A review of financial distress research methods and recommendations for future research.

Ward, Terry J.

INTRODUCTION

Since Beaver's seminal study in 1966, substantial research has been published concerning the ability of accounting information to predict financial distress (bankruptcy). This area of research uses financial distress as a criterion for evaluating the usefulness of accounting information. Although researchers have used this predictive ability criterion to test many accounting variables, much of the empirical research in this area has centered around the use of a predictive ability criterion to compare the usefulness of accrual and cash flow information.

The underlying theory justifying financial distress as a criterion for testing accounting information is based on a future cash flows emphasis. The Financial Accounting Standards Board (FASB) adopted a cash flows concept approach in the development of the conceptual framework, establishing the ability to help investors and creditors predict future cash flows as the primary criterion for evaluating the usefulness of accounting information. Since future cash flows are constructs, researchers developed proxies for future cash flows; financial distress was chosen by a number of researchers as an acceptable proxy. The original studies in this area of research used a dichotomous failed and nonfailed proxy for financial distress, while studies since the mid-1970s used primarily a dichotomous bankrupt and nonbankrupt response measure for financial distress.

This study provides a methodological review of prior financial distress research. The purpose of this paper is to provide a framework for improving the validity of future financial distress studies by discussing the limitations found in prior financial distress research and making recommendations concerning future accounting research. (1)

LANDMARK FINANCIAL DISTRESS STUDIES

The prediction of financial distress (bankruptcy) has been the subject of a substantial amount of research. Most of the financial distress studies fall into one of two categories based on the study's primary objective. Either the primary purpose of the study is to use financial distress as a criterion variable to test the importance of accounting information or the primary purpose is to develop the most accurate model possible to help users predict bankruptcy. The two seminal works of Altman (1968) and Beaver (1966) are typical of studies in each of these two categories.

Applied Bankruptcy Models

Altman attempted to improve conventional ratio analysis by developing a multivariate model using multiple discriminant analysis (MDA). His Z-score model is still used today to predict bankruptcy. The primary objective of studies such as Altman's "landmark study" is to develop the most accurate model or tool to enable users to effectively, and economically, predict bankrupt firms. (2) These researchers are primarily interested in the predictive accuracy of the model and how the model compares to a naive model. The information content of a specific variable is not their primary concern. (3)

Ohlson (1980) extended Altman's methodology three ways: (1) he used multivariate logit analysis to determine probabilistic estimates of failure instead of a simple deterministic classification of a firm as bankrupt or nonbankrupt; (2) he collected a larger, and more representative, sample of 105 bankrupt firms and 2,058 nonbankrupt firms; and (3) he only included firms with published financial data released prior to the announcement of bankruptcy to reduce the amount of sampling bias. Ohlson found that firm size was a significant negative predictor of bankruptcy; bankrupt firms tend to be smaller than nonbankrupt firms. This result shows the need to control for firm size in financial distress studies. Ohlson also showed that, for some firms, the financial reports for the preceding year are issued after the announcement of bankruptcy. Thus, the financial reports include information about a firm's bankruptcy, resulting in the overstatement of the predictive power of the accounting models. To correct this problem, the researcher must either drop these firms from the sample or substitute the financial statements from the previous year for the most current year of interest.

Zavgren (1985) also used logistic analysis in a similar study of forty-five bankrupt and forty-five nonbankrupt firms. Unlike Ohlson, Zavgren theoretically justified the inclusion of seven variables found to load highest on separate factors. Both Zavgren and Ohlson found that their bankruptcy models could be used to predict a firm's likelihood of failure for up to five years in advance with reasonable accuracy. (4)

Financial Distress--Naive Cash Flow Variable

The following studies analyzed the usefulness of cash flow variables using a predictive ability criterion. However, each of these studies used a crude approximation of cash flow from operations, net income plus depreciation and amortization.

Beaver (1966) conducted a univariate analysis of the ability of thirty ratios (four of which were cash flow ratios) to distinguish between seventy-nine failed and seventy-nine nonfailed firms. Beaver matched the failed firms with the nonfailed firms based on industry and asset size. He defined a failed firm as one that had experienced one of the following events: bankruptcy, bond default, overdrawn bank account, or nonpayment of a preferred stock dividend. Beaver used simple classification techniques based on various cutoff points to distinguish between the failed and nonfailed firms. He found that: (1) accounting information can be used to predict failure as early as five years before failure; (2) different ratios predict with different degrees of accuracy; (3) accounting information predicts nonfailure better than failure; and (4) the ratio cash flow/total debt was the best performing variable, with an eighty-seven percent classification rate for year one and a classification rate of seventy-eight percent five years prior to failure. However, Beaver's greatest contribution was his suggestion that financial distress could be used as a framework for evaluating the usefulness (predictive ability) of accounting information (Altman et al., 1981).

