Determinants of Ex-Ante Banking System Distress: A Macro-Micro Empirical Exploration of Some Recent Episodes
Author:
Ms. Brenda Gonzalez-Hermosillo
Search for other papers by Ms. Brenda Gonzalez-Hermosillo in
Current site
Google Scholar Close

Contributor Notes

Author’s E-Mail Address: bgonzalez@imf.org
Publication Date:
01 Mar 1999
eISBN:
9781451845167
Language:
English
Keywords:
WP; coverage ratio; bank distress; performing loan; equity capital; bank fragility; Odds ratio; balance sheet data; insider loan; hearding behavior
Download PDF (4.7 MB)

This paper empirically analyzes the contribution of microeconomic and macroeconomic factors in five recent episodes of banking system problems in the U.S. Southwest (1986–92), Northeast (1991–92), and California (1992–93); Mexico (1994–95); and Colombia (1982–87). The paper finds that a low capital equity and reserve coverage of problem loans ratio is a leading indicator of bank distress, signaling a high likelihood of near-term failure. Distress is shown to be a function of the same fundamental macro-micro sources of risk that determine bank failures. Focusing on distress has the advantage that the fragility of the banking system can be assessed before a crisis actually occurs.

Abstract

This paper empirically analyzes the contribution of microeconomic and macroeconomic factors in five recent episodes of banking system problems in the U.S. Southwest (1986–92), Northeast (1991–92), and California (1992–93); Mexico (1994–95); and Colombia (1982–87). The paper finds that a low capital equity and reserve coverage of problem loans ratio is a leading indicator of bank distress, signaling a high likelihood of near-term failure. Distress is shown to be a function of the same fundamental macro-micro sources of risk that determine bank failures. Focusing on distress has the advantage that the fragility of the banking system can be assessed before a crisis actually occurs.

I. Introduction

Although financial crisis have a long history, in the past two decades many countries have experienced episodes of significant financial sector distress.2 Several recent periods of banking system distress have been associated with currency crises. Perhaps the most acute among the recent experiences are the financial problems encountered by some emerging markets. The current banking system problems that begun in the mid-1990s in Asia (including Thailand, Indonesia, and Korea) have also made apparent the possibility of regional contagion. In Latin America, severe banking crises also occurred in Mexico, Argentina and Venezuela in the first half of the 1990s, and in Chile and Colombia in the 1980s. But banking crises are not events reserved only for emerging economies. Episodes of significant banking system distress have been evident in Japan since the mid-1990s, in the Nordic countries during the early 1990s, and in the United States during the mid-1980s and early 1990s. Although with less intensity, banking problems have also recently afflicted countries like France and Italy.

These recent events have led to an explosion in the past few years in the number of studies that have taken different angles to try to explain the factors that contribute to financial crises. Clearly, finding basic early-warning systems of banking crisis and understanding its dynamics are critical, particularly in the current context of financial globalization in which countries can be affected by the financial problems in other countries. The current empirical literature seems to be largely divided into two camps: studies which primarily focus on the role of bank-specific data (largely in the context of CAMEL variables)3 to explain bank failures, and studies focusing on the contribution of macroeconomic variables to explain banking crises.

Despite the clear advances in the literature of banking failures, many issues remain to be resolved. For example, why is it that, despite the fact that all banks in a country are hit by the same macroeconomic shock, not all banks fail? Do banks that fail have different characteristics than non-failed banks? If so, are some of those characteristics different several periods before actual failure? Are there some indicators that act akin to a pressure gauge through which near term failure is being signaled? Can these latter indicators be used to assess the degree of distress in the banking system before the crisis actually occurs? How to measure “moral hazard”? How does bank contagion occur? Are banking crises fundamentally different in advanced economies vis-à-vis developing countries?

This paper attempts to contribute to the literature of banking failures and banking system crises by analyzing the role of both micro and macro factors in the context of a simple model of bank failures based on market, credit and liquidity risks. By focusing on these notions of risk, rather than on actual definitions of variables, the analysis of different episodes of banking problems can be made broadly comparable despite the fact that specific circumstances and accounting systems usually vary among countries. The paper also attempts to examine some potential sources of contagion. By analyzing the experiences of five recent episodes of banking system distress in an advanced economy and in some developing countries—specifically, in the United States: the Southwest (1986–92), the Northeast (1991–92), and California (1992–93); in Mexico (1994–95); and in Colombia (1982–87)—it proposes a basic method by which to asses the ex-ante fragility of the banking system, before the bank failures actually occur. The main objective of this paper is to find parsimoniously the common factors across these different episodes of banking system distress.

The paper is organized as follows. Section II critically reviews the recent advances in the literature of bank failures and banking crises, and discusses the main issues still unresolved. Section III reviews the macroeconomic background for the episodes of banking system distress examined. It also examines the micro characteristics of banks that failed vs. banks that did not fail. Section IV discusses the basic theoretical framework in which bank fragility is a function of market, credit and liquidity risks (which can be associated with macroeconomic factors) and can be influenced by contagion factors. The different stages in the typical life cycle of bank failures and the main variables examined are also discussed. Section V briefly reviews the empirical methodology used to estimate the probability of failure and the time of failure. The empirical results for each episode are discussed in Section VI. The section begins by proposing an ex-ante measure of bank distress based on the ratio of capital equity plus loan reserves minus nonperforming loans to total assets (coverage ratio). The results suggest that the framework proposed captures reasonably well the episodes of bank fragility in all the cases examined. For completeness, a simple traditional CAMEL model with and without capital and nonperforming loans (the latter is a typical proxy for asset quality) was estimated. The results from this exercise suggest that capital and nonperforming loans account for a significant amount of the explanatory power in traditional CAMEL models. Based on the different models, the predicted values for each individual bank are then aggregated (and weighted by their relative asset shares) to arrive at estimates of fragility for the overall banking system—both based on (ex-post) failure and (ex-ante) distress. Section VII concludes and reviews some of the empirical regularities across regions that most clearly emerge from the analysis. In particular, market risk and liquidity risk were generally important in determining bank distress and eventual failure. Contagion seemed to be present in some cases but its impact was usually small. In addition, a high ratio of nonperforming loans to total assets and a low ratio of capital to total assets seemed to be strong indicators of bank distress.

II. Review of the Literature and Current Issues

Although the literature on bank failures and banking crises is indeed extensive, there seem to be two separate broad streams in the empirical literature: the “micro” and the “macro” camps.4 First, the “micro” approach typically focuses on individual banks’ balance sheet data, possibly augmented with equity price data, to predict bank failures. This literature was particularly prevalent in the 1980s, but many studies have also been published in the 1990s. Most studies have been in the context of U.S. banks. Typically, these studies use different empirical methodologies and financial ratios to produce an evaluation of the condition of banks consistent with the CAMEL rating system often used on-site by regulators. The list of variables that has been suggested is quite extensive and often involves complex accounting. (Appendix I provides a summary of the variables used in some representative studies.) A few of those studies have introduced some macroeconomic/regional factors as explanatory variables, but usually without a formal link to the rest of the analytical framework.

The second stream in the empirical literature, one which has seen significant growth in recent years, focuses on macroeconomic (sometimes, including institutional) variables to explain banking crises.5 These studies typically focus on a large sample of countries which are known ex-post to have had a banking crisis during a certain period. The macroeconomic factors associated with episodes of banking sector problems highlighted in this literature include cyclical output downturns, adverse terms of trade shocks, declines in asset prices (e.g., in equity and real estate markets), rising real interest rates, boom-bust cycles in inflation, credit expansion, losses of foreign exchange reserves, and capital inflows. Interestingly, some of the studies that have recently attempted to examine the macroeconomic causes (only) of the current banking crises in Asian countries find that these models would have largely missed these crises (Demirgüç-Kunt and Detragiache (1998b), and Hardy and Pazarbasioglu (1998)).

Often, the argument by the “macro” camp is that, while it would be useful to include individual bank data in the analysis, bank balance sheet data are typically difficult to obtain as no single source of data exists to date. Furthermore, such data are often based on complex accounting principles that vary among countries. These factors can make cross-country comparisons difficult, particularly if some of the typical variables used in most (full-fledged) -CAMEL studies were to be constructed for cross-country comparisons.6

What is needed? First, an integrated approach of the micro” and “macro” camps. The macro approach has the advantage that many countries can be easily analyzed at once—partly because the data are centralized and are usually readily comparable, but also because dealing with bank-specific data requires a fair amount of institutional knowledge (e.g., accounting practices in that country, timing of failures, etc.). However, the macro approach is based on the ex-post knowledge that there was a crisis, the period in which it began and when it ended.7 Even if the timing and the duration of crises could be pinpointed precisely, studying a new episode of banking crisis would need to await for the actual crisis to occur—as there is no generally accepted measure of fragility that can be analyzed prior to the actual crisis.8 Therefore, the second element that is needed is some measure by which to gauge bank fragility, in an ex-ante basis, before the actual failure. Even so, perhaps the major shortcoming of the macro approach is that, by not including individual banks in the analysis, it does not explain why it is that almost without exception not all banks fail even if hit by the same macro shock.9 What makes some banks survive the shock? What are the channels through which macro shocks affect banks? By recognizing that there are generally two types of banks, sound and unsound, the micro approach focuses on the characteristics of these two groups of banks.

