Deposit insurance scheme and banking crises: a special focus on less developed countries

Transcription

1 Deposit insurance scheme and banking crises: a special focus on less developed countries Anichul Hoque Khan Department of Economics, Simon Fraser University Hasnat Dewan Department of Economics, Thompson Rivers University Abstract This study analyzes the effect of an explicit deposit insurance scheme (EDIS) on the probability of a banking crisis in a country while taking the country s overall economic development into account. The panel data for the period of includes all the countries having the data on an EDIS. The major finding is that if the interaction between a country s overall economic development and its use of an EDIS is not controlled for, the country s use of the EDIS increases the probability of a banking crisis. This increase is greater the more inefficiently designed the EDIS is. However, once the interaction between the overall economic development and the use of the EDIS is controlled for, it is found that the EDIS itself is not a significant factor of a banking crisis. In that case, the less developed the country is that is using an EDIS, the higher the probability of a banking crisis. Keywords Deposit insurance scheme, Banking crises, Less developed countries, Logit model JEL Classification E52, G22, G32 Corresponding address: 8888 University Drive, WMC 3602, Burnaby, BC V5A 1S6, Canada; Tel ; Fax: ; anichul.khan@gmail.com. Corresponding address: 900 McGill Road, Kamloops, BC, V2C 5N3, Canada; Tel ; Fax: ; hdewan@tru.ca.

2 1 Introduction This paper explores the relation between a country s use of deposit insurance scheme and the probability of a banking crisis while taking the country s level of economic development into account. A country s deposit insurance scheme can be explicit, called explicit deposit insurance scheme (EDIS henceforth), or it can be implicit, called implicit deposit insurance scheme (IDIS henceforth). It is explicit in this sense that the rules and regulations are defined in advance to secure the funds deposited in banks up to some limit. And it is implicit because there is nothing explicit but the government usually steps in after any bank failure to rescue the depositors and the banking sector on an ad hoc basis. 1 There is controversy among the researchers in financial economics about the effect of an EDIS on financial market discipline. Many of the researchers believe that the use of an EDIS is an optimal policy to protect a banking sector when its stability is threatened by depositor runs (e.g. Diamond and Dybvig 1983; Bhattacharya et al. 1998). 2 However, some researchers argue that its use may be a source of moral hazard problem, as the banks under the scheme are then encouraged to finance high-risk, high-return projects (e.g. Kane 1989; Laeven 2002; Schich 2008), which may cause additional banks failures, leading to systemic instability in the banking sector resulting in a banking crisis. There has been attempt by Demirguc-Kunt and Detragiache (2002; DKD henceforth) to resolve the controversial impact of an EDIS. 3 Their study finds that the use of an EDIS in general increases the probability of a banking crisis and this probability is greater when the EDIS is designed inefficiently (namely the scheme with voluntary bank membership, government administered scheme, scheme having fund ex ante, scheme with government provided funds, and scheme with a provision of high coverage). Their study also shows that the probability of a banking crisis is even greater if the EDIS s use interacts with the utilizing country s weak institutional environment. 4 Their sample, however, leaves out a large number of less developed countries (LDCs), especially the LDCs with an EDIS (their sample of 61 countries includes only 16 LDCs with an EDIS). 1 See Kyei (1995), Garcia (1996), DKD (1999, 2002) and Demirguc-Kunt et al. (2005). 2 Golembe (1960) finds that the goal of DIS is more focused on restoring confidence in the liquidity of bank deposits than to protect small depositors. 3 Demirgúc-Kunt and Kane (2002) also focus in resolving the ambiguities of the EDIS s impact on banking crises. 4 Other studies such as Cull et al. (2002) and Laeven (2002) also support this finding. 2

3 More than 75% of the countries in the world are classified as LDCs and a growing number of these countries have an EDIS in use. Also, a large number of LDCs have experienced banking crises in the recent years (Caprio and Klingebiel 2003), of which the number of LDCs with IDIS is as large as the number with EDIS. Thus the inclusion of a larger number of LDCs in a sample could result in findings different from those of DKD (2002) concerning an EDIS s impact on a banking crisis. A finding could be that an IDIS contributes to a banking crisis; or if an EDIS is in place, the main reason for a banking crisis is its interaction with the country s inadequate institutional framework. A LDC s institutional framework can be inadequate as a result of a weak government, financial setbacks, unskilled personnel, high corruption, heavy bureaucracy, a weak law and order tradition, unemployment pressure, and so on. In this case, the inclusion of additional LDCs, especially those with EDIS, in a sample could make the estimation results more robust than the results found by DKD (2002). With a large dataset that includes all countries with the information of deposit insurance scheme, i.e. whether it uses an EDIS or an IDIS, this study addresses the following questions. What effect does a country s use of an EDIS have on a banking crisis if the interaction of the EDIS s use with the country s overall economic development is not taken into account? Does this effect differ when the EDIS is inefficiently designed from the moral hazard point of view? What about the EDIS s effect on a banking crisis when the interaction of the EDIS s use with the country s overall economic development is taken into account? Then we compare this interaction effect with the effect of interaction between the EDIS s use and the country s level of corruption, as we recognize that corruption implies merely a single kind of institutional deficiency rather than the structural institutional deficiency. The panel data includes about 150 countries for the period of To obtain the empirical findings of the research questions mentioned above, this study has used a logit estimation model with yearly fixed effects and a set of control variables. The major finding is that if the interaction between the use of an EDIS and the using country s level of economic development is not taken into account, only in that case, as concluded by DKD (2002), the use of the EDIS increases the probability of a banking crisis and this probability is even greater if the EDIS has inefficient design. 3

