: Illustrative evidence q4 2004q4 2006q4 2008q4 2010q4 affected unaffected affected unaffected (a) Number of bank failures (b) Predicted probability of default q4 2004q4 2006q4 2008q4 2010q4 2002q4 2004q4 2006q4 2008q4 2010q4 affected unaffected affected unaffected (c) z-score (d) Loan loss provisions/ assets The upper left graph shows the number of bank failures per year for the group of affected banks and the group of unaffected banks (each group comprises a total of banks). Note that the 2009 bank failures only include the fourth quarter of 2009, and the bank failures in 2012 only include the first, second and third quarter of The upper right graph shows banks predicted probabilities of defaults. The mean values for affected banks and unaffected banks are represented by a solid and dotted line, respectively. The lower left graph shows banks z-scores, where a high value represents a more stable bank. The lower right graph shows banks loan loss provisions over assets. The three latter graphs cover the period from the second quarter of 2001 to the second quarter of The two vertical lines surround the fourth quarter of 2008 in which the U.S. Emergency Economic Stabilization Act was introduced. The dotted line depicts the previous quarter when banks may have begun to anticipate the adoption of the law. 4

5 The contribution of our study to the literature is twofold. First, we add to the large literature that examines the effect of insured deposits on bank stability. 4 Important theoretical contributions include Merton (1977) and Diamond and Dybvig (1986). 5 The existing empirical literature typically conducts cross-country studies to examine how the existence of explicit deposit insurance affects the stability of a country s banking system (e.g., Demirgüç-Kunt and Detragiache, 2002; Gropp and Vesala, 2004). To our knowledge, our study is the first empirical assessment of the relation between deposit insurance and bank stability that provides withincountry evidence for a large economy using bank-level data. 6 We are able to restrict our study to U.S. banks because the legislative event did not affect all banks equally in practice. This allows us to examine how affected banks change their risk-taking and lending decisions within a homogeneous macroeconomic area. Both perspectives, the cross-country perspective from the existing literature and the within-country perspective in our study, provide important insights and complement each other. Second, we demonstrate that the increase in insured deposits during the 2008 financial crisis, which may have had beneficial short-term effects, has long-term effects on banks risk-taking incentives and can amplify financial system distortions. This finding adds to the long-standing as well as the more recent assessments of several researchers that the U.S. deposit insurance needs to be changed (e.g., Berlin et al., 1991; Pennacchi, 2006, 2009). The paper proceeds as follows. Section 3 describes the U.S. Emergency Economic Stabilization Act of October 2008 as regards its consequences for banks insured deposits. Section 4 presents our identification strategy, the data and summary statistics. Section 5 explains the empirical model and shows the estimation results. Section 6 analyzes whether risk taking of banks differs for banks with high or low capital levels subject to state guarantees or not after the change in the deposit insurance scheme. Finally, Section 7 concludes. All tables appear in the appendix. 4 For an overview, see, for example, Demirgüç-Kunt et al. (2008). 5 More broadly related to the effects of deposit insurance are the seminal papers by Diamond (1984) and Gorton and Pennacchi (1990) that study the function of financial institutions to create liquid assets for uninformed customers. In a recent paper, Morrison and White (2011) argue that socially too few deposits are made in equilibrium because of market failures, and they find in their set-up that deposit insurance should be funded out of general taxation to improve welfare. 6 Related to our study is Wheelock (1992) who examines the voluntary membership in the Kansas state insurance system and its consequences for bank failures in the 1920s. 5