The micro camp, and specifically the studies based on CAMEL-type analysis, tend to produce-satisfactory estimation results—in part because there are usually a significant number of potential proxies for the explanatory variables. However, this approach not only often fails to perform appropriately if economic conditions are changed,10 but some of the explanatory variables that are typically used to predict bank failures are themselves endogenous. For example, a common variable used to measure asset quality is nonperforming loans, which itself is a result of poor loan decisions and/or deteriorated economic conditions.11 Banks do not fail because they have a large portion of troubled loans, they fail because of their earlier investment decisions whose outcomes may be also influenced by changed economic conditions—a high level of nonperforming loans are the result of those same fundamental causes.12 A similar argument can be made about capital. Minimum levels of capital (including credit risk-weighted capital-asset ratios) serve as a cushion to absorb shocks. However, ex-post (especially when the bank is in financial trouble), capital is the residual between its (market value) assets and the institution’s nonownership liabilities.13 Indeed, a bank is insolvent when the market value of its capital equity becomes negative. Measures of earnings also typically pose a problem in CAMEL models because, although low earnings are probably indicative of financial distress, soaring earnings do not necessarily mean that a bank is sound. As discussed below in what regulators have found to be the typical life cycle of bank failures, often banks that eventually fail as a result of their high risk-taking were very profitable several periods before their actual failure. As is the case with the macro approach, there is clearly a need in the micro camp to develop generally robust measures of ex-ante bank fragility.

In addition, although there is some anecdotal evidence of banks’ moral hazard behavior (or “looting” as described in Akerlof and Romer (1993)), the supporting empirical evidence seems to be generally lacking.14 Finding proxies for moral hazard is obviously not an easy task. As well, none of the two empirical approaches seems to have been able so far to link bank failures with contagion effects, despite the importance of this source of systemic risk.15 Finally, although it is often assumed that banking crises in developing countries are intrinsically different than in advanced economies, neither approach has provided a definitive empirical answer to this question. All these issues remain to be solved in the literature.

III. Banking System Problems in the United States, Mexico, and Colombia: The Macro-Micro Background

A. Banking System Problems in the United States (1980s to early-1990s)

During the 1980s and early 1990s, more than 1,600 commercial and savings banks insured by the Federal Deposit Insurance Corporation (FDIC) failed (i.e., were closed or received FDIC financial assistance)—far more than in any other period since the 1930s.16 Most of the bank failures occurred between 1986 and 1992, peaking during 1987–89. Bank failures represented approximately 9 percent of both total bank assets, and of the total number of banks existing at the end of 1979 plus all banks chartered during the subsequent 15 years (Federal Deposit Insurance Corporation (1997)). Although unquestionably costly, perhaps at a national level this episode of banking system distress may not have been of “crisis proportions” relative to the size of the U.S. economy when compared, for example, to the recent banking crises in some emerging economies.17

However, bank failures in the United States during this period were highly concentrated in a few states—some of which included the country’s largest banking markets. In some of those states, bank failures accounted for a sizable proportion of the total assets of the state banking system and of the number of existing banks in the region. The states most affected by bank failures include Texas, Oklahoma and Louisiana in the Southwest; New Hampshire, Massachusetts and Connecticut in the Northeast; and California (Table 1). The figures at the state level would seem to be of comparable proportions to several other known episodes of banking crises.18 Interestingly, commercial bank failures in these three U.S. regions—the Southwest, the Northeast, and California—largely occurred in different periods (Figures 1 and 2).19 Prohibitions against interstate branching in several states during this period (including Texas, Colorado, Illinois, Kansas and others) limited the banks’ ability to diversify geographically and made some of the state banking crises akin to sub-national banking crises within the United States.20 “Geographically confined crises were translated into a national problem” (Federal Deposit Insurance Corporation (1997), p. 13). These geographically confined crises are examined below.

Figure 1.
Figure 1.

U.S. Commercial Bank Failures by Region (1986–1995)

(Number of banks)

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

Source: National Authorities.1/ Texas, Oklahoma and Louisiana.2/ New Hampshire, Connecticut and Massachusetts.
Figure 2.
Figure 2.

U.S. Commercial Bank Failures by Region (1986–1995)

(Share of total U.S. failures)

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

Source: National Authorities.1/ Texas, Oklahoma and Louisiana.2/ New Hampshire, Connecticut and Massachusetts.
Table 1.

U.S. Bank Failures, 1980–94

article image
Source: Federal Deposit Insurance Corporation (1997). Notes: FDIC-insured commercial and savings banks that were closed or received FDIC assistance during 1980–94. Based on the total number of banks at the end of 1979, plus banks newly chartered in 1980–94.

The Southwest

The macro/regional setting

Commercial bank failures in the Southwest (defined here to comprise Texas, Oklahoma and Louisiana because the banking problems in the region were concentrated in those states) largely occurred during 1986–92, peaking in 1988–89. The banking problems in this region followed a decline in regional economic activity (Figure 3) which was largely associated with weak agricultural prices and plunging oil prices. The decline in oil prices since 1981, and their virtual collapse in 1986, represented a major shock for the energy-producing southwestern states (Figure 4). The regional real estate market had been booming as a result of the hitherto strong energy market. As oil prices began drifting downward in 1981, southwestern banks sought new investment opportunities in the then-booming real estate markets, particularly in the commercial real estate market. Real estate markets began a prolonged decline in the mid-1980s with construction activity (Figure 5) and real estate prices plunging.21

Figure 3.
Figure 3.

U.S. State Personal Income - Southwest

(Annual percent change)

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

Source: U.S. Bureau of Economic Analysis; National Authorities.
Figure 4.
Figure 4.

Oil Prices - West Texas Intermediate

(Dollars per barrel)

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

Source: International Monetary Fund, International Financial Statistics database; National Authorities.
Figure 5.
Figure 5.

U.S. Non-Residential Construction Building Permits - Southwest

(Index, 1985 = 100)

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

Source: U.S. Census Bureau: National Authorities.
Characteristics of failed banks vs. non-failed banks

Against a common macro/regional environment and shocks, not all banks located in the same area of distress are affected equally by those events and, indeed, not all banks typically fail. What makes banks that fall into crisis different from those that do not? Are some of those characteristics evident several periods before the actual crisis? Are other characteristics particularly apparent shortly before the crisis?22

In the case of southwestern banks, the main characteristics of banks which eventually failed relatively to the banks that did not fail (Figure 19 in Appendix II) and that were evident through much of the period of study include: banks that failed had a significantly higher ratio of commercial and industrial loans (which include loans to the energy sector) relative to total assets; a higher proportion of loans in the commercial and residential real estate sector, and in consumer loans; a higher ratio of loans to assets; a lower proportion of liquid assets and higher expenses; paid a higher average interest rate on deposits and received a lower average yield on loans. In contrast, the ratio of nonperforming loans to total assets (and relative to total loans) for banks that failed were somewhat higher at the beginning of the period than for banks that did not fail, but showed a dramatic deterioration as the crisis intensified. Similarly, equity capital ratios (and profitability) were somewhat lower at the beginning of the period for banks that failed, but deteriorated dramatically as the crisis intensified. The capital and reserve coverage of nonperforming loans (coverage ratio) of banks that eventually failed declined rapidly, becoming negative in 1987—just before the peak of the southwestern bank crisis in 1988–89.23

The Northeast

The macro/regional setting

The banking problems in the Northeast (defined here to comprise New Hampshire, Connecticut and Massachusetts, also because a large part of the bank failures in the region occurred in these states) became prominent during 1991–92. Supported by a strong regional economy, the real estate markets in the Northeast boomed during the 1980s. However, regional economic activity weakened late in the decade (Figure 6) as a result of a slowdown in military spending as the Cold War came to an end, and a decline in activity in the computer industry which was largely concentrated in New England. The real estate market boom of most of the 1980s turned into a bust. In the early 1990s, an oversupply of real estate projects led to a decline in construction activity (Figure 7) and real estate prices went into a sharp decline. The northeastern banks had been aggressive participants in the hitherto prosperous real estate markets of the 1980s.

Figure 6.
Figure 6.

U.S. State Personal Income - Northeast

(Annual percent change)

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

Source: U.S. Bureau of Economic Analysis; National Authorities.
Figure 7.
Figure 7.

U.S. Non-Residential Construction Building Permits - Northeast

(Index, 1985 = 100)

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

Source: U.S. Census Bureau: National Authorities.
Characteristics of failed banks vs. non-failed banks

Northeastern banks that eventually failed, relatively to banks that were not intervened, also exhibited certain characteristics several years before the period of distress (Figure 21 in Appendix II), including: higher shares of commercial and industrial loans, and of commercial real estate loans; a lower share of residential real estate loans and consumer loans; a higher loans to assets ratio; and a lower liquidity ratio. Nonperforming loans and equity capital between these two groups of banks were nearly undistinguishable (in fact, the ratio of equity capital to total assets was slightly higher for banks that eventually failed) several years before the period of distress surfaced, and it was not until the crisis escalated that nonperforming loans and equity capital deteriorated dramatically. Diminished profitability also seemed to be a late indicator of impending problems for those banks that eventually failed. The coverage ratio of banks that eventually failed declined sharply in 1990 when it became negative—about one year before the peak period of failures in 1991–92.