4 However, once the interaction of the EDIS s use with the country s level of economic development is controlled for, the impact of the EDIS itself is different. Then, the less developed a country is that has instituted an EDIS, the higher the probability of a banking crisis. This probability is even greater if the EDIS provides high coverage by extending the coverage to inter-bank deposits. Interestingly, the EDIS itself is not a significant factor of a banking crisis when the interaction of the EDIS s use with the country s level of economic development is controlled for. It implies that a country s overall economic development is very important for the effective functioning of an EDIS. In addition, this study finds that the ineffective functioning of an EDIS due to inadequate institutional infrastructure of a LDC cannot be reversed by improving only a single indicator of institutional weakness, such as the level of corruption. This is proven by the fact that the use of an EDIS separately is still a significant factor to increase the probability of a banking crisis even after the interaction between the EDIS s use and the using country s level of corruption is controlled for. The remainder of the paper is structured as follows. The next section presents the theoretical background and the hypotheses of this paper. Section three explains the data and the variables. Section four outlines the estimation model, the results and the robustness of the results. Section five concludes. 2 Theoretical background and the hypotheses Friedman (1962) correctly states that a fractional reserve banking system is inherently unstable. A fractional reserve bank cannot withstand sustained depositor runs without any external help. Many countries have established different forms of EDIS to insure bank deposits with the hope of averting a banking crisis. However, incorporating an EDIS into a country s depositors safety net is not without a controversy. According to Diamond and Dybvig (1983) an EDIS entails optimal policy to protect bank stability in an environment where the stability could be threatened by self-fulfilling depositor runs, which leads to bank failures and thus to a banking crisis. Note that a bank failure occurs when a bank is unable to meet its obligations to its depositors or other creditors because it has become insolvent or too illiquid to meet its liabilities, and a banking crisis occurs when a sufficient number of banks fail within a given period and threaten 4

5 the rest of the banks (e.g. González-Hermosillo 1996). 5 However, the contagion effect in banking sector or the depositor runs could be triggered by depositors because of information asymmetry in the financial markets (e.g. Hwang et al. 2009). Following Allen and Gale (1998) it can be established that even if the depositors are capable of gathering necessary information, there could still be depositor runs because of actual deterioration in the quality of assets in banks. In such a case the performance of a bank s management, with an aim of ensuring its assets are of good quality, is important for the protection of the bank from depositor runs. However, a bank failure may occur, in the absence of depositor runs, simply because of high non-performing loans which indicate the insolvency of the bank. Establishing an EDIS does not ensure the cessation of depositor runs. Following González-Hermosillo (1996), the view point can be established that the depositors decisions to withdraw their deposits depend on the coverage limit of their deposits (the lower the coverage, the higher the chance of self-fulfilling runs) and also on the effective endowments which are and will be available in the deposit insurance fund. It also depends on the expected probability of simultaneous failures of a significant number of other banks. While the role of an EDIS to prevent depositor runs is ambiguous, monetary economists find its presence generates a moral hazard problem. With an EDIS s presence, following Demirguc-Kunt and Kane (2002), it is affirmed that the regulators which control the supervisory rudiments can have insufficient incentive to monitor banks and there can be weak enforcement of prudential regulations as well. As a result, the bankers can take high risk in the disbursement of loans. They can even engage in fraudulent activities, such as making loans to themselves or to their close associates at lower than the market price or at weak terms and conditions. The EDIS s presence can cause the moral hazard problem on the side of depositors as well, as the depositors can be less careful in choosing the better performing banks knowing that their deposited funds are insured. 5 Gonzaleze-Hermosillo (1996) follow that in a broad sense a bank failure is said to occur when the regulator recognizes the bank as insolvent and decides to liquidate it, or assist in order to keep it in operation. Different instruments are usually followed to assist an ailing bank. These are, for instance, (1) merger or acquisition of the bank with other healthy banks, (2) direct injections of additional capital or other recapitalization schemes, and (3) different restructuring schemes. The restructuring schemes usually include change in bank-management, assisted generalized rescheduling of loan maturities, and removal of banks' bad loans. 5