6 2 The Emergency Economic Stabilization Act of October The Emergency Economic Stabilization Act of October 2008 The Emergency Economic Stabilization Act was a reaction to the financial crisis of 2007 and 2008, which is considered the worst such crisis since the Great Depression. It was signed into law on October 3, The rationale of the act was to restore consumer confidence in the banking system and financial markets. The central item of the act is the Troubled Asset Relief Program (TARP) that allowed the Treasury Department to spend $700 billion to support troubled institutions (Shah, 2009). 7 The program then enacted the Capital Purchase Program (CPP), which enabled the government to purchase equity directly from troubled institutions. The act also includes further stabilizing actions, e.g., allowing the FED to pay higher interest rates on banks deposits held as reserve requirements, or foreclosure avoidance and homeowner assistance with regard to mortgage payments (see, e.g., Nothwehr, 2008). Importantly for our study, Section 136 of the Emergency Economic Stabilization Act provided a piece of reform that affected deposit insurance: It raised insured deposits per account from $100,000 to $250,000, effective as of October 3, Though initially temporarily until December 31, 2009, the increase of insured deposits was prolonged through January 1, 2014 on May 20, 2009 by the Helping Families Save Their Homes Act. Finally, the introduction of the Dodd-Frank Act, signed into law on July 21, 2010, made the temporary increase in insured deposits permanent. 4 Identification strategy and empirical model To assess how an increase in insured deposits affects banks risk takings, we have to consider potential parallel macroeconomic and industry-wide factors that affect all banks, independent of the regulatory change. It would be misleading to simply test how (or whether) banks adapt 7 The amount was reduced to $475 billion with the Dodd-Franck Act of

7 their risk takings after the regulatory change. Rather, we need to explore how affected banks adapt their risk taking relative to the risk taking of a counter-factual that is similar to the affected banks. We therefore construct a group that includes the affected banks, i.e., the treatment group, and a group that includes banks representing the counter-factual behavior, i.e., the control group. The identification of the treatment group and the control group is the main challenge of our study. In particular, we need to make sure that banks in both groups are similar before the event. We proceed in using a difference-in-difference estimation technique with time and bank fixed effects. Our identification strategy is based on variation introduced by the regulatory change and as such the intensity of banks being effectively affected by the change in the deposit insurance scheme initiated through the Emergency Economic Stabilization Act of October In particular, we observe that some banks reported relatively high increases of their insured deposits following the regulatory change, while other banks reported relatively low increases. We classify the former banks as affected banks, which form our preliminary treatment group, and we classify the latter banks as unaffected banks, which form our preliminary control group. Important for our identification strategy, we have to consider that the external variation which determines the selection of banks into both groups is not purely random. The affected banks are those with relatively many deposit accounts above $100,000, while the unaffected banks are those with relatively few such accounts. We need to make sure that this difference between banks in both groups does not create a selection bias for our estimation results. In principle, the following could be relevant for our estimations: First, the difference between banks may be uncorrelated to the dependent variables we are interested in. In this case, there is no problem and we get unbiased estimation results. Second, the difference between banks may have a static level effect on the dependent variables we are interested in. For example, banks in the treatment group and control group have different business models, such that banks in the treatment group are systematically more likely to fail by certain percentage points relative to banks in the control group. In this case, a difference-in-difference estimation technique also provides unbiased estimation results because it explores how differences between both groups change post an event, not the different levels itself. Furthermore, bank fixed effects in 7

8 our estimation account for potential static differences between banks. Third, the difference between banks may have a dynamic effect on the dependent variables we are interested in. For example, banks in the treatment group are systematically more risk loving during recessions and more risk averse during booms relative to banks in the control group. In this case, we are limited in controlling for this bias in our estimation model. Therefore, we need to rule out that this problem exists by restricting our sample to banks in the treatment group and banks in the control group that are similar over the business cycle. Accordingly, we reduce our treatment group and control group to a matched sample of affected banks and unaffected banks that show similar characteristics over many variables and over time. In the following of this section, we first describe the data used in this study (Subsection 4.1). We then describe in more detail our definition of affected and unaffected banks (4.2). Next, we explain how we construct a matched sample of banks (4.3). The subsequent subsections provide the summary statistics (4.4) and further evidence on the similarity of treatment and control group over the business cycle (4.5). 4.1 Data Our data comes from different sources of the FDIC as well as the U.S. Department of the Treasury. First, we use quarterly data on U.S. commercial banks from the FDIC Statistics on Depository Institutions, which includes detailed balance sheet, income statement and deposits data for all FDIC-insured banks in the United States since Note that the data refer to individual FDIC-insured institutions, which may be part of larger bank holding companies. Second, we use information on bank failures in the period 1993 to 2012 from the FDIC Failed Bank List to estimate probabilities of default. Finally, we use data about which banks received TARP from the TARP Investment Program Transaction Reports, which are available on the website of the U.S. Department of the Treasury. Our main analyses cover the period up to ±10 quarters around the introduction of the Emergency Economic Stabilization Act in Q4 2008, i.e., the period from Q to Q In Q our preliminary sample consists of 8,463 commercial banks. Previous studies that also used the FDIC data found that some of the data is erroneous or includes banks that 8