California

The macro/regional setting

California’s banking problems were particularly prevalent during 1992–93, but there were some occurrences of failure beginning in the mid-1980s. Although many banks failed in this state, they were generally small (in contrast to the Southwest and the Northeast). The prosperous 1980s took a turn by the end of the decade as the California economy slowed (Figure 8). Bank problems surfaced in precisely the sectors that had led to the economic boom of the 1980s: defense-related manufacturing associated with the Cold War, construction activity, and the real estate markets. During most of the 1980s, California had been also a major recipient of Japanese investments and these inflows of capital had contributed to a booming property sector. As the Japanese economy fell in recession in 1990, Japanese investors and banks (which themselves were beginning to face “a large portfolio of nonperforming assets, and pressures from financial markets and Japanese regulators to deal with these issues” (Federal Deposit Insurance Corporation (1997), p. 396)) significantly cut their lending and investments in U.S. real estate markets. California had been an important recipient of the Japanese investments in the U.S. property market. When Japanese investments declined, this also had an important impact on the California economy.24 The real estate market and construction activity suffered a significant decline in California during the late 1980s and early 1990s (Figure 9).

Figure 8.
Figure 8.

U.S. State Personal Income - California

(Annual percent change)

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

Source: U.S. Bureau of Economic Analysis; National Authorities.
Figure 9.
Figure 9.

U.S. Non-Residential Construction Building Permits - California

(Index, 1985 = 100)

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

Source: U.S. Census Bureau; National Authorities.
Characteristics of failed banks vs. non-failed banks

Banks that eventually failed in California, relatively to those that did not fail, exhibited certain characteristics several years before the period of distress (Figure 23 in Appendix II), most notably: a higher loans to assets ratio; a lower liquidity ratio; and higher expenses. Nonperforming loans deteriorated sharply as the crisis intensified. Capital ratios (and profitability) also worsened as the crisis unfolded. The coverage ratio of banks that eventually failed declined sharply in 1991, approaching zero and becoming negative about one year before the peak period of failures in 1992–93.

What role did interest rates play?

At the national level, U.S. nominal and real interest rates generally declined from the mid-1980s to the early 1990s (Figure 10), which helped the condition of most banks by lowering the cost of funding and improving the creditworthiness of borrowers. The decline in interest rates was accompanied by an upward-sloping yield curve in the early 1990s which “increased the value of bank security portfolios and raised net interest margins on new loans, reducing the number of bank failures” (Federal Deposit Insurance Corporation (1997), pp. 410–11).

Figure 10.
Figure 10.

U.S. Real Interest Rates

(Percent per anum)

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

Source: International Monetary Fund, International Financial Statistics database.

B. Banking System Problems in Mexico (1994–95)

The macroeconomic setting

The recent Mexican financial crisis is regarded by many as the prototype of crisis that could threaten other emerging markets whose capital markets are becoming more integrated with global markets because of the evident connection between banking crisis and currency crisis, and also because of the apparent contagion to other economies.25 The Mexican economy was shaken by the collapse of the peso in December 1994, with interest rates increasing manyfold and the economy contracting drastically in the months that followed (Figures 11 and 12).26 Two relatively small banks received an infusion of capital and had their management replaced by the regulatory authorities, apparently in connection with fraudulent activities, during the third quarter of 1994. But it was not until after the currency collapse that several banks were intervened by the authorities.27 By end-1995, more than ½ of all the Mexican banks had received financial support from the government, including the two largest banks which had a significant amount of their problem loans removed from their books and purchased by the authorities in the final quarter of 1995.

Figure 11.
Figure 11.

Mexico: Nominal Exchange Rate

(Pesos per US dollar)

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

Source: International Monetary Fund, International Financial Statistics database; Press Releases.
Figure 12.
Figure 12.

Mexico: Short-term Real Interest Rates

(Percent per anum)

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

Source: International Monetary Fund, International Financial Statistics database.

Characteristics of failed banks vs. non-failed banks

Banks that eventually failed in Mexico, relatively to those that did not fail, exhibited certain characteristics evident several periods before the crisis (Figure 25 in Appendix II), most notably: a higher proportion of residential loans; and lower profitability. Banks that eventually failed saw their level of nonperforming loans deteriorate rapidly prior to the crisis, while they had much lower capital ratios than was the case for banks that did not fail.28 The coverage ratio of failed banks was also significantly lower than for non-intervened banks.

C. Banking System Problems in Colombia (1982–87)

The macroeconomic setting

The banking crisis in Colombia was spread over several years, with some banks failing in 1982 and during 1986–87. Although the number of banks that failed in Colombia was relatively small, especially when compared to more recent crises in other emerging economies, failed banks represented about 20 percent of the total assets in the banking system.

Most of the explanations available concerning the causes of Colombia’s banking crisis point to the sudden end of the coffee boom of the late 1970s (for example, Rojas-Suârez and Weisbrod (1995)). In Colombia, economic conditions in the 1970s and early 1980s were heavily dependent on coffee exports. The price of coffee in international markets (based on Brazilian coffee) fell from over US$2 per pound in 1977 to 83 cents in 1981. Coffee prices recovered to US$1.90 per pound in 1986, before they fell again to 90 cents in 1987. Colombia’s export prices followed closely this pattern (Figure 13).

Figure 13.
Figure 13.

Colombia: Price of Exports

(Index, 1980 = 100)

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

Source: International Monetary Fund, International Financial Statistics database; National Authorities.

Characteristics of failed banks vs. non-failed banks

Banks that eventually failed in Colombia, relatively to those that did not fail, showed certain characteristics (Figure 27 in Appendix II), including: declining ratios of liquidity and deposits from the public relative to total assets; rising average interest rates paid on deposits and interest received on loans; and increasing expenses. Nonperforming loans rose sharply prior to the period of failures in 1986–87. The capital equity ratios of failed banks were slightly lower than of non-failed banks before 1986–87, but they rose sharply at the time of the bank interventions—presumably in connection with some rehabilitation plans. The coverage ratio of failed banks declined sharply and became negative in 1984— two years before the 1986–87 wave of bank failures.29

IV. Model of Bank Distress and Main Variables

A. The Basic Framework

The literature on early-warning systems of bank failure has primarily relied on bank-specific variables for clues about the soundness of individual banks, while the analysis of macroeconomic factors has been essentially the domain of the literature on banking (and financial) crises. Relatively few studies have attempted to integrate these two approaches. One of those attempts is the basic framework proposed in González-Hermosillo (1996), which suggests that bank fragility is essentially a function of liquidity risk, market risk and credit risk.30 In turn, those risks are conditioned by macroeconomic conditions and can be influenced by the overall fragility of the banking system. In particular, based on a simple two-period balance sheet framework,31 the probability of an individual bank becoming unsound Fzi can be expressed as a function

F z i = F ( x , y , k ) ( 1 )

where deposit flows during a given period are given by x (normalized by the stock of total deposits or assets), y constitutes the bank’s (normalized) net asset income and k can be viewed as the bank’s optimal (normalized) level of capital required to minimize the expected costs of insolvency. The values of x and y are known in period t but are only known in a probabilistic form for period t + 1. In this framework, k is chosen by bank’s management as the optimal level of initial capital (but could also be imposed by the regulatory authorities as the minimum level of capital)—however, the ex-post market value of capital is clearly the difference between (market value) assets and non-ownership liabilities.

Equation (1) is generalized to explore the fundamental sources of risk. Assuming a zero expected recovery rate of defaulted loans, the expected net asset income at the beginning of the period can be expressed as a function of market risk β* (Γ) im and default risk τ (Γ, im):

y = y ( β * ( Γ ) i m , τ ( Γ , i m ) ) ( 2 )

In particular, the realized net asset income of a bank at the end of the period will depend on: the realized (exogenous) market return im adjusted by the β* (·) of the bank’s asset portfolio (β*(·) ≥ 0) chosen by bank management;32 and on the occurrence of default by the borrowers τ (·). The vector Γ is assumed to encompass macroeconomic variables related to the state of the business cycle (e.g., output growth, housing market activity, etc.). Assuming that borrowers are always willing (but sometimes unable) to pay, the likelihood of default by borrowers may change with the business cycle and with changes in the market interest rate.

This basic model can be augmented in two respects. First, if banks were assumed to follow a “herding” type of behavior vis-à-vis other banks, the choice of β* (·) could also be a function of the fragility of the overall banking system FT (which is essentially derived by aggregating the fragilities of individual banks, weighted by their asset shares). Thus, for example, in an effort to maintain market share, banks may increase their degree of risk-taking if they perceive that other banks are doing the same thing—particularly if there are deposit guarantees which would reduce the banks’ potential costs of assuming such added risks. Second, banks can also encounter higher default risk (even if economic conditions are favorable) if borrowers are able, but not willing, to repay their loans. This may reflect poor credit controls by banks’ management. Alternatively, this may be also the result of banks taking deliberate extreme credit risks (i.e., engaging in “looting” or, more generally, moral hazard) by extending loans “in circumstances in which no reasonable person would expect a future positive payoff in any future state of the world, but for which the present payoff was very high” (Akerlof and Romer (1993), pp. 28–29).