6 Thus, as long as a country s institutional structure is not developed enough to check the moral hazard problem caused by an EDIS, it can be assumed that the country s use of the EDIS increases the likelihood of a banking crisis. This increase is even more if the EDIS is inefficiently designed. An EDIS is considered inefficiently designed when it allows higher scope of moral hazard problem (the next section discusses more elaborately the designs of EDIS). In general, an EDIS that allows higher coverage to depositors, generates higher promising insurance to bankers, and is weak from the legal point of view increases the moral hazard problem (e.g. DKD 2002; Demirguc-Kunt and Kane 2002; Demirguc-Kunt et al. 2005; Kyei 1995). The development of the institutional framework of a country is vital to allow an EDIS to work efficiently for bank stability. While the institutional framework in a LDC is insufficiently competent for an EDIS s effective functioning (see Chen and Fishe 1993) 6, it is sufficiently competent in a developed country for the effective functioning of an EDIS (see Gropp and Vesala 2001) 7. The well developed institutional framework can generate enough reliability and the awareness among the depositors about the insurance of their funds under an EDIS and at the same time can check the moral hazard problem caused by bank managements and regulators. It is also important to note that the bank managements in a LDC can be more aware about the scope of misusing an EDIS than the awareness of the depositors about the pros and cons of having the EDIS. Thus, the more a country is economically developed, the more competent its institutional framework and the better its institutional performance, then the greater the confidence the general people (or depositors) will have about the reliability of the institutional performance. The tendency of moral hazard or fraudulent activities of the people in a developed country is also lower because of state supported social programs for its citizens even in cases of economic shock. In addition, immoral activities by any banker or regulator in a developed country is less likely as s/he knows it will be difficult to be successful ex post because of non-cooperation from the other related officials and institutions (e.g. court, police). 6 Chen and Fishe (1993) show EDIS s use to be ineffective in LDCs. 7 Gropp and Vesala (2001) show EDIS s use to be relatively more effective in EU countries. 6

7 The hypotheses for empirical examination Based on the above discussions, we formulate the following hypotheses for the empirical examination: 1 Ceteris Paribus, a banking crisis is more likely in a country with an EDIS than with an IDIS, when the interaction of the country s overall economic development with the EDIS is not taken into account. The likelihood of the crisis is even greater the more inefficiently designed the EDIS is. 2 Ceteris Paribus, the less developed is a country that has adopted an EDIS, the more likely a banking crisis. Once the interaction between the level of economic development and the EDIS is controlled for, the EDIS itself is no longer a significant factor of banking crises. 3 The data 3.1 The data of deposit insurance schemes We use Demirguc-Kunt et al. (2005) for the data source of deposit insurance schemes. They built their database from the earlier studies by Demirguc-Kunt and Sobaci (2001), Garcia (1999), Kyei (1995) and Talley and Mas (1990). Our database is further complemented and improved with the data from various countries and online sources as well as based on the survey by Barth et al. (2004). The cross-country data of using an IDIS or an EDIS, including the designs of EDIS, are reported in Table A1 and Table A3 of the appendix. EDIS and IDIS An increasing number of countries have been adopting an EDIS since the 1980s. According to the International Association of Deposit Insurers (IADI), as of January 2010, 106 countries have instituted some form of EDIS. 8 We assume that a country uses an IDIS, if it has not instituted an EDIS, because the country establishes a de facto insurance system for the depositors and the banks ex post to any bank failures (see Kyei 1995; DKD 1999, 2002; Demirguc-Kunt et al. 2005). The proportions in our dataset between the countries with an EDIS and the countries with an IDIS and between the high-income countries and the middle-tolow income countries are very different from those in the dataset used by DKD 7

8 (2002). In Table A1 of the appendix there are 150 countries, 44% of which have instituted an EDIS and about 40% of the countries with an EDIS belong to the high income group. These figures are much different in DKD (2002), in which 60% of the countries of their sample of 61 countries have an EDIS, and about 55% of the countries with an EDIS belong to the high income group. A simple dummy variable called IDIS-EDIS is used in our regression models to examine the effect of an EDIS on banking crises. The variable takes the value 0 for the country-period with IDIS and 1 for that with EDIS. We hypothesize the variable to be positively associated with the probability of a banking crisis as long as a variable representing the interactions between the EDIS s uses and the using countries levels of economic development is not controlled for. An EDIS can be formed (or designed) in many different ways. However, we consider only the forms which were studied by DKD (2002) in order to keep the scope of this study within a reasonable limit. 9 Broadly those are the regulation about an EDIS s administration and funding, banks membership and premium in the scheme, and about the scheme s coverage to depositors. These different forms of EDIS and the corresponding variables in our models are defined below: EDIS s administration: An EDIS can be administered solely by private banks, or jointly by banks and the government, or exclusively by the government. In a privately administered system a competitive market premium is expected to be set optimally and execution of the closure rule is also expected to be appropriate (Carisano 1992). On the other hand, a government administered scheme allows a moral hazard problem on the part of banks (Demirguc-Kunt and Kane 2002; DKD 2002) as it implicitly provides higher promising insurance to the bankers but with relatively weaker bank monitoring and inappropriate bank closure rules. In our dataset 12 schemes are administered privately, 18 jointly, and 34 by the government. The variable called EDIS s administration takes 0 for IDIS, 1 for EDIS with private administration, 2 for joint administration, and 3 for governmental administration. Banks membership: A bank s membership in an EDIS can be compulsory or voluntary. If the membership is voluntary, there can be an adverse selection 8 Retrieved from on 8/8/