Deakin (1972) extended the studies by Altman (1968) and Beaver by incorporating the fourteen "strongest" variables suggested by Beaver and the multivariate methodology recommended by Altman. The author obtained a sample of thirty-two failed and thirty-two nonfailed firms. Deakin matched the firms based on industry, asset size, and year of financial data. Like Beaver, Deakin sampled the firms based on a failed and nonfailed classification scheme; a failed firm was one that was bankrupt, insolvent, or liquidated. Using multiple discriminant analysis (MDA), Deakin found that the ratio cash flow/total debt was very important to the discriminant model. The fourteen variable model predicted failure as far as three years in advance with fairly high accuracy. Similar to Beaver and Deakin, Blum (1974) used a failed and nonfailed classification to determine the predictive ability of financial ratios. However, the primary purpose of Blum's study was to "develop a Failing-Company Model to aid the Antitrust Division of the Justice Department in assessing the probability of business failure" (Altman et al. 1981, p. 227) using linear MDA. Blum defined a failed firm as any firm meeting one of the following criteria: (1) failure to pay debts when due, (2) debt accommodation agreement with creditors to reduce debts, or (3) occurrence of bankruptcy. The only cash flow variable in the model, cash flow/total debt, was generally ranked high in the predictive models.

Criticism of Failed and Nonfailed Measure of Financial Distress

The prior studies by Beaver, Deakin, and Blum were criticized because of the heterogeneous failed and nonfailed sample selection scheme used. In a two-group failure classification, firms within a group should be homogeneous and representative of the population of failed enterprises (Altman et al., 1981). Thus, studies by Altman, Haldeman, and Narayanan (1977), Norton and Smith (1979), and Mensah (1983) tested the ability of financial ratios and cash flows to predict bankrupt and nonbankrupt firms. The authors used various stepwise linear and quadratic MDA models. Except for Altman, Haldeman, and Narayanan, the results of these bankruptcy studies validated the results of prior studies that used a broader definition of failure. Mensah found cash flow/net worth to be the most important ratio, while cash flow/total assets and cash flow/total debt were both found important by Norton and Smith. (5)

FINANCIAL DISTRESS--REFINED OPERATING CASH FLOW VARIABLE

The following studies tested a more refined measure of operating cash flow, with the most recent studies also testing additional cash flow variables other than operating cash flow. In each of these studies, the primary purpose was to test the usefulness of cash flow information using a predictive ability criterion.

The study by Largay and Stickney (1980) of the 1975 bankruptcy of W.T. Grant was the catalyst behind the explosion of cash flow studies during the 1980s. In this study, Largay and Stickney observed the trends of certain accrual ratios (profitability, turnover, liquidity, and solvency ratios) and cash flow from operations (CFO) for ten years preceding bankruptcy. The authors found that the company's CFO provided a more accurate and timely signal of W.T. Grant's eventual failure than traditional accrual ratios.

However, two studies by Casey and Bartczak (C&B) (1984, 1985) appeared to contradict the study by Largay and Stickney. In the 1984 study, C&B tested whether CFO or accrual ratios could best predict bankruptcy. The authors obtained a matched-pair sample (matched on industry) of bankrupt and nonbankrupt firms for the period 1971-1982. The sampling scheme resulted in a total sample of sixty bankrupt and 230 nonbankrupt firms. Half the firms (thirty bankrupt and 165 nonbankrupt firms) were used to develop the classification models, while the other firms were used as a holdout group. The authors generated two groups of models with lag periods of one to five years before bankruptcy. The first group of models was composed of univariate cash flow models for each lag period. The cash flow variables tested were: CFO (working capital provided by operations, plus or minus changes in the noncash working capital accounts); CFO divided by current liabilities; and CFO divided by total liabilities. The other group of models was composed of multivariate accrual models generated from linear MDA. Each accrual model included the following six accrual ratios: net income/total assets, cash/total assets, current assets/current liabilities, net sales/current assets, current assets/total assets, and total liabilities/owners' equity. C&B found that neither cash flow variable had higher classification rates than the combined six accrual ratios.

In their subsequent study, C&B (1985) attempted to determine whether CFO could increase the accuracy of accrual-based ratios to distinguish between bankrupt and nonbankrupt firms. The study was essentially the same as the previous one except the approach was different; the objective was to determine if CFO had incremental predictive ability. The authors developed accrual, cash flow, and mixed (cash flow variables added to the accrual models) models for lag periods of one to five years. To obtain test statistics on each variable, C&B also generated mixed models using conditional stepwise logit analysis. (6) C&B concluded that operating cash flows "do not provide incremental predictive power over accrual-based ratios" (p. 395).

The research by C&B seems to validate the FASB's position that accrual information better predicts future cash flows, but contradicts the FASB position that cash flow information taken together with accrual information should better predict future cash flows. However, C&B only tested CFO, and they used working capital from operations (WCFO) to calculate CFO. Drtina and Largay (1985) found that using WCFO to calculate CFO might create confounded results because of the inconsistencies among the way firms define WCFO and the diversity in reporting practices. C&B also failed to control for firm size, either by matching or by incorporating a size variable in the models.