9 are not viable. Therefore, we follow Berger and Bouwman (2009) and exclude banks that (1) have no commercial real estate or commercial and industry loans outstanding, (2) have zero or negative equity capital, (3) hold assets below $25 million or (4) hold consumer loans exceeding 50% of gross total assets. 8 Accordingly, we are left with 5,696 banks. We additionally exclude banks for which we find missing values for our variables of interest, i.e., TO BE ADDED. This leaves us with 5,307 banks. Furthermore, we want to exclude biases from newly founded banks and therefore require banks existence two years before the Emergency Economic Stabilization Act was enacted, which further reduces the sample to 5,305 banks. As described in the following subsection in more detail, our identification strategy is based on the reported change in insured deposits per customer for each bank from the second to the third quarter of Our sample therefore only includes banks that did not fail before the third quarter of 2009, i.e., a total of XXX banks. Further, our identification strategy only considers banks in the top quartile (treatment group) or lowest quartile (control group) of reported changes. This cuts our sample in half, i.e., we get a total of XXX banks. Finally, as described in Subsection 4.3 in more detail, we apply propensity score matching and require that banks in the treatment group and control group have very similar characteristics before the regulatory change (1% caliper). After our matching procedure, the final data set includes 2,208 banks, of which 1,104 are classified as affected and 1,104 are classified as unaffected. While our main difference-in-difference estimations cover the period from Q to Q2 2011, we also use data of the period Q to Q to estimate probabilities of default for each bank and quarter. For this extended period, our estimation includes on average 10,000 banks per quarter and a total of 767,649 bank-quarter observations. 4.2 Definition of affected and unaffected banks We define our treatment and control group based on the change in insured deposits initiated through the Emergency Economic Stabilization Act of The increase in insured deposits is reflected in the reporting of FDIC insured banks, which includes the total amount of 8 Some further exclusion criteria used by Berger and Bouwman (2009) are not relevant for our sample. 9

10 deposits and the amount of insured deposits per bank and quarter. Note that although the act was effective as of the fourth quarter of 2008, banks were required to continue reporting their amount of insured deposits based on the old $100,000 amount per customer for several quarters, and only starting with the Q reporting period adapted the new $250,000 amount per customer for the calculation and reporting of insured deposits. Thus, we use the reported amounts of insured and total deposits of the third quarter of 2009 to differentiate between treatment and control group. We construct the sample of affected and unaffected banks following two steps. First, we calculate the change in the ratio of insured deposits over assets from Q2 to Q Second, we assign banks to the treatment group ( affected ) if this change is in the top quartile of the sample. Banks that exhibit a change in the ratio of insured deposits over assets in the lowest quartile belong to the control group ( unaffected ). Intuitively, affected banks have relatively more customers in their portfolio with deposits exceeding $100,000 before the introduction of the act. Figure 2: Change of insured deposit ratios q4 2004q4 2006q4 2008q4 2010q4 2002q4 2004q4 2006q4 2008q4 2010q4 affected unaffected affected unaffected (a) Insured deposits/assets (b) Insured deposits/deposits The figure shows the development of the mean values of the ratio of insured deposits to assets (a) and the ratio of insured deposits to deposits (b) from the second quarter of 2001 to the second quarter of The two vertical lines surround the fourth quarter of 2008 in which the Emergency Economic Stabilization Act took place. The mean value for affected banks is represented by a solid line. The mean values for unaffected banks is represented by a dotted line. Fig. 2 shows the ratio of insured deposits to total assets (a), and the ratio of insured deposits to total deposits (b) before and after the Emergency Economic Stabilization Act was introduced. 10