9 problem as the sounder banks may opt out of the system (see Demirguc-Kunt and Kane 2002; DKD 2002; Garcia 1999). Thus, the EDIS with voluntary bank membership can increase the banking sector s instability. In our dataset there are only 6 countries with an EDIS where bank membership is voluntary. The variable called Banks membership takes the value 0 for IDIS, 1 for EDIS with compulsory membership, and 2 for voluntary membership. EDIS s funding: An EDIS can be funded ex ante or unfunded but funding is ex post to any bank s failure. In a funded scheme banks pay premiums into a recognized fund. In an unfunded scheme either the government reimburses depositors of the failed bank, or the insured financial institutions are called upon to contribute (see Kyei 1995). However, a scheme with ex ante funding makes the scheme more promising against any contingency and hence it widens the scope of moral hazard problem on the part of banks as they take excessive risks in disbursing loans (DKD 2002; Demirguc-Kunt and Kane 2002). In our dataset 11 countries with an EDIS are funded ex post and 55 funded ex ante. The variable called EDIS s funding takes the value 0 for IDIS, 1 for EDIS with funding ex post, and 2 for EDIS with funding ex ante. EDIS s fund-source: Funding for a scheme can come from three sources exclusively from banks, jointly from banks and the government, and exclusively from the government. The best option for a scheme is to collect funds exclusively from banks (see DKD 2002). Then the moral hazard problem on the part of banks is minimized as the banks themselves pay for their own actions. In our dataset the funding source of EDIS is banks in 19 countries, it is joint in 43 countries, and it is the government in only one country. The variable called EDIS s fund-source assigns the value 0 for IDIS, 1 for EDIS funded by banks, 2 for it joint and 3 for the government funded scheme. Banks premium: The premium paid by a bank in an EDIS could be risk adjusted or fixed. The bank paying premium based on its risk-taking is efficient because it reduces the moral hazard problem on the part of the bank (see Freixas and Rochet 1997, p. 267; Kyei 1995; Garcia 1999; DKD 2002). On the other hand, if the premium is fixed for the bank, the bank takes more 9 The reason is that this study is an extended version of DKD (2002), but with many new findings. 9

10 risk in its business, such as in the allocation of loans. Our dataset has 14 countries with an EDIS where the premium system is risk-based and it is fixed in 51 countries. The variable called Banks premium takes the value 0 for IDIS, 1 for EDIS with risk-based premium, and 2 for that with fixed premium. Low coverage is the best option for an EDIS (e.g. Demirguc-Kunt and Kane 2002; DKD 2002; Garcia 1999), since higher coverage in a scheme allows for a greater scope of moral hazard on the part of banks. The higher the coverage to deposits, the higher the amounts insured, and thus the higher the scope for banks to take excessive risks in dealing with loans. 10 The coverage of an EDIS can be extended in the following forms: Foreign currency deposits covered: An EDIS can increase its coverage by providing coverage to foreign currency deposits. In our dataset 22 countries with an EDIS extend no coverage to foreign currency deposits while 42 do cover foreign currency deposits. The variable called Foreign currency deposits covered assigns the value 0 for IDIS, 1 for EDIS without the coverage to foreign currency deposits, and 2 for those with the coverage. Inter-bank deposits covered: An EDIS can extend its coverage to inter-bank deposits. In our dataset there are 51 countries with an EDIS without the coverage of inter-bank deposits and 13 with the coverage. The variable called Interbank deposits covered assigns the value 0 for IDIS, 1 for EDIS without the coverage of inter-bank deposits, and 2 for those with the coverage. Coverage with co-insurance: An EDIS can increase its coverage by assigning co-insurance in addition to the original coverage. In our dataset there are 50 countries with an EDIS that have not assigned the co-insurance and 15 that have assigned. The variable called Coverage with co-insurance takes the value 0 for IDIS, 1 for EDIS without the co-insurance, and 2 for those with co-insurance. 10 Thiratanapong (2007) shows that depositors responsiveness to bank risk taking has gone up since the 1997 crisis in Thailand, but a blanket coverage by the government in the aftermath of the crisis has to some extent weakened market discipline. 10

11 Average coverage of deposits to GDP-per-capita ratio: As a rule of thumb, the IMF typically proposes an amount equivalent to 1 to 2 times of GDP per capita as an appropriate coverage limit for an EDIS (Garcia 1999). However, it has been noticed in our dataset that there are 31 countries with an EDIS in which the ratio is below 2 and in another 31 the ratio is more than that. We take the average ratio for the latest 5 years available if the adoption date of the EDIS is older than 5 years, otherwise the average from the date of adoption. 11 The variable called Average coverage / GDP-per-capita is not in fact a dummy variable as it takes the value 0 for IDIS, but the remaining values are the ratios of deposits under coverage to GDP per capita. Average coverage of deposits to deposit-per-capita ratio: Another way of measuring an EDIS s coverage limit is the ratio of average coverage of deposits to deposit-per-capita. We include this variable in our model as well because the ratio in terms of deposits per capita may be different from the ratio in terms of GDP per capita. Of the available data, there are 31 countries with an EDIS in which the average ratio is below 5 and in 29 countries the ratio exceeds 5. Similar to the Average coverage / GDP-per-capita variable, we take the average ratio of the latest 5 years available if the adoption date of the EDIS is older than 5 years, otherwise the average is from the date of adoption. 12 The variable called Average coverage / deposit-per-capita is also not a dummy variable as it takes the value 0 for IDIS and the remaining values are the said coverage ratios. We hypothesize each of the above-mentioned EDIS design variables to be more strongly positively related with the probability of a banking crisis than the positive relation of the simple IDIS-EDIS dummy. This hypothesis is affirmed as long as a variable representing the interactions between the EDIS s uses and the using countries different levels of economic development is not controlled for. 11 Note that we drop Mexico as an outlier from the sample when we use this variable in the regression. The reason is that the coverage ratio for Mexico is extremely high compared to other countries and the estimation result for this variable changes substantially if Mexico is included. 12 For the similar reason mentioned in the earlier footnote, Mexico is again dropped from the sample as an outlier. 11