In a study similar to the studies of C&B, Gentry, Newbold, and Whitford (GN&W) (1985) compared thirty-three bankrupt and thirty-three nonbankrupt firms to determine if funds flow information could predict bankruptcy. GN&W matched the firms on size, industry, and sales. The authors used a separate sample of twenty-three weak and twenty-three nonweak firms to validate the models instead of a separate holdout group of bankrupt and nonbankrupt firms. The weak firms were identified by the "creditwatch list issued by the Wells Fargo Bank" and "from various financial services who rate candidates for financial failure" (p. 158). Instead of testing CFO, the authors tested seven funds flows (each scaled by total net cash flows) based on Helfert's (1982) "cash-based funds flow model" and six accrual ratios. The seven funds flows tested were operations, working capital, financial, fixed coverage expenses, capital expenditures, dividends, and other asset and liability flows. The authors used various statistical techniques to generate the models (individual and mixed models) to test the funds flow and accrual ratios (MDA, probit, and logit techniques); however, the logit models provided the best results. Gentry, Newbold, and Whitford found that only the dividend funds flow component was significant both one year and three years before bankruptcy. They also found that the components of CFO (funds from operations, working capital, and fixed coverage expenses) failed to improve the classification of failed and non-failed firms. However, the authors did not specifically test CFO.

In a subsequent study, GN&W (1987) extended the 1985 study by comparing the predictive ability of accrual ratios and funds flow components. The authors broke up the net working capital from operations variable into five funds flow components (changes in accounts). The five funds flow components were accounts receivable, inventory, other current assets, accounts payable, and other current liabilities. The authors also tested six accrual ratios common in prior financial distress studies. GN&W found that outflows of certain components indicated a healthy company. Inventories, dividends, and receivables were inversely related to bankruptcy; nonfailed companies showed net outflows for inventories, receivables, and dividends while failed companies showed inflows leading up to the bankruptcy. Thus, this study offers evidence that the particular inflows and outflows of funds may be more important in predicting bankruptcies than a net cash flow such as CFO.

Aziz, Emanuel, and Lawson (AE&L) (1988) tested the predictive ability of six cash flow variables on a sample of forty-nine bankrupt and forty-nine nonbankrupt firms for the period 1971-82 (the models were generated with data collected from 1966-1981). The cash flows tested were operating cash flow, net capital investment, taxes paid, liquidity change, stockholders' cash flow, and lender cash flows. AE&L selected the cash flows based on a cash flow identity developed by Lawson (1978, 1985) and scaled the cash flows by the book value of the firm to avoid the problem of heteroscedasticity. Lawson's cash flow identity differs from Helfert's (1982) cash-based model in that Lawson's model is for "firm valuation," while Helfert's model is to provide analysis by "area of management attention" (AE&L, p. 419-420). MDA and logistic regression were used to generate the models (lagged for 1-5 years) to test the cash flow variables. AE&L found that the cash flows, taxes paid, operating cash flow, and lender cash flow were significant predictors in two of the five years before bankruptcy, with taxes paid significant in all five years. A comparison of Lawson's model with the GN&W (1985) and Altman, Haldeman, and Narayanan (1977) (ZETA) models indicated that Lawson's model was superior to the GN&W model for the only year compared and was superior to the ZETA model for three or more years before bankruptcy.

Aziz and Lawson (A&L) (1989) extended the research of AE&L in two ways: (1) they tested the incremental usefulness of cash flow information by combining the cash flow variables with the five accrual ratios in Altman's Z-score model and (2) they used a holdout sample of twenty-six bankrupt and sixty-seven nonbankrupt firms to test the validity of the models generated. The authors also extended prior research by incorporating formal hypothesis testing. Based on classification accuracy, A&L confirmed the results of C&B and GN&W, concluding that cash flow based models do not improve on the existing model's (accrual) overall accuracy. However, concerning the holdout sample, cash flow based and mixed models exhibited superior predictive accuracy, thus suggesting cash flows may be more stable across time.

The results of the above studies are surprising, since the main stated benefit of cash flows is their incremental usefulness in helping accrual information to predict insolvency (Staubus, 1989). The only study showing cash flow based components have incremental predictive content (Gentry, Newbold, and Whitford, 1987) found that certain changes in accounts that comprise working capital have predictive content.

LIMITATIONS OF PRIOR FINANCIAL DISTRESS STUDIES

This section contains a discussion of the limitations of prior financial distress studies that compared the predictive ability of accrual and cash flow information. Where appropriate, methodology studies addressing these limitations are also discussed.

Response Measure for Financial Distress

A limitation of prior financial distress studies is the use of a dichotomous financial distress measure. The use of a dichotomous classification of financial distress is an overly simple representation of the financial distress process and is unlikely to capture the true underlying construct. The financial distress of a firm is an unobservable continuum. Firms are not simply bankrupt or nonbankrupt but possess certain degrees of financial distress that vary from day to day and period to period. However, since researchers do not have the capability to observe this continuum, they select events to operationalize this construct. The finance literature stresses the belief that many events indicate different degrees of financial distress, and companies may go through many of these events before bankruptcy occurs (e.g., Dewing (1953), Donaldson (1969), and Newton (1975)).