Depositors’ behavior regarding their desired flows of deposits into, or out of, the bank(s) where their deposits are held can be viewed as a function or

x = x ( u , ( F z i | γ * ( γ max , e , F T ) , Ω ) ) ) ( 3 )

where u refers to the depositors’ exogenous and stochastic needs for bank deposit transactions (e.g., reflecting liquidity needs, payments requirements, savings decisions, etc.). Fzi is the expected probability that the bank in question (bank i in this case) will fail given the anticipated effective level of deposit guarantees, γ*, and given the information set available at time t, Ω, used to form expectations. In the case of explicit deposit insurance schemes, the level of γ* is a function of the statutory maximum level of deposits per account covered by the deposit insurance program γmax, the (exogenous) endowment available in the deposit insurance fund e, and the expected probability FT that there may be a significant number of banks in the system also failing. If many banks fail during the same period, the endowment may not be sufficient to cover all the ailing banks. If depositors suspect that the bank where they hold their deposits is in trouble, they may withdraw their money from that bank and place it in another bank perceived to be sound. If they suspect that many banks are in trouble but they are unable to distinguish which banks are actually unsound (i.e., there is asymmetric information), and there is uncertainty as to whether the resources in the deposit insurance fund are sufficient to cover all losses, there may be deposit runs—even from banks that would be otherwise sound.33 Sudden deposit withdrawals would increase liquidity risk for banks and could lead to failure if assets cannot be made readily liquid, while maintaining their market value, to cover for the deposit withdrawals. In the presence of asymmetric information, the fragility of the overall banking system can affect even sound banks if indiscriminate deposit runs develop—thus, contagion can occur.

B. The Life Cycle of Bank Failures

In practice, it is of course often difficult to determine what may constitute “excessive” risk-taking by banks because they are in the business of taking risks. Too much risk can result in financial distress, but too little may also threaten the bank’s long-term viability. Excessive risks can be extremely profitable at first. Changed economic conditions following the initial investments frequently reveal (ex-post) the actual magnitude of those risks.

In interviews with U.S. regulators, for example, it was found that bank failures seem to have a life cycle which roughly corresponds to three phases (see Federal Deposit Insurance Corporation (1997)). In the first stage, there may be rapid loan growth, loan concentrations emerge, and lending is aggressive (internal controls in the growth area are weak, and underwriting standards are lenient). The increased lending may be funded by a volatile lending source. This growth could occur throughout the entire institution or within a specific asset type. The growth will generate added revenue from increased loan fees and interest income. In addition, because these are largely new loans, initially there are no delinquencies, so that the growth is almost always accompanied by growth in income and capital (assuming retained earnings). If the rapid growth draws the attention of the regulators, management usually points to the “excellent earnings” and contribution to capital that the growth has provided. In the United States, this stage of development seems to take up to two years.

In the second stage, the institution has rising loan-quality problems. Associated expenses may-far exceed the industry averages. Nonrecurrent sources of income are used to maintain the same level of profits that existed during the growth phase. Eventually profits begin to decline, and inadequate reserve levels become apparent. At this point the bank may have a high loans-to-assets ratio. Management may still believe that the problem is manageable. This stage may take an additional one to two years.

In the final stage, deteriorating asset quality is a serious problem. If the institution is large, the capital markets (if sufficiently developed) may have recognized that the institution has inadequate loan-loss reserves and are unwilling to provide fresh capital and funding. At this point, major changes in the banks’ operations are necessary if they are to avoid failure. Dividends may be cut, expenses (mostly personnel) are slashed, and assets are sold to cover losses and expenses. This crisis phase may last up to a year and results either in failure of the bank or, if fundamental changes are made, in its eventual recovery.

Based on this analysis of the life cycle of bank failures, it would appear that different indicators tend to surface at different times in the cycle. In developing a procedure to identify the ultimate sources of risk for U.S. banks, research at the Federal Deposit Insurance Corporation (1997) attempted to separate “condition” or “ex-post risk” variables—as indicators of the current strength or weakness of a bank—from some of the fundamental sources of risk. They find that a bank in a weak condition would typically have low equity capital and net-income ratios, and high nonperforming loan ratios. On the basis solely of those condition variables, they found little evidence in 1982 to distinguish banks that failed from those that did not fail five years later. They then examined nine sources of (ex-ante) risk: loans-to-assets ratio; large deposits to total liabilities; return on assets; asset growth; loan growth; operating expenses to total expenses; salary expenses per employee; interest on loans and leases to total loans and leases; and interest and fees to loans and leases. Their study, covering all U.S. banks for the period 1980–88 and based on the explanatory power of individual variables using logit regression, showed that the best long-range predictor of failure is bank’s loans-to-assets ratio.34 However, the predictive power of most variables frequently changed across different periods. Furthermore, it is not clear from the study what were the criteria for choosing those particular variables (other than perhaps because of their individual predictive power), nor was the joint explanatory power of different variables put together apparently explored.

C. Main Variables

In a similar spirit to the FDIC approach, this study also attempts to identify fundamental sources of risk, which are largely the result of bank management’s behavior in the past, and proximate indicators of fragility (or bank “condition” in FDIC s terminology). However, the choice of variables here is consistent with the theoretical framework discussed above in which bank distress is a function of market risk, credit risk and liquidity risk, often related to economic conditions and the fragility of the overall banking system. Furthermore, the focus is on finding “common ground” across different episodes of banking system problems (rather than fitting models based on different idiosyncratic variables in each case but with the highest predicting power). Obviously, accounting systems are often different among countries and the relevant measurements of alternative types of risk may vary according to the circumstances specific to each episode. However, to the extent possible, an effort was made to maintain conceptual equivalences across the different episodes of banking crises examined.

Table 2 summarizes the indicators of bank failure and distress examined in this study, and their expected signs. The hypothesized proximate indicators of fragility are nonperforming loans, capital equity, and the combined equity capital and loan reserve coverage of nonperforming loans—which are similar to most of the FDIC s condition variables.35 The proxies for the fundamental sources of risk were chosen based on the specific type of risks examined: market, default or liquidity risk. In general, high market risk would occur if a bank’s loan portfolio is heavily concentrated on booming sectors that may be subject to boom-bust dynamics (including, for example, sectors vulnerable to foreign exchange fluctuations), in areas that are highly dependent on cyclical economic conditions (e.g., real estate, sectors dependent on commodity prices or the stock market), or in sectors with returns significantly higher that the market rate of return (investments with high β*). The variables proxying for market risk are likely to vary depending on the specific circumstances in each region. For example, a high exposure to commercial and industrial loans was probably only an important factor in the case of banks in the U.S. Southwest because this category of loans includes loans extended to the energy-dependent sector.

Table 2.

Indicators of Bank Failure/Distress

article image
article image
1/ Nonperforming loans are defined in the United States as the ratio of loans past due 90 days or more, plus nonaccrual loans and repossessed real estate loans; in Mexico (until 1997) as the ratio of loans past due 30 days or more (past due interest on that installment only, not including the entire unpaid principal balance); and in Colombia as the ratio of loans past due 180 days or more.

There are several proxies for credit or default risk. In general, a high loan-to-assets ratio is probably associated with fast lending growth and weak internal credit controls. High yields paid on loans may indicate that the bank is originating high-risk loans, but low yields may also mean that risk is not priced properly. High interest rate spreads may also mean that the bank is taking risky loans; however, low interest rate spread may reflect the bank’s efficiency. The sources of liquidity risk depend on whether some type of deposits are more volatile and whether a bank has enough liquidity to respond to large deposit withdrawals. The variables include large deposits, interbank deposits and deposits from the public, liquid assets and the average interest rate paid on deposits. The proxies for moral hazard are two. First, insider loans because of their inherent potential conflict of interest (however, this data are only available for the United States). Second, the ratio of interest income on loans, fees and leases to total assets because—consistent with the observation made by Akerlof and Romer (1993) for the case of some financial institutions in the U.S. Southwest during the 1980s—moral hazard may be associated with banks loading up on up-front fees, commissions and high interest rates on loans extended even though the long-term viability of the loans is not expected to be favorable.36

The regional and macroeconomic variables were generally based on measures of economic activity and real interest rates. In addition, other variables representative of the specific circumstances of each episode of banking system distress were also included (e.g., oil prices in the case of the U.S. Southwest, currency depreciation in Mexico, and coffee-related export prices in Colombia). The banking sector variable, as a proxy for contagion based on the theoretical framework discussed above where contagion can occur through deposit runs and/or banks’ hearding behavior, used was the ratio of total loans in the banking system of the country or region relative to output.37 Banking system loans growing significantly faster than output would be generally associated with increased fragility. This may be the result of banks following “herding behavior” patterns in their lending. The increased fragility of the banking system could also lead to deposit runs if depositors are unable to distinguish which banks remain solvent. However, it may also be that banks may find some safety in large numbers if regulators are less able (or willing) to intervene additional banks as more banks run into financial difficulties.

Since all the variables chosen are essentially proxies, some additional bank-specific variables (including measures of profitability, cost efficiency, size and dummies for banks part of a holding company structure) were also included to explore their potential role in explaining bank failure and distress. Before discussing the empirical results, the empirical methodology is briefly reviewed.