12 3.2 Sample selection and the other variables The period covered in our sample is To select the sample countries, we begin by taking all the countries enlisted by the IMF in their IFS (International Financial Statistics) data series. Then we drop the countries and/or periods out of the sample because either they are centrally planned or socialist states, 13 subservient states, 14 states affected by civil war, 15 or because the data is missing on the necessary control variables. 16 Furthermore, parts of the study period for some countries are dropped because of their transitional state 17 and because of non-systemic banking crises 18 (see Table A1 and A2 of the appendix). To be robust we use the data of an EDIS, including its designs, from the date it was first enacted if not revised, otherwise from the date of its latest revision. For instance, the EDIS of Finland was most recently revised in 1998, and thus we use the data of Finland since Note that about one-third of the countries reported in Table A1 revised the designs of their EDISs after they were first instituted, but unfortunately the data do not offer much information on the revisions undertaken. Defining banking crises The data on banking crises come from Caprio and Klingebiel (2003). 19 A banking crisis is said to occur in a country when a sufficient number of banks fail within a given period and threatens the rest of the banks in the country. 20 Studies 21 find that a banking crisis is difficult to identify because of limited information. For instance, many banking crises are accompanied by depositor runs, but the potential for depositor runs cannot be observed directly. On the other hand, banking crises in recent years are associated not only with depositor runs, but also 13 China, Cuba, Laos, and Vietnam are dropped because of their centrally planned economies. 14 Hong Kong, Micronesia, and Taiwan are not completely sovereign states yet. 15 Afghanistan, Somalia, and Sudan have been affected by civil conflicts throughout most of our study period. 16 Albania, Azerbaijan, Belarus, Brunei Darussalam, Isle of Man, Kiribati, Liechtenstein, Marshal Islands, and Serbia Montenegro are dropped due to lack of data on the specified control variables. 17 We exclude only the first two years for transitional economies when they transitioned from the centrally planned economy to the market economy. 18 Periods of non-systemic banking crises are dropped because: 1) a non-systemic crisis may affect the systemic crisis and the explanatory variables in our regressions; 2) we are interested in estimating the probability of a systemic banking crisis given that the economy is otherwise in tranquil period. 19 There could be small scale (or non-systemic) banking crises and systemic banking crises. The banking crises, discussed in this paper, are systemic banking crises unless otherwise specified. 20 See Gonzalez-Hermosillo (1996) for details. 21 See Glick and Hutchison (1999). 12

13 with deterioration in asset quality and subsequent government intervention. However, the availability of such information is also limited. Due to information and conceptual limitations, most studies have employed a combination of events to identify and date the occurrences of banking crises. Following Caprio and Klingebiel (2003), we find that at least one of the following conditions exists in a systematic banking crisis: 1 The ratio of non-performing loans to total assets in the banking system is at least 10 percent. 2 The amount of assets of insolvent banks is at least 10 percent of the amount of total assets of all banks. 3 The cost of rescue operation was at least 2.5 percent of the GDP. 4 Banking sector problems resulted in government intervention with measures such as large-scale nationalization of banks, or prolonged bank holidays, or bank closures, or mergers of a large number of banks. The variable called Bank crisis takes the value 0 when an economy is in a tranquil state and 1 when it experiences a systemic banking crisis. The years during which the systemic crises occurred are excluded from the sample, because the crises themselves may affect the behaviour of the explanatory variables in our model. Economic development One of the most important explanatory factors for this study is the factor representing countries levels of overall economic development. The less developed a country is economically, the more inadequate the country s institutional structure, and thus the poorer the performance of the institutions. This is due to weaker access to modern technology, higher share of unskilled personnel, greater limitation of financial capabilities, higher level of corruption and bureaucracy, lower standard of law and order tradition, higher unemployment pressure, lower level of overall consciousness among the people because of lower education level, weaker government to resist some domestic or international pressures to implement often some inefficient measures, and so on in a LDC. When the overall institutional efficiency is weak, one cannot expect top-quality business practices from a bank with its available funds, nor can one expect high returns from the projects financed by bank loans. Thus the probability of a 13