A truer test of the usefulness of a financial distress model would be the model's ability to distinguish between marginally distressed firms, not just between healthy firms and firms in very serious financial distress (Jones, 1987). Events such as loan/interest default and failure to pay dividends may be of more interest to investors and creditors because the ability to predict these events one year in advance would provide the users with an earlier warning signal than predicting bankruptcy one year in advance.

Researchers have also questioned the use of bankruptcy as a proxy for financial distress because bankruptcy is a legal event and not an economic event (Dietrich, 1984). Financial distress results from economic occurrences. Only economic events should truly capture the level of financial distress of a firm. Legal recognition of bankruptcy may occur sometime after the firm is economically insolvent, or occur even though the company is not economically insolvent. Also, the economic conditions of bankrupt firms are likely not similar to other types of distressed firms. Thus, using a legal event as a proxy for economic conditions may produce misleading results. Even the firms selected by legal status may form a heterogeneous set because some firms voluntarily choose bankruptcy, and others do not. Thus, the economic conditions of these firms may be quite different (p. 84).

Gilbert, Menon, and Schwartz (GM&S) (1990) addressed the appropriateness of using a bankrupt and nonbankrupt response scale to test the importance of accounting information by replicating the studies of C&B (1985) and Altman (1968) using two separate samples of firms. The authors found that operating cash flow could significantly distinguish between bankrupt and distressed firms. However, a healthy and bankrupt model performed poorly when used to distinguish bankrupt firms from distressed firms (defined as firms having consecutive losses) for the holdout sample. This result suggests that cash flow information may be more useful in distinguishing between events of financial distress other than bankruptcy. However, GM&S failed to: (1) look at other economic events of financial distress such as loan defaults and failure to pay dividends; (2) develop multi-state models of distress to better capture the predictive ability of cash flow and accrual information; and (3) control for the size of the firms, either by matching or by including size as an independent variable.

Lau (1987) corrected for many methodological limitations of the prior dichotomous predictive models by using a five-state response scale to approximate the continuum of corporate financial health instead of the conventional bankrupt and nonbankrupt dichotomy. The states included: (1) financial stability, (2) omitting or reducing dividend payments, (3) default of loan interest or principal payments, (4) protection under Chapter X or XI of the Bankruptcy Act, and (5) bankruptcy and liquidation.

Lau considered these states of financial distress to be on an ordinal scale. According to Lau, "states one to four are states of increasing severity of financial distress" (1987, p. 128). However, a limitation of Lau's study is that the statistical technique she used did not incorporate the ordinal structure of her dependent variable into her model; Lau used a nominal based logit model. For ordinally scaled dependent variables, nominal logistic models are inferior to ordinal logistic regression models (Agresti, 1984). According to Agresti (1984), ordinal procedures have the following advantages over nominal procedures:

(1) Ordinal methods have greater power.

(2) Ordinal data description is based on measures similar to those (e.g., correlations, slopes, means) used in ordinary regression for continuous variables.

(3) Ordinal analyses can be used in a greater variety of models, most of which have simpler interpretations than the standard models for nominal variables.

(4) Ordinal models can be applied in settings where the standard nominal models are trivial or else have too many parameters to be tested for goodness of fit (p. 3).

Lau attempted to overcome these limitations by employing a rank-scoring rule. However, the rank-scoring rule did not allow for statistical tests of significance for each variable. An ordinal logistic procedure is not limited by these problems. (7)

The sampling technique used by Lau to identify the loan default firms also created additional sample selection bias. Lau used the SEC 10-Ks of firms that had filed for bankruptcy or had C-rated bonds during 1977 to 1980 to identify firms that defaulted on loans during 1976 or 1977. Thus, firms that recover from financial distress after experiencing a loan default were excluded from her original (1976) and holdout (1977) samples.

Ward (1994) extended Lau's multi-state methodology by using ordinal logistic regression to generate the prediction models for a four-state ordinal response scale. This study modeled the severity of a firm's financial distress by using four ordinal states of financial distress as the dependent (response) variable and by employing ordinal logistic regression. The study did not include firms that declare bankruptcy and then liquidate as a fifth state of financial distress because the authors identified only four liquidation firms for the developmental sample. The ordinal states of financial distress (dependent variable) used in this study were:

DIST = 0 if firm was healthy (no event of financial distress during 1988 or 1989),

1 if firm experienced a greater than forty percent reduction in cash dividend per share during 1988 or 1989 after a history of successive cash dividends per share paid,

2 if firm experienced a loan principal/interest default or debt accommodation (extension of cash payment schedules, reduction in principal, or reduced interest rates) during 1988 or 1989, and

3 if firm filed, or was forced to file, for Chapter XI protection during 1988 or 1989.

Firms used to generate the ordinal four-state regression models were not pooled across many years but were all selected from one year, 1988 (developmental sample). The study used financial data for 1984/85 (year three models), 1985/86 (year two models), and 1986/87 (year one models) to generate financial distress regression models from the developmental sample of 1988 firms.