V. Empirical Methodology

The probability of bank failure was estimated based on the fixed-effects logit model for panel data suggested by Chamberlain (1980). Let i = 1, 2, …., n denote the groups (banks in this case) and t = 1, 2,…, Ti the observations for the ith group. Let yit be the dependent variable taking on values of 0 or 1, and xit be a row vector of explanatory variables in a model of the form

y it = α i + β ′ x it + ε it ( 4 )

Although constant over time, the parameter αi may be different for different cross-sectional units. To account for heterogeneity, letyi = (yit, … yiT) be the outcomes for the ith group as a whole and k1i = Σ yit (the summation from t = 1 to Ti) be the observed number of ones for the dependent variable in the ith group. Fixed-effects logit regression maximizes a conditional likelihood function based on the probability of a possible value of yi, conditional on k1i = Σ yit (the summation from t = 1 to Ti).38

The (time-varying) survival time models estimated were based on monotonie hazards. Let T represent the duration of stay in the state of no-failure (time to exit) and t a realization of T. Assuming that the random variable Γ has a continuous probability density function f(t), the cumulative probability distribution is given by

F ( t ) = ∫ o t f ( s ) ds = Prob ( T ≤ t ) ( 5 )

the survival function is given by

S ( t ) = 1 − F ( t ) = Prob ( T ≥ t ) ( 6 )

and the hazard function can be written as

λ ( t ) = lim dt → 0 Prob ( t ≤ T < t + dt | T ≥ t ) dt = − d ln S ( t ) / dt ( 7 )

which is the conditional probability that a bank that has occupied the state of no-failure for a time t leaves it in the short interval of length dt after t. Alternatively, the hazard function is the instantaneous rate of leaving the state of no-failure per unit of period at t.

Specifically, the hazard function can take the form

λ ( t ) = λ 0 ( t ) e β 1 x 1 + β 2 x 2 + … + β k x k ( 8 )

where λ0 (t) is the baseline hazard (the individual heterogeneity). In the case of the proportional hazard model with time-dependent explanatory variables, first suggested in Cox (1972), the partial likelihood estimator provides for a method of estimating the vector β without providing a direct estimate of λ0 (t).39 The Weibull hazard assumes that λ0(t) = ptp-1, where p is the shape of the parameter to be estimated from the data. The exponential hazard model is a particular case of the Weibull function in which p = 1, and hence λ0 (t) = 1.40

VI. Empirical Evidence

A. “Distressed” Banks

From the informal examination of the broad characteristics of banks that failed vs. those that did not failed in the five episodes of banking system problems examined, it would appear that a notable increase in the nonperforming loans (relative to total assets and/or total loans, and adjusted or not for loan reserves) of fragile banks shortly preceded the period of banking crisis. Often, a decline in the capital equity ratios of fragile banks also preceded the period of crisis.41 To allow for a general framework, the main indicator of distress was constructed by combining these elements: the ratio of capital equity and loan reserves minus nonperforming loans to total assets (coverage ratio).

Focussing on the coverage ratio as the main indicator of distress has several advantages. It allows for the possibility that two banks with an equally high ratio of nonperforming loans to assets would be in a different standing if one has set aside reserves to cover for a significant amount of the problem loans, or if it has a higher level of equity capital. In the cases examined, the coverage ratios of banks which were never intervened seemed to be generally bounded at the upper-end (usually at around 8–12 percent)42 because equity capital, constituting resources that are expensive for a bank to maintain inactive, is usually not much higher than the minimum level required by the authorities. As the coverage ratio declines and approaches zero, the bank’s own resources in the form of equity capital and loan reserves become increasingly insufficient to cover for nonperforming loans.43 As this happens, a bank would be increasingly more fragile and likely to be in distress.

Banks were considered to be in “distress” if their coverage ratio was lower than a certain threshold.44 In the case of U.S. banks the coverage ratio threshold was set at zero, so that banks were considered to be in “distress” when their coverage ratios were zero or negative because their own resources in capital equity and reserves for problem loans would be insufficient to cover for nonperforming loans. In the cases of Mexican and Colombian banks, the threshold would be expected to be higher because of the more narrow definition of nonperforming loans than in the United States.45 Arbitrarily, the coverage ratio threshold for -Colombia and Mexico was set at 1.5 percent.46

In general, the data examined suggest that banks which failed typically showed earlier signs of distress and often had multiple periods in which distress was apparent before the actual intervention. However, not all banks which became distressed were necessarily intervened—likely because of corrective actions that may have diminished the need for intervention or due to improving economic conditions. Intervention is an extreme one-time event and its timing is largely determined by the regulators. As discussed below, when aggregating the incidence of bank distress for the overall system (weighted by banks’ asset shares) it was generally found that the rise in this indicator of fragility typically preceded and magnified the actual ex-post (asset weighted) incidence of failures in the banking system.

In order to compare the characteristics of failed and non-failed banks vs. banks classified as “distressed” or not, banks were regrouped based on whether they experienced episodes of distress or not. Banks’ characteristics (based on their medians) are presented in Appendix II: Figure 20 refers to banks in the U.S. Southwest, Figure 22 to banks the U.S. Northeast, Figure 24 to California, Figure 26 to Mexico and Figure 28 to banks in Colombia. The figures suggest that the characteristics of distressed banks were broadly consistent with those of banks that eventually failed; and similarly for non-distressed and non-failed banks. (The figures related to the characteristics of failed and non-failed banks were previously discussed and can be also found in Appendix II).

To ascertain formally what were the predominant factors that determined the probability and timing of bank failure, and whether in fact those same factors would also explain bank distress, the following section focuses on the econometric analysis.

B. Probability of Bank Failure and Time of Failure

A set of models were estimated for each region with banking problems. The first simple models examined the individual explanatory power of the ratio of nonperforming loans to total assets (NPLA) and the ratio of equity capital to total assets (EQ) in determining the one-period ahead probability of intervention and time of intervention. Second, a model based on bank-specific variables (proxying for market risk, credit risk, liquidity risk and moral hazard), but excluding NPLA and EQ was examined. Third, the latter was augmented to include macro/regional variables, and a banking system variable (proxying for contagion).47 The dependent variable takes the value of one if the bank was intervened at time t + 1 and zero otherwise. The explanatory variables correspond to time t.

As discussed below, in almost every case, NPLA and EQ were fairly good predictors of bank failure—though NPLA was generally the best—which is consistent with the impression gathered from the figures examined that NPLA and EQ (or, more generally, the coverage ratio (COVR)) seem to be good proximate indicators of failure. In a second step, the same models based on bank-specific variables and the augmented full models were also estimated for the case in which banks became distressed. In this case, the (contemporaneous) dependent variable takes the value of one if a bank is distressed and zero otherwise.

As discussed earlier, the probability of failure is estimated based on a fixed effects logit model.48 The timing of failure was estimated by fitting a non-parametric (time-varying) Cox proportional hazard model. A parametric (time-varying) Weibull distribution—and an exponential distribution if the maximum-likelihood estimator of p in the Weibull function was close to 1—with monotone hazard rates were also estimated.49 The estimates based on the Cox proportional hazard model were generally consistent with those obtained from estimating a parametric Weibull hazard model and, hence, only the latter are reported.50

Empirical results for the U.S. Southwest

For the case of the three regions in the United States, the data is annual covering the period December 1985 to December 1992. The maximum survival time is (right) censored at 96 months.51 In the Southwest, 2,946 commercial banks in the states of Texas, Oklahoma and Louisiana were examined, of which 647 banks were intervened during the period 1986–93.

The empirical findings for bank failures in the Southwest are reported in Table 3.1. In models (1) and (2), a higher ratio of nonperforming loans to total assets (NPLA) and a lower ratio of capital equity to total assets (EQ) are associated with a higher probability of failure and a higher hazard rate (or, equivalently, reduced survival time). The model χ2 are high in both models, but the pseudo R2 is higher in the case EQ.52 However, the odds ratio (or exponentiated coefficient eβ) of NPLA is larger; such that a one percentage point rise in NPLA would increase the probability of failure by more than 1 ½ percentage points, while a one percentage point increase in EQ would reduce the probability of failure by about ⅓ of a percentage point.53 Similarly, a one-unit rise in NPLA would increase the hazard of failure (or reduce the survival time) by 1.14, while a one-unit increase in EQ would reduce the hazard (or increase the survival time) by 0.9.54 The estimated p values greater than one for the Weibull hazard models indicate positive duration dependence; that is, the likelihood of failure at time t, conditional upon survival up to time t, is increasing in t.

Table 3.1.

Estimation Results—U.S. Commercial Bank Failures in the Southwest

article image
article image
Notes: The dependent variable takes the value of one if a bank is intervened at time t+1 and the value zero otherwise. Probability of failure is estimated by fixed-effects logit. The odds ratio and hazard ratio depict exponentiated coefficients. The z statistics are given in parenthesis. P|z| is the test of the underlying coefficient being zero. One, two, and three asterisks indicate significance levels of 1, 5 and 10 percent, respectively.