14 banking crisis because of higher non-performing loans is relatively higher in a country with lower economic development regardless of the country s adoption of an IDIS or an EDIS. A variable called Less development is generated to examine the effect of the level of overall economic development on banking crises. The variable assigns the values of 1, 2, 3, and 4 for high income, upper-middle income, lower-middle income, and low income countries respectively. Thus all the developed nations belong to the group of high income and all the LDCs belong to the groups of upper-middle to low income countries. Countries with their respective groups are reported in Table A1 (data source is Demirguc-Kunt et al. 2005). What happens if a country s use of an EDIS is interacted with the country s low economic development? As already discussed, it is possible that the EDIS increases the probability of a banking crisis because it is operated in a LDC, where the EDIS interacts with the weak institutional structure, widening the scope of moral hazard on the part of banks. To examine the effects of such interactions on banking crises, we generate the interaction variables by multiplying each of the variables for EDIS with the variable for countries levels of economic development. We also use cross-country time series indexes of corruption and generate - similar kind of interaction variables with each of the variables for EDIS. This is done to examine how much a country s level of corruption s interaction with an EDIS, in comparison to the country s level of development s interaction, can explain the probability of a banking crisis. Corruption indexes range from zero to six and increase with the quality of institutions, i.e., the higher the index value, the lower the level of corruption. The data source of corruption index is the International Country Risk Guide. It should be noted that we have the data until 1997, after which we apply the 1997 s corruption index to the years 1998 through Other control variables When testing the effect of an EDIS on banking crises, we use a set of control variables which are similar to that of DKD (2002). Expected signs of the variables impacts on banking crises and their sources are reported in Table 1. 14

15 In addition, we use a variable called Remains of the past crises to take into account the impact of the past crises. One theory suggests that the farther a crisis is from its first occurrence, the more properly the crisis is addressed, the lower the residue of the crisis, which effectively reduces the recurrence of a crisis. A contrary view is that the farther the crisis is from its first occurrence, the lower the residue of the crisis, the less careful the central authority is about the prudential regulation of the banking sector and about the moral hazard problem on the side of bank managements, the greater the risks the bank managements take again in disbursing loans which consequently increases the recurrence of a banking crisis. The latter view is in line with Kindleberger (2000, pp. 13) that states: some time must elapse after one speculative mania that ends in crisis before investors have recovered sufficiently from their losses and disillusionment to be willing to take a flyer again. Explanatory variables Table 1. Expected signs of the control variables on banking crises Expected signs References and short explanations GDP growth - Kaminsky and Reinhart (1999) and Gorton (1988): cyclical output downturns increase banking sector problems; DKD (1999, 2002, 2005): -ve relation with occurrences and costs of BC Terms-oftrade change - Kaminsky and Reinhart (1999) and Gorton (1988): terms of trade deteriorations increase banking sector problems; DKD (1999): -ve relation with occurrences and costs of BC Inflation + Obstfeld (1986): 2 nd generation theory; Komulainen and Lukkarila (2003): +ve relation with CC; DKD (1999, 2005): +ve relation with occurrences and costs of BC. Depreciation + DKD (1999, 2002): +ve relation with occurrences and costs of BC. Real interest + DKD (1999, 2002, 2005): +ve relation with occurrences and costs of BC. M2/reserves + Calvo (1998): high ratio causes vulnerability; DKD (1999, 2002, 2005): +ve relation with occurrences and costs of BC; Komulainen and Lukkarila (2003): + ve relation with CC Credit + DKD (2002, 2005): +ve relation with occurrences of BC growth t-2 GDP/Capita - DKD (1999, 2005): -ve relation with occurrences of BC NB: GDP = gross domestic production; BC refers to banking crises and CC to currency crises ; -ve = negative, +ve = positive. Thus, the effect of the variable Remains of the past crises on a new banking crisis is theoretically inconclusive. Note that the past crises considered are the crises taken place within and around the 1980s and the 1990s. All the control variables, their formulations and data sources are reported in Table A4 of the appendix. 22 As can be seen from the data series on corruption, the corruption index of a country does not change that frequently. 15

16 4. Empirical evidence 4.1 Estimation model A logit estimation model is used. The following model is fitted to estimate probability: Prob(y = 1 x) = e 1 e β x β x F( x) The dependent variable is banking crisis dummy y, and y = 1 for a country in a year if the country experiences a banking crisis in that year, otherwise y = 0. Vector x is the set of explanatory variables we have discussed in the previous section. The parameter vector reflects the impacts of changes in x on the probability. The notation F(.) is the cumulative logistic distribution. The log-likelihood function for the logit model can be written as j S lnl lnf( b) ln 1 F( x b) x j j, j S where S is the set of all country-years j such that y j = 1, i.e. country-years with banking crises. Thus the set of j not belonging to S implies the set of all country-years not-experiencing banking crises. We take the robust standard errors (see Huber 1967; White 1980), according to which the standard errors are adjusted for countries clustering. We include yearly fixed effects by taking yearly dummies in the regression models to allow the possibility that the probability of a banking crisis may change cross-year independently of the explanatory variables. Country fixed effects could also be included, but the inclusion of country fixed effects in the logit estimation model would require omitting all the countries from the panel that did not experience banking crises. Since many of the countries in our dataset did not experience any banking crisis during the study period and thus would be omitted from the panel, we do not estimate the country-fixed effects. When interpreting the regression results it is important to remember that an estimated coefficient does not indicate the direct increase in the probability of a banking crisis given a one-unit increase in the corresponding explanatory variable. Instead, the coefficient reflects the effect of the change in the explanatory variable 16