Ward and Foster (1997) subsequently tested whether or not using the inability of a firm to pay debts when due, loan default/accommodation, as a response measure produces different results than using legal bankruptcy as the response measure. They developed two financial distress prediction models from data using developmental samples of 1988 and 1989 firms. Each model was run three times, with data lagged one, two, and three years before the event. The two models differed only in the way the dependent variable was defined. The first model used a legal definition of bankruptcy, healthy versus bankrupt, as the dependent variable, while the second model used inability to pay debts when due, healthy versus loan default/debt accommodation, as the dependent variable.

The study's results suggested that a loan default/accommodation response measure was more valid response measure than bankruptcy for determining information useful to lenders in predicting future insolvency (inability to pay obligations when due) of a firm. The authors concluded that future researchers use either a dichotomous loan default/accommodation dependent variable to generate the prediction models, or sample both loan default/accommodation and bankrupt firms to generate financial distress prediction models.

Violations of Statistical Assumptions

Most financial distress studies violated important assumptions of the statistical techniques used. MDA assumes that the predictor variables are randomly drawn and normally distributed, and linear MDA also assumes equal variance/covariance matrices for each group (Altman et al., 1981). Many researchers failed to test these assumptions. Those researchers who did test for violations of MDA assumptions found that these two assumptions were usually violated; financial ratios do not tend to be normally distributed.

Logistic procedures are not restricted by assumptions concerning the predictor variables. Thus, they are generally preferred over MDA (Press and Wilson, 1978). However, logistic procedures assume that the response variables are randomly drawn (Altman et al., 1981) and require sample sizes of at least 10(S+1) to generate unbiased estimates, where S is the number of predictor variables in the model (McFadden, 1974; and Freeman, 1987).

Sufficient Sample Size

Noreen (1988) used simulations to compare the performance of probit and ordinary least squares regression (OLS) models in predicting bankruptcy. (8) Noreen found that, for a sample size of 10(S+1), probit incorrectly rejected the null hypothesis of no effect at a rate twice the normal level and greater than OLS. However, the results reversed when the sample size was increased to 20(S+1) (OLS incorrectly rejected the null hypothesis more than probit).

Stone and Rasp (1991) subsequently extended the research of Noreen by using simulated data and actual data to compare the performance of OLS regression models and logit regression models. Like Noreen, the authors found that a sample size of 10(S+1) led to biased logistic estimators; these logistic estimators were also more biased than the OLS estimators. However, the authors found that the chi-square statistics for the overall models incorrectly tended to reject the null hypothesis of no effect while the individual parameter test statistics (t-test statistics) tended to be conservatively biased against rejecting the null hypothesis.

All of the early cash flow studies that used logistic procedures, except the studies of Casey and Bartczak (1984, 1985), violated the requirement of a sufficient sample size for some, if not all, of their logistic models. The violations were especially severe for the study by GN&W (1987) that found funds flows were significant predictors of bankruptcy. GN&W needed a minimum sample size of 180 firms to meet the sample size requirement of 10(S+1) for their mixed models. Since the sample size used by GN&W is substantially less than the recommended sample size, their results could be affected by the bias caused by using small sample sizes to generate the logistic parameter estimates.

Choice-Base Sampling Bias

Financial distress studies may also suffer from stratification bias because the researchers nonrandomly selected the bankrupt and nonbankrupt firms. Instead of randomly selecting a sample of firms from the population and then identifying the bankrupt and nonbankrupt firms, researchers first identified the bankrupt firms to obtain a sufficient sample size of these firms. Then, the researchers matched the bankrupt firms with nonbankrupt firms, normally on a one to one basis. Consequently, the samples were not representative of the population (bankrupt firms were oversampled). This sample bias is often called "choice-base sampling bias" (Manski and Lerman, 1977).

Zmijewski (1984) empirically demonstrated the existence of choice-based sampling bias in a binary probit bankruptcy model. He also showed that a weighted probit model eliminated most of the sampling bias and that sampling a larger percentage of nonbankrupt firms lessened the bias. Zmijewski also found that the bias did not change the statistical results or overall accuracy of the model.

Manski and McFadden (1981) and Cosslett (1981) illustrate weighted procedures to correct for choice-based sampling bias in binary logit models. However, the impact of choice-base sampling bias in logit models depends on the type of statistical model used. A weighted model is needed to correct for the bias in the conditional logit models illustrated by Manski and McFadden and Cosslett. However, Maddala (1991) shows that a weighted model is not needed for the binary logit model illustrated by Berkson (1944). For a logit model, choice-base bias is not present in the parameter estimates, only the intercept is affected. Thus, the bias simply affects classification accuracy and not statistical tests of parameter estimates. Maddala recommends a simple adjustment of the intercept to obtain accurate classifications.

Due to stratified sampling, all the bankruptcy studies using logistic (probit) procedures violated the assumption of random selection of the response variables. However, none of the probit studies used weighted models to correct for the choice-based sampling bias, and only Casey and Bartczak (1984, 1985) sampled a larger percentage of nonbankrupt firms.