11 In each sub-figure we present separate average values for affected banks and unaffected banks. In both sub-figures we observe that the ratio of insured deposits to assets (deposits) between affected and unaffected banks converges as a consequence of the change in regulation. Both sub-figures also show that our identification does not hinge on the denominator of this ratio. By construction we see a sharp change for the affected banks relative to the group that we consider unaffected by the event. Using total assets or deposits as the denominator only changes the level, but not the relative impact of the event on the control and treatment group. This also holds for our later results. 4.3 Matching estimation To disentangle a potential selection bias from the treatment effect, we first estimate a logit model that explains the probability that a bank is materially affected by the legislative change. The propensity score matching technique uses bank-specific characteristics and constructs similar samples before the event. In particular, we match banks based on key characteristics which determine bank stability closely following Wheelock and Wilson (2000), which reflect banks capital adequacy, asset quality, management, earnings and liquidity. We estimate the model in averaging the data for the period from Q to Q3 2008, i.e., the two and a half years before the act was introduced. 9 In particular, we estimate the following matching equation: Affected i = β 0 + β 1 Equ i + β 2 Loans i + β 3 RELoans i + β 4 CILoans i + β 5 ORE i + β 6 Inc i + β 7 Liqu i + β 8 NP L i + β 9 T A i + β 10 RoA i + ϵ i (1) The explanatory variables in this equation are: total equity to total assets (Equ), total loans to total assets (Loans), real estate loans to total loans (RELoans), commercial and industry loans to total loans (CILoans), other real estate owned to total assets (ORE), income earned, not collected on loans to total assets (Inc), liquidity measured as the difference between federal 9 Note that subprime loan problems surfaced for the first time before the 2001 recession. These problems quickly subsided during the period from 2002 to 2005, which was accompanied by a recovery of the economy. DiMartino and Duca (2007) shows that this pattern gradually deteriorated thereafter. Given this development we conduct the matching two years before the law was introduced. 11

12 funds purchased minus federal funds sold standardized by total assets (Liqu), non-performing loans to total assets (NPL), the size represented by total assets measured in logarithmic form (TA) and the net interest income (plus total non-interest income (RoA). We restrict the sample to commercial banks. For our main specification, we use the nearest neighbor matching procedure (on a one-toone basis), which selects the banks closest in terms of their propensity scores. We conduct the analysis without replacement, which means that a neighbor can only be used once. We require common support and impose a tolerance level of 1%, the so-called caliper. The caliper is equivalent to choosing an individual from the comparison group as a matching partner for a treated individual that lies within the caliper (propensity range) and is closest in terms of propensity score. 4.4 Summary statistics We present mean values and standard deviation for our main variables and further bank characteristics in Table 2 for the period from Q to Q The table differentiates between affected and unaffected banks in order to compare characteristics of both samples before the event. Since our difference-in-difference estimation is based on a matched sample, which uses affected banks as the treatment group and unaffected banks as the control group, we technically insure that banks in both groups have similar characteristics. Additionally, as suggested by Imbens and Wooldridge (2009), Table 2 also reports normalized differences to compare both groups with respect to important bank characteristics. Normalized differences are a scale-free measure of the difference in distributions, and calculated as the difference in averages by treatment status, scaled by the square root of the sum of the variances (Imbens and Wooldridge, 2009, p. 24): = X 1 X 0 V AR1 + V AR 0, (2) where X 1 and X 0 represent the mean values of the group of affected banks and unaffected 10 See Table 10 for a description of the FDIC data that we use in our study. 12

13 banks, respectively, and V AR 1 and V AR 0 represent the respective variances. As a rule of thumb, groups are regarded as sufficiently equal and adequate for linear regression methods if normalized differences are basically in the range of ± Overall we find that both groups resemble quite similar characteristics before the event. 12 Characteristics regarding size (e.g., total assets), structure of the asset side (e.g., real estate loans, C&I loans) in both groups are all quite comparable between affected and unaffected banks referring to normalized differences well below ±0.25. With respect to risk taking (e.g., probability of default, z-score, total risk-based capital, provisions and charge-offs), we observe that both groups behave quite similar before the introduction of the Emergency Economic Stabilization Act that changed the amount of insured deposits in Q Normalized differences well above ±0.25 are only evident for the share of insured to deposits and assets, which we use for identifying treatment and control group. In summary, Table 2 shows that dividing the sample in affected and unaffected banks by the change in insured deposits leaves us with two sub-samples that resemble similar characteristics before the event. The descriptive statistics show that our identification procedure is appropriate and that our results are not likely to be disturbed by differences in groups. [Table 2] 4.5 Similarity of treatment and control group over the business cycle In this section we provide further evidence that our results are not driven by dynamics resulting from differences between the treatment group and the control group over the business cycle. As such, we inspect whether the trend of the z-score and the probability of default is parallel for both groups in Section Additionally, in Section we observe whether the normalized differences are within the required range for different risk proxies comparing both groups of banks. 11 Note the difference with the t-statistics, which is sensitive to the sample size, and which tests whether there is a significant difference in means, not, as normalized differences, whether linear regression methods tend to be sensitive the the specification. 12 This is not surprising because we use propensity score matching to construct our sample. 13