Model (3) is based on bank-specific variables (proxying for market, credit and liquidity risks, and for moral hazard) and the full model (4) augments regional/macro variables and a banking sector variable (proxying for contagion).55 In general, higher commercial real estate loans (LCOMRE) and residential loans (LRESI) increase the probability of failure, suggesting that -these loans were quite risky. Higher commercial and industrial loans (LCI) reduce the survival time.56 The existence of high credit risk is somewhat ambiguous. On the one hand, a higher average yield on loans (LNYIELD) is associated with a lower probability of failure and an increased survival time—suggesting that credit risk may have been priced properly. Although a higher loans-to-assets ratio (LAS) is correlated with a lower probability of failure, it is also associated with a reduced survival time (this latter effect would be consistent with high credit risk coming from a fast expansion of loans to assets). With respect to liquidity risk, the variables had most of the expected effects. In particular, an increase in interbank deposits (DEPH3), large deposits (DEPLGE) or in liquid assets relative to total assets (SEC) are associated with a lower probability of failure. However, although higher level of liquid assets is associated with increased survival time, higher levels of interbank deposits and large deposits seem to reduce banks’ survival time—possibly because a higher ratio of these deposits relative to total assets leaves banks more exposed to sudden deposit withdrawals. There is no evidence that higher insider loans (INSL) increased the probability of failure or reduced banks’ survival time in the Southwest. However, a higher ratio of interest income and fees to total assets (INTAS) had an significant positive correlation with a higher probability of failure and a reduced survival time—consistent with the hypothesis that moral hazard can be associated with high fees and current payments.

Model (4) introduces macroeconomic and banking sector variables (the latter as a proxy for contagion). As expected, an increase in regional economic activity (SPERYCH) would increase the survival time of banks. However, a puzzling result is that lower real interest rates (INTRS) seem to be consistent with a high probability of failure. The latter result may be reflecting the fact that U.S. interest rates generally declined throughout the entire period of study, largely as a result of a credit crunch.57 Notwithstanding this puzzle, lower real interest rates are associated, as expected, with an increased survival time (interestingly with the largest hazard ratio). A lower price for oil increased the probability of failure, but high oil prices reduced the survival time—the latter may be associated with asset “bubbles” since the risk of a burst is higher as the boom stage develops further. Finally, although the coefficients are small, a higher ratio of banking system loans relative to the region’s personal income (STLNPI) is correlated with a lower probability of failure, but also with a reduced survival time—the former may be consistent with regulatory forbearance (in the sense that having a large number of banks already in difficulty could reduce the probability that an additional weak bank would be also intervened) and the latter with possible contagion. Adding the macroeconomic and banking system variables greatly increases the explanatory power of the model compared to model (3) where only bank-specific variables are considered.

Models (5) and (6) repeat the same exercise just described, except that the dependent variable becomes bank “distress” instead of failure. The number of occurrences of distress during the 1985–92 period in the U.S. Southwest is nearly three times the number of actual bank failures and some banks experienced multiple occurrences of distress (i.e., a bank’s coverage ratio may be below the threshold repeatedly during several periods). The results are broadly consistent with models (3) and (4) discussed above. The most notable differences are the following: higher commercial real estate loans (LCOMRE) and residential loans (LRESI) are now also associated with a reduced survival time (when examining bank failures, a higher level of these loans was associated only with a higher probability of failure); an increase in regional economic activity (SPERYCH) is now also correlated with a lower probability of distress (before it was only with an increased survival time); higher real interest rates (INTRS) are now associated with a higher probability of distress (before the relationship seem to be inversed and puzzling); and the signs for interbank deposits (DEPJJB) are now conflicting in models (5) and (6) when explaining the banks’ probability of distress.

The predicted probabilities of failure and survival time for each bank and in each period, based on the models described, were then aggregated and weighted by the banks’ asset share relative to the whole banking system in the region for each period. This imposes a difficult test, because it is not only important to infer correctly the actual number of failures (specifically, the correct number of predicted zeros and ones), but also to predict the precise banks that failed in each period according to their relative size. The expected probability of failure based on the coverage ratio (ph_covr) would have overestimated somewhat the actual (asset-weighted) probability of banking failures in the Southwest, but it would have given a clear indication that the banking system was rapidly weakening one or two years before the peak of the banking crisis (Figure 14.1). The expected probability of failure based on the ratio of nonperforming loans to total assets (ph_npla), and based on the ratio of capital equity to total assets (ph_eq), would have predicted similar developments. The expected probability of failure based on bank-specific variables (ph_bs), and the models augmented by the regional/macro (ph_rm) and the banking system variables (ph_f), would have shown a deterioration of more comparable magnitudes to the actual events; however, they would have predicted a new wave of crisis emerging at around the time when the actual crisis was dissipating.58

Figure 14.1.
Figure 14.1.
Figure 14.1.

U.S. Banking System Failures - Southwest

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

The predicted intervention times (based on a log expected time parameterization, as noted earlier) for individual banks were aggregated similarly by taking asset-weighted averages. The expected failure times are given in arbitrary time units, in connection with the arbitrary measurement of time and the maximum (right) censoring of banks. Hence, expected failure times should be interpreted with caution as more indicative of general patterns than of exact dates. In general, a higher degree of bank fragility would be associated with reduced time until failure or, equivalently, reduced survival time (failure would be expected to occur sooner). Based on the coverage ratio, the predicted (asset weighted) failure time for the banking system (et_covr), shows a significant decline prior to the peak of the crisis (signaling increased fragility) and an improvement as failed banks exit the sample (Figure 14.1). Models based on bank-specific variables (et_bs), augmented by regional/macro variables (et_rm) and a banking sector variable (et_f) appear to be much less foretelling of a forthcoming crisis—although, this may reflect in part the fact that the data does not go back far enough prior to the surfacing of the banking crisis. However, the (asset weighted) expected failure time does increase significantly as weak banks that fail are removed from the system.59

Lastly, the same exercise of predicting the fragility of the banking system by adding up the asset-weighted fragility of individual banks was performed based on the expected probability of distress and time of distress (Figure 14.2). The number of occurrences of distress during the 1985–92 period is nearly three times the number of actual failures. The asset-weighted distress occurrences seem to precede and magnify the asset-weighted incidence of failures; in other words, there were more incidences of distress than of failure, and the incidences of distress were generally apparent before the actual bank interventions. The estimated probability of distress based on the three models (ph_bs, ph_rm and ph_f) seem to capture reasonably well the degree of distress in the banking sector. The chief difference with the predicted probability of intervention is that the surge in the predicted probability of failure toward the end of the period is much more modest when focusing on bank distress. The expected time of distress’ paths are also broadly consistent with the case of interventions, except that the expected failure time based on bank-specific variables (et_bs) is flatter when focusing on distress that when examining actual bank interventions.

Figure 14.2.
Figure 14.2.

U.S. Banking System Distress - Southwest

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

Basic CAMEL approach: U.S. Southwest

How would these results compare with the standard CAMEL approach? To shed some light on this question, a simple model based on typical CAMEL proxies was estimated to predict bank interventions. For simplicity, the basic model relies only on one proxy for each variable (many studies often use several proxies at the same time): the ratio of capital equity plus loan-reserves to total assets (CA) for capital adequacy; the ratio of nonperforming loans to total assets (NPLA) for asset quality; the ratio of loans to assets (LAS) for management quality; net income (NI) for earnings; and the ratio of liquid assets to total assets (SEC) for liquidity (model (1) in Table 3.3). In model (2), an experiment is performed in which NPLA and CA—the variables used here to define “distress”—are dropped.60 While the fit of model (1) is reasonably good, it would seem that CA and NPLA indeed play a fairly significant role in explaining bank failures. The pseudo R2 falls dramatically and the model χ2 is significantly lower in model (2).

Table 3.2:

Estimation Results—U.S. Commercial Bank Distress in the Southwest

article image
Notes: The dependent variable takes the value of one if a bank’s coverage ratio is less than or equal to zero and the value zero otherwise. The z statistics are given in parenthesis and are based on robust (Huber and White) standard errors which account for correlated observations in grouped data, P|z| is the test of the underlying coefficient being zero. One, two, and three asterisks indicate significance levels of 1,5 and 10 percent respectively.
Table 3.3.

Estimation Results—U.S. Commercial Bank Failures in the Southwest (CAMEL)-

article image
Notes: The dependent variable takes the value of one if a bank is intervened at time t + 1 and the value zero otherwise. Probability of failure is estimated by fixed-effects logit. The odds ratio and hazard ratio depict exponentiated coefficients. The z statistics are given in parenthesis and are based on robust (Huber and White) standard errors which account for correlated observations in grouped data. P|z| is the test of the underlying coefficient being zero. One, two, and three asterisks indicate significance levels of 1, 5 and 10 percent, respectively.

The predicted (asset-weighted) probability of banking system failure based on the full CAMEL model (model (1)) is depicted as ph_caml, while model (2), excluding NPLA and CA, is represented by ph_cam2 in Figure 14.3. Model (1) seems to provide a clearer signal (though somewhat overstated) of impending crisis. The expected failure times are not significantly different in both models (et_caml and et_cam2).

Figure 14.3.
Figure 14.3.

U.S. Banking System Failures - Southwest (CAMEL)

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

Empirical results for the U.S. Northeast

In the Northeast, 261 commercial banks in the states of New Hampshire, Connecticut and Massachusetts were examined, of which 41 banks were intervened during the period 1986–93.

The empirical findings for bank failures in the Northeast are reported in Table 4.1. Consistent with the findings obtained for the U.S. Southwest, a higher ratio of nonperforming loans to total assets (NPLA) in model (1), and a lower ratio of capital equity to total assets (EQ) in model (2) are associated with a higher probability of failure and a higher hazard rate (reducing the survival time). A one percentage point rise in NPLA would increase the probability of failure by more than 3 ½ percentage points, while a one percentage point increase in EQ would reduce the probability of failure by about ⅓ of a point. Similarly, a one-unit rise in NPLA would increase the hazard of failure by 1.2, while a one-unit increase in EQ would reduce the hazard by 0.8. The estimated p values indicate positive duration dependence.