Pooling of Firms Across Time and Selection of Holdout Sample

A third limitation of prior bankruptcy studies is that the researchers pooled firms across a large period to obtain enough bankrupt firms. Conditions probably are not stable across the different years used to develop the samples. Thus, sampling across years tends to increase the variation in the sample, resulting in a lower likelihood of finding significant results. Also, studies that used holdout samples to verify the predictive accuracy of their models selected the holdout group of firms from the same period used to develop the models. These ex post discriminations are true predictions only if stationarity exists. Otherwise, one should draw the holdout sample from a future period distinct from the original sample period. Evidence of ex ante predictive power requires intertemporal validation and not just cross validation (Altman et al., 1981).

Mensah (1984) found that financial distress prediction models were not stationary across time, thus questioning the practice of pooling bankrupt firms across wide periods of time. This result also indicates that one must obtain the holdout sample from outside the period used to develop the model to get an accurate picture of the predictive ability of the model.

Period Used to Obtain Data

The period used to generate the predictor models is also very important. Cash flows and accruals differ because various accrual methods create a difference in the timing of the recognition of revenues and expenses. This difference should have grown as the profession has moved closer to the all-inclusive definition of income. Using factor analysis, Gombola and Ketz (1983) found that cash flow variables scaled by total debt, total assets, and equity loaded on a separate factor from accrual variables after the mid 1970s, thus suggesting that cash flow variables provided information not in other financial data. Gombola et al. (1987) subsequently tested whether the failure of prior studies to obtain sampling data points after the mid 1970s could be confounding the results of these studies. The authors divided their sample into pre-1972 and post-1972 sampling groups. They concluded that cash flow variables were not more relevant in the post-1972 model. However, their samples may have been too small to detect a difference, and the authors failed to develop models from the post-1981 period.

In a study concerning the intertemporal divergency among operating cash flow, working capital, and income from operations, Franz and Thies (1988) found that income and working capital from operations had diminished as a percentage of operating cash flow over the period from 1967 to 1985. Their results also showed that most of the decrease in income as a percentage of cash flow occurred from 1981 to 1985 (decrease from .5099 to .3493 versus a decrease of .5519 to .5099 for the period from 1967 to 1981, respectively) (1988, p. 24). This increasing gap between income and cash flow suggests that cash flows may provide additional information in predicting financial distress, and that models developed from the post-1981 period are more likely to detect the incremental usefulness of cash flows as predictors of financial distress. However, all of the previous cash flow studies generated models with at least half their data points falling before 1975, and all of the studies' models were developed from pre-1982 data.

CONCLUDING REMARKS--RECOMMENDATIONS

The above limitations of prior financial distress cash flow studies indicate the need for additional research. However, the validity of financial distress studies would be improved if researchers incorporate the following recommendations in future studies:

(1) Obtain sample sizes of at least 10(S+1), and possibly 20(S+1) for all logit (probit) models tested to lower the amount of sampling bias.

(2) For binary studies, use logit regression to eliminate the effects of choice-base sampling bias on parameter estimates. If probit or conditional logit is used, the researcher should incorporate weighted probit and weighted conditional logit models to limit the effects of choice-base sampling bias in binary models.

(3) Sample so that firms used to generate the predictive models are not pooled across a large period but are selected from one year (or as short a period as possible).

(4) Holdout firms should be obtained from a year (period) other than the year (period) used to generate the models.

(5) For cash flow research, either use actual cash flows as reported on the statement of cash flows or estimate cash flows from post-1981 data.

(6) For dichotomous financial distress studies, replicate research using dichotomous measures of financial distress other than bankrupt and nonbankrupt, especially loan default/ accommodation.

(7) Develop ordinal multi-state measures of financial distress whenever possible.

The development of an ordinal multi-state response variable is a natural extension of prior financial distress research. Prior research has failed to concentrate on obtaining a better measure of financial distress; researchers assumed that bankruptcy was an appropriate response measure for evaluating the usefulness of accounting information. However, as stated earlier, the use of a dichotomous measure, especially bankruptcy, as the proxy for financial distress suffers from many weaknesses. The severity of financial distress of a firm can be modeled using ordinal levels of financial distress. Whether bankruptcy is the most appropriate proxy for financial distress is an empirical question that needs to be addressed.

If a dichotomous response is used to evaluate accounting information, a more appropriate dichotomous response may be loan default/accommodation versus a healthy state. The use of loan default/accommodation as the criterion variable for evaluating the usefulness of accounting information also makes more sense from an applied perspective. Lenders are primarily interested in whether a firm defaults on a loan. Most firms that default on loans do not become bankrupt. Even for those default/accommodation firms that do become bankrupt, predicting the default or accommodation would be of more interest to the lender because default normally occurs before bankruptcy. Thus, models that predict loan default/accommodation should provide lenders more time to react.

REFERENCES

Agresti, A. (1984). Analysis of Ordinal Categorical Data. New York: John Wiley and Sons, Inc.