14 4.5.1 Parallel trend assumption To adequately employ a difference-in-difference estimation technique we have to guarantee that the parallel trend assumption is satisfied. As such, we have to guarantee that the developments of the probability of default and the z-score for the treatment and control group follow similar trends before the event. Analogous to previous studies we graphically inspect the trend of mean probability of default and z-scores for both groups in Fig. 1, and we confirm that the parallel-trend assumption holds. Further, in Subsection 4.4 and shown in Table 2, we observe that the groups of affected and unaffected banks do not differ regarding most characteristics Relevance of the business cycle to bank risk Using a propensity matching technique as in section 4.3, we alleviated concerns that banks differ between the group of banks that benefited from the change in the law and the group that did not. Furthermore, we would like to show that banks are similar across the business cycle with regard to different risk proxies. Doing so, we would like to rule out that banks that benefited from the change in the deposit insurance scheme were more risky ex-ante. In Fig. 3 and 4 we graphically compare the development of our main variables of interest and different risk proxies for separately for the treatment and the control group. Additionally, we also show the evolution of the normalized differences. In Figure 3 4 we observe that the values of the normalized differences are between ±0.25 for the period 2000 to 2007, which is equivalent to non-significant differences between groups (Imbens and Wooldridge, 2009). Thereby, concerns are alleviated that banks risk taking differed for banks in the treatment and control group before the event. For all variables we observe that groups are similar before the event with trends slightly more diverging thereafter. [Fig. 3] 14

15 5 Empirical model and results 5.1 Baseline estimation By applying a conditional difference-in-difference estimation technique which is eqivalent to a matching of groups procedure beforte estimating the treatment effect, we estimate whether higher risk taking by banks is systematic and can be attributed to the change in regulation. Formally, we estimate the following equation: Risk it = β 0 + β 1 Event t + β 2 (Event t Affected i ) + τ γ + ν i + ϵ it (3) The short-hand Risk it is a proxy for bank risk of bank i at time t. The different measures that we use for Risk it are common in the literature and explained in more detail in the following subsection. Our event window is the fourth quarter of Accordingly, the variable Event t is a time dummy with a value of zero for all quarters before the Emergency Economic Stabilization Act was introduced (t < Q4 2008) and a value of one for all quarters after the event (t Q4 2008). The variable Affected i is a dummy variable of bank i that is one if the bank experienced a sharp increase in insured deposits per customer (belongs to the top 25% quantile) and thus belongs to the treatment group, and zero otherwise (belongs to the bottom 25% quantile). Hence, the interaction term Event t Affected i is one if both the variable Event t and the variable Affected i amount to one, and zero otherwise. The corresponding coefficient β 2 is the main interest. It captures the average effect of the introduction of the act on the bank stability of affected banks. The variable τ γ represents yearly time fixed effects (TE). Further, we are concerned that unobserved differences between banks might influence our results. Thus, we include fixed effects (FE) ν i for each bank i. Finally, ϵ it is the idiosyncratic error term. To account for heterogeneity among banks, we use clustered standard errors at the bank level. For robustness, we re-estimate our baseline estimation without bank fixed effects. The variable Affected i that otherwise interferes with bank fixed effects then enters the equation. 15