17 on the probability function as given in the above expression. However, the sign of the coefficient does indicate the direction of the change. 4.2 Estimation results In our regressions, the sample size is largest when we use the simple IDIS-EDIS dummy to estimate the effects of EDIS. The sample sizes are smaller and unbalanced when the EDIS-design variables are used, because of the variables missing values. Therefore, we also estimate the effects of EDIS by using the IDIS-EDIS dummy with the same samples that are used to estimate the effects of different EDIS-designs. The reason for this estimation is to determine whether the estimated coefficient of an EDIS s design is more precise than the coefficient of the simple EDIS. The estimation results of the simple IDIS-EDIS dummy used in different regression models with different sample sizes are reported in the corresponding columns at the bottom of the result-tables. These results are distinguished with bold and italic style (due to space limitation, the coefficients of the control variables are not reported). The quality of a regression model is assessed by the following values: (1) loglikelihood value the higher the value, the better the model; (2) Wald Chisquared value the higher the value, the better the model; (3) joint significance level (Prob > Chi-square) the lower the value, the stronger the model s joint significance level; (4) model s goodness of fit (Pseudo R-squared value) the higher the value, the better the model; and (5) Akaike s Information Criterion (AIC) 23 the lower the value, the better the model. We also report the values of the estimation models overall prediction accuracy and the prediction accuracy of crises EDIS and banking crises At first we estimate the effect of an EDIS controlling the other necessary variables including the yearly dummies, but excluding the variable of interaction between the EDIS and the countries levels of economic development. The estimation results are reported in Table 2 (results of the yearly dummies are not reported 23 In the general case, AIC = 2k 2ln(L), where k is the number of parameters in the statistical model and L is the maximized value of the likelihood function for the estimated model. 24 We consider a crisis (no-crisis) correctly estimated if the predicted value of the crisis (no-crisis) more (less) than the crude probability of crisis, i.e. the number of crises divided by the total number of observations. In order to get the overall prediction accuracy, we consider both crises and no-crises estimated correctly. 17

18 lacking enough space). Many of the observations are lost because of the missing data of the control variables. A few observations are dropped since they are outliers with respect to the parameters of some explanatory variables used (see Table A5 of the appendix). The number of countries included in the samples for different regression models ranges from 98 to 104 (see Table A6 of the appendix for the lists of the countries used). IDIS-EDIS EDIS s administration Table 2. Estimating the effect of an EDIS on banking crises Dependent var. = Bank crisis (1) (2) (3) (4) (5) (6) GDP growth *** (0.048) *** (0.048) *** (0.047) ** (0.050) *** (0.048) *** (0.048) Inflation * (0.0071) * (0.0071) * (0.0071) (0.0067) * (0.0070) * (0.0072) Terms-of-trade change (0.0076) (0.0077) (0.0078) (0.0077) (0.0076) (0.0077) Depreciation ** (0.0068) ** (0.0068) ** (0.0068) * (0.0066) ** (0.0067) ** (0.0069) Real interest (0.0010) (0.0010) (0.0010) (0.0009) (0.0010) (0.0011) M2/reserves (0.0038) (0.0038) (0.0032) (0.0036) (0.0042) (0.0031) Credit growth t-2 (0.0087) (0.0094) (0.0088) (0.0100) (0.0086) (0.0092) Less development *** (0.1358) *** (0.1309) *** (0.1316) *** (0.1357) *** (0.1320) *** (0.1308) Remains of the past crises * (2.8622) (3.1156) (2.9256) (3.3213) * (2.8832) (2.9878) 1.091*** (0.5%) (0.4107) ** (0.1454) ** (0.3061) Banks membership EDIS s fund-source ** (0.2114) EDIS s funding Banks premium Constant *** (1.127) *** (1.119) *** (1.065) *** (1.106) *** (0.2063) *** (1.120) * (0.2171) *** (1.126) obs Log likelihood Wald chi Prob > chi Pseudo R AIC No. of crises % overall correct % crises correct ** 2% 1.011** 2% 0.939** 3% 1.091*** 1.011** 2% IDIS-EDIS (0.419) (0.419) (0.428) (0.4107) (0.419) Log likelihood Pseudo R