Altman, E. I. (1968). Financial ratios, discriminant analysis and the prediction of corporate bankruptcy. The Journal of Finance, 23 (September): 589-609.

Avery, R. A. Eisenbeis & J. F. Sinkey, Jr. (1981). Application of Classification Techniques in Business, Banking, and Finance. Greenwich, CT: JAI Press, Inc.

Avery, R. G. Haldeman & P. Narayanan. (1977). ZETA ANALYSIS: A new model to identify bankruptcy risk of corporations. Journal of Banking and Finance, 1 (June), 37-72.

Aziz, A., D. C. Emanuel & G. H. Lawson. (1988). Bankruptcy prediction--An investigation of cash flow based models. Journal of Management Studies, 25 (September), 419-437.

Aziz, A. & G. H. Lawson. (1989). Cash flow reporting and financial distress models: Testing of hypotheses. Financial Management, 18 (Spring), 55-63.

Ball, R. & G. Foster. (1982). Corporate financial reporting: A methodological review of empirical research. Journal of Accounting Research, 20 (Supplement), 161-233.

Beaver, W. H. (1966). Financial ratios as predictors of failure. Journal of Accounting Research, 4 (Supplement): 71-127.

Berkson, J. (1944). Application of the logistic function to bio-assay. Journal of the American Statistical, 39 (September), 357-365.

Blum, M. (1974). Failing company discriminant analysis. Journal of Accounting Research, 12 (Spring): 1-25.

Casey, C. J. & N. J. Bartczak. (1984). Cash flow--It's not the bottom line. Harvard Business Review, 62 (July-August): 61-66.

Casey, C.J. & N.J. Bartczak. (1985). Using operating cash flow data to predict financial distress: Some extensions. Journal of Accounting Research, 23 (Spring): 384-401.

Cosslett, S. R. (1981). Efficient estimation of discrete-choice models. In Structural Analysis of Discrete Data with Econometric Applications, edited by C. F. Manski and D. McFadden.

Cambridge, MA: The MIT Press.

Deakin, E. (1972). A discriminant analysis of predictors of business failure. Journal of Accounting Research, 10 (Spring): 167-179.

Dewing, A. S. (1953). The Financial Policy of Corporations. New York; Ronald Press.

Dietrich, J. R. (1984). Discussion of methodological issues related to the estimation of financial distress prediction models. Journal of Accounting Research, 22 (Supplement): 83-86.

Donaldson, G. (1969). Strategy for Financial Mobility. Boston, MA: Harvard Business School Press.

Drtina, R. E. & J.A. Largay, III. (1985). Pitfalls in calculating cash flow from operations. The Accounting Review, 60 (April), 314-326.

Franz, D. P. & J. B. Thies. (1988). Intertemporal divergence among cash flow, working capital, and income from operations. Review of Business and Economic Research (Spring): 18-28.

Freeman, D. H. (1987). Applied Categorical Data Analysis. New York: Marcel Dekker, Inc..

Gentry, J. A., P. Newbold & D. T. Whitford. (1985). Classifying bankrupt firms with funds flow components. Journal of Accounting Research, 23 (Spring): 146-160.

Gentry, J. A., P. Newbold & D. T. Whitford. (1987). Funds flow components, financial ratios, and bankruptcy. Journal of Business Finance and Accounting, 14 (Winter): 595-606.

Gilbert, L. R., K. Menon & K. B. Schwartz. (1990). Predicting bankruptcy for firms in financial distress. Journal of Business Finance and Accounting, 17 (Spring): 161-171.

Gombola, M. J. & J. E. Ketz. (1983). A note on cash flow and classification patterns of financial ratios. The Accounting Review, 58 (January): 105-114.

Gambola, M. J., M. E. Haskins, J. E. Ketz & D. D. Williams. (1987). Cash flow in bankruptcy prediction. Financial Management (Winter): 55-65.

Helfert, E. A. (1982). Techniques in Financial Analysis. Homewood, Ill.: Richard D. Irwin.

Holmen. J. S. (1988). Using financial ratios to predict bankruptcy: an evaluation of classic models using recent evidence. Akron Business and Economic Review, 19 (Spring): 52-63.

Jones, F. L. (1987). Current techniques in bankruptcy prediction. Journal of Accounting Literature, 6 (Spring): 131-164.

Largay, J. A. & C.P. Stickney. (1980). Cash flows, ratio analysis and the W. T. Grant Company bankruptcy. Financial Analysts Journal, 36 (July-August): 51-54.

Lau, A. Hing-Ling. (1987). A five-state financial distress prediction model. Journal of Accounting Research, 25 (Spring): 127-138.

Lawson, G. H. (1978). The rationale of cash flow accounting. In Trends In Managerial and Financial Accounting, ed. C. V. Dam. Martins Nijhoff.

Lawson, G. H. (1985). The measurement of corporate performance on a cash flow basis: A reply to Mr. Egginton. Accounting and Business Research, 15 (Spring): 99-107.

Maddala, G. S. (1991). A perspective on the use of limited-dependent and qualitative variables models in accounting research. The Accounting Review, 66 (October): 788-807.