16 5.2 Results for different risk proxies Predicted probabilities of default First, we use banks predicted probabilities of default as a proxy for bank risk, Risk it. We thereby provide evidence whether the predicted probabilities of default of affected banks increase significantly relative to unaffected banks after the regulatory increase of insured deposits in the fourth quarter of Because banks predicted probabilities of default are not directly observable, we follow three steps: First, we run a probability model with historical bank failures to explore how different bank characteristics determine a bank s probability of default. Second, we use these regression results to calculate predicted probabilities of default for all banks in our sample. Finally, we use the predicted probability of default as dependent variable in Eq. (3). Our probability model is based on a sample of all U.S. banks registered at the FDIC in the period of , i.e., a total of about 10,000 banks on average per quarter, of which 574 banks failed during this period. 13 In particular, we estimate the following linear probability model: Failure it =β 0 + β 1 Equ it + β 2 Loans it + β 3 RELoans it + β 4 CILoans it + β 5 ORE it + β 6 Inc it + β 7 Liqu it + β 8 NP L it + β 9 T A it + β 10 RoA it + τ γ + ν i + ϵ it (4) The explanatory variables are the same as the ones we use for the propensity score matching in the previous section and based on Wheelock and Wilson (2000). The variables τ γ and ν i represent yearly time fixed effects and bank fixed effects, respectively. We use a linear probability model instead of a nonlinear probability model because we want to include bank and time fixed effects in our model. Thus, we can correctly address that bank failures may be explained by bank- and time-specific factors that are not captured by our explanatory variables. With a nonlinear regression model, the introduction of many dummy variables leads to i) practical problems because the presence of many dummy variables makes 13 The total number of banks for the year 1993 was over 13,000, and then declined subsequently to about 7,000 for the year The total number of observations for our estimation is 767,

17 the estimation much more difficult, 14 and ii) the so-called incidental parameters problem (see, e.g., Greene et al., 2002; Fernández-Val, 2009). 15 We are aware that using a linear model to predict bank failures comes with the problem that we do not capture the shape of the distribution of bank failures correctly, and we may predict probabilities of default of less than 0% and above 100%. However, we weight this bias less severe than the potential problem from neglecting fixed bank and time effects or the potential incidental parameters problem. Furthermore we are not directly interested in the level of the coefficients of the explanatory variables, but in their relative weights. 16 From the (unreported) regression we observe that most of the variables determine bank failures in an intuitive way: banks with a larger capital buffer, with more loans relative to total assets, larger banks, banks with higher liquidity and banks that are more profitable are significantly less likely to fail whereas banks with higher non-performing loan ratios and risky real estate investments have a significant higher probability of failure. Next, we calculate the predicted probability of default for each bank and each quarter. See Figure 1 for the development of the mean predicted probability of default for the groups of affected and unaffected banks, and Table 2 for related summary statistics. Finally, we estimate Equation (3) for the sample of affected banks (treatment group) and unaffected banks (control group), using the predicted probability of default of each bank and quarter as dependent variable. Table 3 shows results for five different specifications of Eq. (3). The first column presents results for an OLS regression without bank fixed effects using ± 8 quarters around the event. The four remaining columns provide results for regressions using bank fixed effects to control for unobserved bank specific effects for periods of ± 4, 6, 8, 10 quarters around the event. Each column of Table 3 includes yearly time fixed effects and standard errors (in parentheses) that are clustered on the bank level. The interaction term Affected i *Event t measures the average difference in banks default probabilities between 14 When we try to fit a fixed effects logit model, the estimations do not converge or drop huge amounts of observations. Only roughly 25,000 from the original 767,649 observations stay in the estimation. 15 According to Fernández-Val (2009), the incidental parameters problem arises because the unobserved individual characteristics are replaced by inconsistent sample estimates, which, in turn, bias estimates of model parameters. 16 Note that despite the practical problems and concerns mentioned above, a recent literature review on the determinants of bank failures by Mayes and Stremmel (2012) shows that logit and probit models are most common in this field of research. 17