19 (Table 2 continued) Dependent variable = Bank crisis (7) (8) (9) (10) (11) GDP growth *** (0.049) *** (0.050) *** (0.049) *** (0.049) *** (0.048) Inflation (0.0067) (0.0068) (0.0067) (0.0064) (0.0063) Terms-of-trade change (0.0074) (0.0075) (0.0074) (0.0077) (0.0077) Depreciation * (0.0065) * (0.0066) * (0.0066) ** (0.0064) ** (0.0063) Real interest (0.0009) (0.0010) (0.0009) (0.0009) (0.0009) M2/reserves (0.0035) (0.0042) (0.0029) (0.0034) (0.0030) Credit growth t-2 (0.0099) (0.0097) (0.0094) (0.0108) (0.0106) Less development *** (0.1380) *** (0.1397) *** (0.1330) *** (0.1317) ** (0.1337) Remains of the past crises * (2.8438) (3.2727) (2.9837) (3.3412) (3.5592) Foreign currency deposits covered ** (0.2196) Interbank deposits covered *** (0.2497) Coverage with coinsurance ** (0.2704) Average coverage/ GDP-per-capita *** (0.0787) Average coverage/ deposit-per-capita *** (0.0181) Constant *** (1.152) *** (1.160) *** (1.116) *** (1.163) *** (1.167) obs Log likelihood Wald chi Prob > chi Pseudo R AIC No. of crises % overall correct % crises correct ** 2% 1.026** 2% 1.113*** 0.940** 3% 0.944** 3% IDIS-EDIS (0.418) (0.418) (0.415) (0.428) (0.427) Log likelihood Pseudo R Numbers in the parentheses are standard errors; ***, **, and * refer to 1%, 5%, and 10% level of significance. The parameter of IDIS-EDIS dummy is strongly significant with the expected positive sign. That is, the probability of a banking crisis is higher in a country with an EDIS than the country with an IDIS. Note that we do not control the EDIS s interaction with the country s level of overall economic development in this model. This finding is already confirmed in the previous empirical studies (e.g. DKD 1999, 2002, 2005). However, the level of significance of an EDIS s effect is much stronger in our study (significant at 0.5% level) than those found in the previous 19

20 studies (e.g. in DKD 2002 it is significant at 8% level). The stronger significance level could be due to our different dataset as mentioned in the first section. It could also be due to our different estimation model as we take the yearly fixed effects and a slightly different set of control variables. The predicted probabilities of crises estimated for the Philippines (1998), Thailand (1997) and Japan (1991), for example, would have decreased from 9.85%, 27.72% and 4.01% to 4.53%, 13.98% and 1.86% respectively if these countries had switched from EDIS to IDIS. In order to infer that the probability of a banking crisis is even higher for any inefficient design of the EDIS, its coefficient should be estimated more precisely than the coefficient of the IDIS-EDIS dummy. The sign has to be positive as well. The coefficient in comparison with that of the simple IDIS-EDIS dummy is considered to be more precisely estimated if (1) the level of significance of its estimation is stronger, (2) the log-likelihood value of the model is higher, and (3) the model s goodness of fit is better (higher pseudo-r-squared value). Accordingly, the more precisely estimated coefficients of EDIS-designs and their interpretations are as follows. EDIS s funding: The probability of a banking crisis in a country is higher if the country uses an EDIS which is funded ex ante than if it is funded ex post. This result is established when the interaction between the country s overall economic development and the EDIS is not controlled for. Thus, this study supports the viewpoint developed by Demirguc-Kunt and Kane (2002) and DKD (2002) that an EDIS s ex post funding is better than the ex ante one, but as long as the said interaction is not considered. The predicted probabilities of crises for South Korea (1997) and the Philippines (1998), for example, would have decreased from 8.71% and 10.40% to 5.10% and 6.26% respectively if these countries had switched from ex ante funding scheme to ex post one. Inter-bank deposits covered: Our study confirms that when the interaction of a country s level of economic development and its use of an EDIS is not considered, the probability of a banking crisis in the country is higher when the EDIS provides coverage to the inter-bank deposits than when it does not. Like the parameter of the simple IDIS-EDIS dummy, this parameter is also more strongly significant in our study (significant at 1% level) than it is in 20

21 DKD (2002; significant at 10% level). When checked, we find that this stronger significance level is due partly to our different dataset, partly to our different estimation model, and partly to our slightly different set of control variables. The predicted probabilities of crises for the Philippines (1998) and Thailand (1997), for example, would have decreased from 13.04% and 34.73% to 6.94% and 20.36% respectively if these countries had chosen not to offer coverage to inter-bank deposits. Average coverage of deposits: The parameters of both of the variables Average coverage/ GDP-per-capita and Average coverage/ deposits-percapita suggest that the probability of a banking crisis increases in a country with its increase of average coverage to deposits. This result only holds if the interaction of a country s level of economic development and its use of an EDIS is not taken into account. This finding is similar with DKD (2002) who found that the probability of a banking crisis increases with the increase of explicit coverage limits to deposits. The predicted probabilities of crises for South Korea (1997) and Turkey (2000), for example, would have decreased from 6.28% and 7.67% to 3.98% and 0.92% respectively if the ratios of the average coverage of deposits to GDP-per-capita in these countries had declined from 3.86 and to 2. The above-mentioned findings remain the same when we test the regression models of Table 2 by taking (1) only the control variables that are significant, or (2) only the control variables that are significant with the least number of missing values. In the first case, the control variables are GDP growth, Depreciation, and Less development; the numbers of observations and the numbers of countries range from 1240 to 1321 and from 109 to 110 respectively. In the second case, the control variables are Depreciation and Less development; the numbers of observations and the numbers of countries range from 1866 to 1992 and from 148 to 149 respectively. The findings remain the same when we add corruption, a measure of institutional environment, as a separate control variable in the models of Table 2 (the numbers of observations then become much smaller ranging from 726 to 771). The results of these additional tests, however, are not reported in this paper but can be obtained upon request from the authors. 21