Table 4.1.

Estimation Results—U. S. Commercial Bank Failures in the Northeast

article image
article image
Notes: The dependent variable takes the value of one if a bank is intervened at time t+1 and the value zero otherwise. Probability of failure is estimated by fixed-effects logit. The odds ratio and hazard ratio depict exponentiated coefficients. The z statistics are given in parenthesis and are based on robust (Huber and White) standard errors which account for correlated observations in grouped data. P|z| is the test of the underlying coefficient being zero. One, two, and three asterisks indicate significance levels of 1, 5 and 10 percent, respectively. Model 4 (a) did not converge. Hence, the micro-macro components were separated.

The results from model (3) based on bank-specific variables and the full model (4) suggest that a higher ratio of commercial real estate loans to total assets (LCOMRE) and residential loans to total assets (LRESI) are associated with a higher probability of failure, and that the former is correlated with a reduced survival time, suggesting (as was the case for Southwestern banks) that real estate loans were quite risky. A higher average yield on loans (LNYTELD) is associated with a lower probability of failure and an increased survival time—suggesting that credit risk may have been priced appropriately.61 A higher level of interbank deposits (DEPIB), large deposits from the public (DEPLGE) and a higher liquidity ratio (SEC) are associated with a lower probability of failure, and with an increased survival time in the cases of DEPIB and SEC. There is no evidence that higher level of insider loans (INSL) increased the probability of failure or reduced banks’ survival time of banks in the U.S. Northeast. An increase in regional economic activity (SPERYCH) lowers the probability of failure and increases the survival time of banks. Lower real interest rates (INTRS), as was evident during most of the period of study, are again surprisingly consistent with a higher probability of failure. Although with a relatively small effect, a higher ratio of banking system loans relative to the region’s personal income (STLNPI) is correlated with an increased survival time—possibly in connection with regulatory forbearance. Regional/macro variables appear to have a significant explanatory power.62

The expected probability of failure based on the coverage ratio (ph_covr), would have overestimated the actual (asset-weighted) probability of banking failures in the Northeast, but it would have given a clear indication that the banking system was rapidly weakening about a year before the peak of the banking problems (Figure 15.1). The overestimation seems to be largely due to the role of nonperforming loans (ph_npla). The expected probability of failure based on bank-specific variables (ph_bs) would have shown a moderate increase prior to the banking problems, but the augmented regional/macro (ph_rm) and banking system variables (ph_f) would have magnified this deterioration—albeit significantly overestimating the actual magnitude of the crisis.

Figure 15.1.
Figure 15.1.
Figure 15.1.

U.S. Banking System Failures - Northeast

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

The expected (asset weighted) failure time based on the coverage ratio (et_covr) shows a significant decline prior to the peak of the crisis and an improvement as banks that fail exit the sample, largely because of nonperforming loans (et_npla). Bank-specific variables (et_bs), and models augmented by regional/macro variables (et_rm) and by the banking sector variable (et_f), suggest similar developments (Figure 15.1).

The number of occurrences of distress during the 1985–92 period is more than three times larger than the number of actual failures. The results from the models based on bank distress (models (5) and (6)) are broadly consistent with the results from the models based on intervention. The chief differences are that an increase in residential loans (LRESI) lengthens the survival time, and the effect of the real interest rate (INTRS) is now not statistically significant. The full model seems to have a significantly greater explanatory power than the one based only on bank-specific variables.

The (asset-weighted) distress occurrences seem to precede and somewhat magnify the (asset-weighted) incidence of failures (Figure 15.2). The estimated probability of distress based on the three models (ph_bs, ph_rm and ph_f) seem to indicate increasing stress in the system before the emergence of the banking crisis, though the augmented models based on macro/regional variables (ph_rm) and the full model including proxies for contagion (ph_f) show a more pronounced deterioration. The expected distress time paths are also generally consistent with those results.

Figure 15.2.
Figure 15.2.

U.S. Banking System Distress - Northeast

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

Basic CAMEL approach: U.S. Northeast

Using a simple CAMEL model, it appears that nonperforming loans (NPLA) and capital (CA) strongly influence banks’ survival time (Table 4.3). The predicted (asset-weighted) probability of banking system failures with all CAMEL variables (ph_caml) magnifies the results obtained from the model that excludes NPLA and CA (ph_cam2) (Figure 15.3). The expected failure times are not substantially different in both models.

Figure 15.3.
Figure 15.3.

U.S. Banking System Failures - Northeast (CAMEL)

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

Table 4.2.

Estimation Results—U.S. Commercial Bank Distress in the Northeast

article image
Notes: The dependent variable takes the value of one if a bank’s coverage ratio is less than or equal to zero and the value zero otherwise. Probability of distress is estimated by fixed-effects logit. The odds ratio and hazard ratio depict exponentiated coefficients. The z statistics are given in parenthesis and are based on robust (Huber and White) standard errors which account for correlated observations in grouped data. P|z| is the test of the underlying coefficient being zero. One, two, and three asterisks indicate significance levels of 1, 5 and 10 percent respectively.
Table 4.3.

Estimation Results—U.S. Commercial Bank Failures in the Northeast (CAMEL)

article image
Notes: The dependent variable takes the value of one if a bank is intervened at time t+1 and the value zero otherwise. Probability of failure is estimated by fixed-effects logit. The odds ratio and hazard ratio depict exponentiated coefficients. The z statistics are given in parenthesis and are based on robust (Huber and White) standard errors which account for correlated observations in grouped data. P|z| is the test of the underlying coefficient being zero. One, two, and three asterisks indicate significance levels of 1, 5 and 10 percent respectively. Model 1(a) did not converge.

Empirical results for California

In California, 562 commercial banks were examined, of which 55 banks were intervened during the period 1986–93. In contrast to the bank failures in the Southwest and the Northeast, most of the banks intervened in California were relatively small and interventions happened generally later.

The empirical findings for bank failures for California are reported in Table 5.1. As in the Southwest and in the Northeast, a higher level of nonperforming loans (NPLA) in model (1) and a lower level of equity (EQ) in model (2) are associated with a higher probability of failure and a reduced survival time, with the effect of NPLA being greater than that from EQ. In California, a one percentage point increase in NPLA would increase the probability of failure by almost 1 ½ percentage points, while a 1 percentage point increase in EQ would reduce the probability of failure by almost ½ of a point. Similarly, a one-unit increase in NPLA would increase the hazard of failure by 1.2, while a one-unit increase in EQ would reduce the hazard by 0.6. The estimated p values indicate positive duration dependence.

Table 5.1.

Estimation Results—U.S. Commercial Bank Failures in California

article image
article image
Notes: The dependent variable takes the value of one if a bank is intervened at time t+1 and the value zero otherwise. Probability of failure is estimated by fixed-effects logit. The odds ratio and hazard ratio depict exponentiated coefficients. The z statistics are given in parenthesis and are based on robust (Huber and White) standard errors which account for correlated observations in grouped data. P|z| is the test of the underlying coefficient being zero. One, two, and three asterisks indicate significance levels of 1,5 and 10 percent, respectively.

The results from model (3) based on bank-specific variables and the full model (4) suggest, as is the case with the other U.S. regions, that higher levels of commercial real state loans (LCOMRE) and residential loans (LRESI) are generally associated with a higher probability of failure. Also similarly to the other two episodes of banking problems in the United States, higher interbank deposits (DEPIB) and large deposits from the public (DEPLGE) are associated with a lower probability of failure, and a higher liquidity ratio (SEC) is correlated with an increased survival time. However, in contrast with the other two regions in the United States, a higher average yield on loans (LNYIELD) is associated with a decreased survival time—suggesting that credit risk was a problem in California. Also in contrast with the other -regions, a higher level of insider loans (INSL), a proxy for moral hazard, is associated with a reduced survival time. The macro/regional effects are similar to the other U.S. regions. An increase in regional economic activity (SPERYCH) increases the survival time of banks while lower real interest rates (INTRS), evident during most of the period of study, are consistent with a high probability of failure. A higher ratio of banking system loans relative to the estate’s personal income (STLNPI) is correlated with a lower probability of failure—albeit having a small effect. Introducing regional/macro and banking sector variables appear to significantly increase the explanatory power of the bank-specific model.

The expected probability of failure based on the coverage ratio (ph_covr), would have overestimated the actual (asset-weighted) probability of banking failures in California, but it would have shown a increase prior to the banking failures, largely on account of the predicted probabilities of failure based on nonperforming loans (ph_npla) (Figure 16.1). The expected probability of failure based on bank-specific variables (ph_bs), as well as the augmented regional/macro (ph_rm) and banking system models (ph_f), would have shown a sharp increase prior to the banking problems.

Figure 16.1.
Figure 16.1.
Figure 16.1.

U.S. Banking System Failures - California

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

In terms of the (asset-weighted) expected failure time for the banking system, the coverage ratio (et_covr) shows a significant decline prior to the peak of the crisis and an improvement as banks that fail exit the sample. Bank-specific variables (et_bs), and models augmented by regional/macro variables (et_rm) and by a banking sector variables (et_f), suggest similar trends (Figure 16.1).