18 affected and unaffected banks after the event relative to the period before the event. In particular the interaction term in Table 3 indicates whether affected banks become riskier (higher default probability) or less risky (lower default probability) after the event relative to the group of unaffected banks. The first column shows a positive and significant interaction effect that indicates that banks more affected by the raise in insured deposits have higher probabilities of default after the event relative to the control group. This results is corroborated by the remaining four columns that include bank fixed effects. Those columns also indicate the economic significance of the effect increases with the window size around the event. While for the shortest period of ± 4 quarters, the increase in default probabilities for affected banks relative to the control group is 15 basis points after the event, this effect increases to 22 basis point for the window of ± 10 quarters. The economic significance is roughly 50%, calculated as the average increase of the probability of default of the treatment group relative to the control group after the event. Note that in calculating this measure we omit the level effect prior to the event. The results in Table 3 support results from the literature that emphasize the detrimental effects of deposit insurance schemes on bank stability (e.g., Demirgüç-Kunt and Detragiache, 2002). With regard to this body of literature that investigates the effects of deposit insurance schemes on bank stability on a cross-country base, our study complements this research by conducting the first within-country analysis for the United States. Furthermore, we are able to identify the direction of this effect exploiting the Emergency Economic Stabilization Act as at least quasi exogenous shock and are thus able to evaluate recent policy changes in the U.S. [Table 3] 18

19 5.2.2 Z-scores Z-scores As a second risk measure we use the bank z-score, which measures a bank s ability to absorb losses by its equity. Following Laeven and Levine (2009), the bank z-score is defined as the natural logarithm of the sum of a bank s return on assets, ROA, and its equity to assets ratio, AR, standardized by the volatility of bank s return on assets, ST D(ROA), i.e., z-score it = log ROA it + AR it ST D(ROA) it. (5) ROA it and AR it reflect book values as reported by the banks. Technically, we calculate ST D(ROA) it as the 12 quarter rolling standard deviation of return on assets for each bank and quarter. Note that the definition comprises the natural logarithm because the measure is otherwise highly skewed. A larger value of the z-score is associated with a more stable bank. Table 4 shows the regression results for Eq. (3) using the banks z-scores as the dependent variable. The set-up of Table 4 follows that of Table 3. In column (1) we observe a negative and significant coefficient of the interaction term, which indicates that affected banks become riskier after the event relative to the control group. In other words, the raise in insured deposits from $100,000 to $250,000 in Q has a destabilizing effect on banks that benefited from this act. The point estimate of 0.17 shows that this effect is also economically relevant. As we are estimating a log-linear model (the left-hand side represents the logarithm of banks z-score) the interaction term is equivalent to a 8% difference in stability between treatment and control group after Q (i.e., banks in the treatment group become 8% less stable). Columns (2) to (5) of Table 4 confirm the results from column (1) using bank fixed affects to control for bank characteristics that might disturb the initial results in column (1). We find that the interaction term remains significant and negative, indicating that affected banks become riskier after the event. Results further indicate that this effect is very persistent. Even for the period of ±10 quarters around the event, we find a significant effect of the interaction term that is statistically and economically similar. Economic significance of these 19

20 effects amounts to roughly 7%. [Table 4] Z-score components To determine the drivers for the z-score results, Table 5 shows results for the volatility of banks returns on assets, SD(RoA), equity-to-assets ratios, and returns on assets, RoA. Column (1) shows that affected banks have a higher volatility after the event in comparison to their unaffected peers, as reflected in the positive and significant coefficient of the interaction term. Column (2), which uses the equity-to-assets ratio as dependent variable, shows a negative and significant coefficient for the interaction term. In Column (3) we report that the profitability of affected banks is significantly lower after the event relative to the unaffected banks. Given that these banks take more risks after the introduction of the law, one is likely to expect higher profits for these banks. In Section we will show that affected banks classify a larger amount of loans as non-performing, which could explain a decline in profits. Furthermore, looking at interest income (total and interest income separately) we observe no significant disparities for affected and unaffected banks 17. Overall, Table 5 shows that the lower z-score values for affected banks relative to unaffected banks after the event stem from all three components of the z-score. [Table 5] Asset quality We further approximate bank risk with five other variables that capture banks asset quality. Here we rely on the ratio of loan loss provisions to total assets, charge-offs to total assets, the ratio of banks non-accrual assets, the amount of other real estate owned to banks total assets and non-performing loans 18 to banks total assets. Higher values for one of these measures 17 Note that we standardize interest income with fixed values of assets one period before the event. 18 Classified as assets past due 90 days still accruing interest. 20