Manski, C. F. & D. McFadden. (1981). Alternative estimators and sample designs for discrete choice analysis. In Structural Analysis of Discrete Data with Econometric Applications, edited by C. F. Manski and D. McFadden. Cambridge, MA: The MIT Press.

Manski. C. F. & S. R. Lerman. (1977). The estimation of choice probabilities from choice based samples. Econometrica, 45 (November): 1977-1988.

McFadden, D. (1974). Conditional logit analysis of qualitative choice behavior. In Economic Theory and Mathematical Economics, edited by P. Zarembka. New York: Academic Press.

Mensah, Y. M. (1983). The differential bankruptcy predictive ability of specific price level adjustments: Some empirical evidence. The Accounting Review, 58 (April): 228-245.

Mensah, Y. M. (1984). An examination of the stationarity of multivariate bankruptcy prediction models: A methodological study. Journal of Accounting Research, 22 (Spring): 380-395.

Newton, G. W. (1975). Bankruptcy and Insolvency Accounting. New York: The Ronald Press Company.

Noreen, E. (1988). An empirical comparison of probit and OLS regression hypothesis tests. Journal of Accounting Research, 26 (Spring): 119-133.

Norton, C. L. & R. E. Smith. (1979). A comparison of general price level and historical cost financial statements in the prediction of bankruptcy. The Accounting Review, 54 (January): 72-87.

Ohlson, J. A. (1980). Financial ratios and the probabilistic prediction of bankruptcy. Journal of Accounting Research, 19 (Spring): 109-131.

Press, S. J. & S. Wilson. (1978). Choosing between logistic regression and discriminant analysis. Journal of the American Statistical Association, 73 (December): 699-705.

Staubus, G. J. (1989). Cash flow accounting and liquidity: Cash flow potential and wealth. Accounting and Business Research, 19 (Spring): 611-169.

Stone, M. & J. Rasp. (1991). Tradeoffs in the choice between logit and OLS for accounting choice studies. The Accounting Review, 66 (January): 170-187.

Ward, T. J. (1994). An empirical study of the incremental predictive ability of Beaver's naive operating flow measure using four-state ordinal models of financial distress. Journal of Business Finance and Accounting (June): 547-561.

Ward, T. J. & B. P. Foster. (1997). A note on selecting a response measure for financial distress. accepted for publication in Journal of Business Finance and Accounting (Tentative Publication Date is Fall 1997).

Zavgren, C. V. (1985). Assessing the vulnerability to failure of American industrial firms: A logistic analysis. Journal of Business Finance and Accounting, 12 (Spring): 19-45.

Zmijewski, M. E. (1984). Methodological issues related to the estimation of financial distress prediction models. Journal of Accounting Research, 22 (Supplement): 59-82.

NOTES

(1.) The decision to limit the literature review to financial distress studies that compared the predictive ability of accrual and cash flow information was made to keep the task manageable.

However, the methodological limitations identified in the financial distress studies discussed in this paper, and the recommendations for correcting these limitations, apply to all studies using financial distress as a criterion to evaluate the predictive usefulness of accounting information.

(2.) Altman et al. (1981) provide a thorough discussion of what they consider to be the "ten landmark studies."

(3.) The distinction between these two objectives was more obvious in the earlier studies because of the limited capabilities of the statistical techniques used (forms of discriminant analysis) to develop tests of significance on particular variables. Later, the implementation of logistic and probit techniques enabled researchers to address both objectives. Still, the primary objective of a financial distress study dictates the type of methodology used.

(4.) Predicting bankruptcy has been a long and fruitful area of research. However, for brevity's sake, the author has only described the earlier studies by Altman, Ohlson, and Zavgren because these studies are typical of bankruptcy studies that did not incorporate cash flow variables into their models. Interested readers should see Ball and Foster (1982), which contains a complete listing of early financial distress studies.

(5.) A more recent study by Holmen (1988) comparing the predictive accuracy of the Altman Zscore model (Altman, 1968) and the cash flow/total debt variable (.03 and .07 cutoff points) used by Beaver (1966) on a sample of eighty-four bankrupt and eighty-four nonbankrupt firms found that the simple cash flow/total debt univariate model predicted bankruptcy with fewer errors than the multivariate Altman Z-score model.

(6.) Most statistical packages for MDA do not generate meaningful test statistics for each variable tested (the slopes using MDA have no intuitively practical meaning). MDA is also primarily a classification technique instead of a predictive technique; thus, MDA is not the most appropriate technique to use when the purpose is to test the ability of independent variables to predict financial distress.

(7.) Since probit and logistic techniques are so similar, both require similar statistical assumptions. Thus, the results from this study for probit analysis should also hold true for logistic techniques.

(8.) Most current nominal logistic statistical packages provide summary test statistics and are not as limited concerning the testing of the incremental predictive ability of variables. However, nominal logistic techniques still have many limitations when compared to ordinal logistic regression.

Terry J. Ward, Middle Tennessee State University