When comparing the results with bank distress (models (5) and (6)), the results are broadly consistent with those obtained from the failure models. The chief differences are that an increase in residential loans (LRESI) now reduces the probability of distress; a higher level of commercial real estate loans (LCOMRE) also reduces the survival time; a higher average yield on loans (LNYTELD) reduces the probability of distress (giving conflicting signals regarding credit risk vis-à-vis the results for intervention); and real interest rates (INTRS) and the ratio of total banking system loans to the state’s personal income (STLNPI) are now not significant. The number of occurrences of distress during the 1985–92 period is more than four times the number of actual failures.

As in the previous cases, the (asset-weighted) distress occurrences seem to precede and significantly magnify the (asset-weighted) incidence of failures (Figure 16.2). The estimated probability of distress based on the three models (ph_bs, ph_rm and ph_f) seem to indicate increasing stress in the system prior to the peak of the bank failures. The expected distress time paths are also broadly consistent with those results.

Figure 16.2.
Figure 16.2.

U.S. Banking System Distress - California

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

Basic CAMEL approach: California

Based on a simple CAMEL model, it appears that nonperforming loans (NPLA) in particular, but also capital (CA), give some of the strongest signals of likely failures (Table 5.3). The predicted (asset-weighted) probability of failure with all CAMEL variables (ph_caml) magnifies somewhat the results obtained from the model that excludes NPLA and CA (ph_cam2) (Figure 16.3). The (asset-weighted) expected failure times in both models show somewhat different paths: while et_caml based on the full model increases after declining briefly prior to the peak of the bank failures, et_cam2 which excludes nonperforming loans and capital is flatter throughout most of the period.

Figure 16.3.
Figure 16.3.

U.S. Banking System Failures -California (CAMEL)

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

Table 5.2.

Estimation Results—U.S. Commercial Bank Distress in California

article image
Notes: The dependent variable takes the value of one if a bank’s coverage ratio is less than or equal to zero and the value zero otherwise. Probability of distress is estimated by fixed-effects logit. The odds ratio and hazard ratio depict exponentiated coefficients. The z statistics are given in parenthesis and are based on robust (Huber and White) standard errors which account for correlated observations in grouped data. P|z| is the test of the underlying coefficient being zero. One, two, and three asterisks indicate significance levels of 1,5 and 10 percent, respectively.
Table 5.3.

Estimation Results—U.S. CommercialBank Failures in California (CAMEL)

article image
Notes: The dependent variable takes the value of one if a bank is intervened at time t+1 and the value zero otherwise. Probability of failure is estimated by fixed-effects logit. The odds ratio and hazard ratio depict exponentiated coefficients. The z statistics are given in parenthesis and are based on robust (Huber and White) standard errors which account for correlated observations in grouped data. P|z| is the test of the underlying coefficient being zero. One, two, and three asterisks indicate significance levels of 1, 5 and 10 percent, respectively.

Empirical results for Mexico

In Mexico, 31 domestic commercial banks (excluding foreign banks) were examined based on quarterly data for the period 1992-Q1 to 1995-Q3, of which 16 banks were intervened by the authorities during 1994–95.63 The banks’ maximum survival time is (right) censored at 45 months.

The empirical findings for bank failures for Mexico are reported in Table 6.1. As in all the previous cases examined, a higher level of nonperforming loans (NPLA) in model (1) and a lower level of equity (EQ) in model (2) are associated with a higher probability of failure and a reduced survival time. In Mexico, a one percentage point increase in NPLA would increase the probability of failure by almost 2 percentage points, while a 1 percentage point increase in EQ would reduce the probability of failure by about ⅓ of a point. Similarly, a one-unit rise in NPLA would increase the hazard of failure by 1.2, while a one-unit increase in EQ would reduce the hazard by 0.6. As in the other cases considered, the estimated p values indicate positive duration dependence.

Table 6.1.

Estimation Results—Bank Failures in Mexico

article image
article image
Notes: The dependent variable takes the value of one if a bank is intervened at time t+1 and the value zero otherwise. Probability of failure is estimated by fixed-effects logit. The odds ratio and hazard ratio depict exponentiated coefficients. The z statistics are given in parenthesis and are based on robust (Huber and White) standard errors which account for correlated observations in grouped data. P|z| is the test of the underlying coefficient being zero. One, two, and three asterisks indicate significance levels of 1, 5 and 10 percent, respectively.

The results from model (3) based on bank-specific variables and the full model (4) suggest that higher residential loans (LRESI) and higher agricultural loans (LAGR) are associated with a reduced survival time. A higher ratio of loans to assets (LAS), a proxy for credit risk, is correlated with a higher probability of failure. Higher deposits from the public (DEPPUB) increase banks’ survival time, while a higher proportion of liquid assets (SEC) reduces the probability of failure. Higher interbank deposits (DEPIB) seem to increase the probability of failure—this effect may be related to size, since the largest banks (which received government assistance toward the end of the period of study) presumably had the greatest access to the interbank market. The ratio of interest and fees to assets (INTAS), a proxy for moral hazard, was not significant. A depreciating domestic currency (an increase in DELEX) would reduce the banks’ survival time, while real interest rates (INTRS) were not significant. A higher ratio of banking system loans relative to GDP (BSLNGDP) is associated with a reduced survival time and, in contrast with the previous cases examined, it has a fairly substantial effect—which would be consistent with the hypothesis of contagion.64 The model’s statistics seem to improve significantly with the inclusion of macro and banking system variables.

The expected (asset-weighted) probability of banking system failures based on the coverage ratio (ph_covr) shows a gradual, but clear increase beginning in early-1993 (Figure 17.1)—close to two years before the actual crisis. Although ph_covr underestimates the actual (asset-weighted) bank failures, this effect occurs largely in 1995-Q4 when the two major Mexican banks received government financial support in the form of purchases of some of their bad loans. The predicted probabilities of failure resulting from nonperforming loans (ph_npla) seem to give a clearer signal than if based on equity capital (ph_eq). The expected (asset-weighted) probability of banking system failures based on the bank-specific model (ph_bs), as well as the augmented regional/macro (ph_rm) and banking system models (ph_f), would have shown a significant increase shortly before the banking crisis.

Figure 17.1.
Figure 17.1.
Figure 17.1.

Banking System Failures - Mexico

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

The expected (asset weighted) failure time based on the coverage ratio (et_covr) is quite flat during most of the period, largely due to the effect of equity capital (et_eq) since nonperforming loans (et_npla) alone would have shown a continuous decline throughout most of the period of study (Figure 17.1). Bank-specific variables (et_bs) and the model augmented by regional/macro variables (et_rm) show some deterioration prior to the actual crisis, but the decline is fairly steady if based on the full model (et_f) which includes the effects of possible contagion.

When compared with bank distress, the results (models (5) and (6)) are generally consistent with those from the failure models. The main differences are that an increase in residential loans (LRESI) also increases the probability of distress; a higher proportion of non-securitized loans (LNONSEC) appear to increase banks’ survival time; loans to assets (LAS), public deposits (DEPPUB) and the exchange rate (DELEX) are no longer significant; and higher interest rates (INTRS) increase the probability of distress. The change in the effects coming from DELEX and INTRS is particularly interesting because bank failures seem to have been affected significantly by the currency depreciation, but not bank distress which was impacted more directly by high real interest rates. The number of occurrences of distress during the period 1992-Q1 to 1995-Q3 is close to four times the number of actual failures.65

As with the previous cases examined, the (asset-weighted) distress occurrences seem to precede and significantly magnify the (asset-weighted) incidence of failures (Figure 17.2). The estimated probability of distress based on the three models (phbs, phrm and ph_f) suggest increasing stress in the system prior to the banking crisis. The expected distress time paths are also broadly consistent with those results.

Figure 17.2.
Figure 17.2.

Banking System Distress - Mexico

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

Basic CAMEL approach: Mexico

Based on a simple CAMEL model, it appears that nonperforming loans (NPLA) would have given the strongest signal of likely failures (Table 6.3). However, the overall fit of model (1) is only moderately better than model (2) which excludes capital and nonperforming loans. The predicted (asset-weighted) probability of failure with all CAMEL variables (ph_caml) is similar to the probability predicted by the model that excludes capital and nonperforming loans (ph_cam2) (Figure 17.3). The expected (asset-weighted) failure times in both models show different paths: while et_caml which includes all the variables increases steadily, without nonperforming loans and capital et_cam2 is flatter and shows a moderate decline prior to the crisis.66

Figure 17.3.
Figure 17.3.

Banking System Failures - Mexico (CAMEL)

Citation: IMF Working Papers 1999, 033; 10.5089/9781451845167.001.A001

Table 6.2.

Estimation Results—Bank Distress in Mexico

article image
Notes: The dependent variable takes the value of one if a bank’s coverage ratio is less than or equal to 1.5 and the value zero otherwise. Probability of distress is estimated by fixed-effects logit. The odds ratio and hazard ratio depict exponentiated coefficients. The z statistics are given in parenthesis and are based on robust (Huber and White) standard errors which account for correlated observations in grouped data. P|z| is the test of the underlying coefficient being zero. One, two, and three asterisks indicate significance levels of 1, 5 and 10 percent, respectively.