21 indicate higher asset risk for banks. Column (1) of Table 6 shows results for banks loan loss provisions over assets as the dependent variable. We observe higher provisions for affected banks after the event relative to their unaffected peers. Column (2) shows similar results for banks charge-offs over assets. This is not surprising given the results for loan loss provisions over assets, because both measures are highly correlated. Next, we evaluate the effect of the change on non-accrual assets over assets. As shown in column (3), we observe that affected bank have a higher volume of nonaccrual assets after the event relative to the control group, which is qualitatively similar to the results above. Column (4) shows results for other real estate owned over assets, which indicates foreclosure activity. We find that this measure significantly increases after the event for the group of affected banks relative to unaffected banks. Finally, column (5) shows that non-performing loans increase for affected banks relative to the control group after the event. In summary, we get a clear indication that the volume of problem loans increased for affected banks after the event relative to unaffected banks. Note that magnitude, sign and level of significance of the coefficients remains within range if the dependent variable is standardized by asset values one period before event (q3 2008). [Table 6] Note further that the trend for loan loss provisions over assets is illustrated in Figure 1, and the trends for the other measures of asset quality are illustrated in Figure 3 in the appendix Asset composition Finally, we use measures as dependent variables that reflect asset composition, which is a more indirect measure of bank risk, but may provide interesting results as regards the investment decision of banks towards assets considered more or less risky. In particular, we estimate total loans over assets, real estate loans over assets, consumer loans over assets, commercial and industrial loans, and commercial real estate loans. 21

22 Panel A of Table 7 shows the results for the impact of the event on the composition of banks assets. First, we find that affected banks do not change their total loan share after the event relative to unaffected banks, as shown in column (1). Next, taking a closer look at the different loan categories, we find that the composition of loans changed after the introduction of the Emergency Economic Stabilization Act. We see that affected banks have higher shares of consumer loans (column (3)) and real estate commercial loans (column (5)) after the event relative to the unaffected banks. These loan types are considered relatively risky. In particular, consumer loans typically include credit card loans. This may explain an overall lower asset quality and higher bank risk of affected banks relative to unaffected banks. [Table 7] If we standardize the dependent variables using the value of total assets one period before event (Q3 2008), results slightly change. In Panel B of Table 7 we observe that total loans, real estate loans and real estate commercial loans increase after the change in the law relative to the control group. In an additional analysis we standardize total assets with asset values one period before the event. We thereby find that the interacted affect Event Affected is positive and significant. This suggests that banks affected by the raise in deposit insurance became larger after Q than banks that did not benefit from the raise. This is in line with results in Table 7 which shows an increase in the volume of most loan categories but not in relation to banks total assets. 5.3 Robustness tests Cross-section results To guarantee that the results are not driven by serial correlation, a typical problem with difference-in-difference estimation Bertrand et al. (2004) suggest to ignore the time structure of the data. Therefore, we average the data before and after the event and rerun the estimation 22

23 for this collapsed sample. In unreported results, we find that our findings are not driven by serial correlation and observe coefficient of similar magnitude and significance Time-placebo estimation We also want to be sure that our results are not driven by time trends unrelated to the change in insured deposits. Therefore, we conduct a placebo estimation in which we shift the event to Q We then rerun our estimation for observations ±8 quarters before and after this 2005 pseudo event. In unreported results we do not find an effect for the 2005 pseudo event in any of our baseline specifications. This finding supports our assumption that our results are not driven by factors unrelated to increase in deposit insurance per customer. 6 Bank bailouts, bank risk taking and bank capitalization In this section we address two further issues that are important for our analysis. First, we have to deal with the fact that right after the raise in deposit insurance banks were able to apply for TARP under which banks receive capital support from the U.S. government. This may affect our results because both events occur at the same time which makes it difficult to identify the effect from the change in insured deposits alone. Furthermore, the government intervention may distort our results in other ways. For example, banks subject to TARP may also have profited from the raise in insured deposits, i.e., gained more insured deposits. Due to the fact that equity plays a central role for several of our risk measures, a possible outcome of our analysis may be that a raise in insured deposits results in (higher capitalized) saver banks which may only stem from the fact that some banks received government support at the same time. To eliminate these concerns we split our sample in banks that receive TARP and banks that did not receive capital support from the government. Second, we are interested whether the effect from the raise in insured deposits is different for banks that were already riskier in the run-up to the financial crisis. We therefore define a bank to be highly capitalized if its total risk-based capital ratio is above the average median of 14.96% for the period 8 quarters before the event and lowly capitalized, if a bank s capital 23