The impact of the response measure used for financial distress on results concerning the predictive usefulness of accounting information.

Ward, Terry J.

ABSTRACT

Researchers testing the usefulness of accounting information in predicting financial distress have used many different responses as proxies for financial distress. They often compare results across these different studies, attempting to make conclusions concerning the usefulness of particular accounting information. However, comparisons are valid only if the various response variables used by the various studies have construct validity; the different response variables all measure the same intended construct, economic financial distress.

The primary purpose of this paper is to determine the validity of various response variables of financial distress by observing the stability of results across three different response variables. Similar results across the different response variables would suggest that researchers could validly compare results of the various financial distress studies. However, results that vary depending on the response variable used would indicate that different response variables may actually measure different constructs, and that the results reported in previous studies may be dependent on the response variable used.

The findings of this study suggest that results very dependent on the response variable used for financial distress. Thus, one cannot validly compare the results of prior financial distress studies that used different measures of financial distress. The results of this study suggest that the various response variables are not equal measure of financial distress. Results seem to suggest that a dichotomous bankruptcy measure may be the poorest measure of economic financial distress.

INTRODUCTION

Since the 1960s, accounting researchers have used an ability to predict financial distress criterion to evaluate the usefulness of competing accounting methods. A major area of accounting information predictive usefulness research in the last three decades has concerned the predictive ability of accrual and cash flow information. The profession stressed the incremental usefulness of cash flow information by requiring that companies present a Statement of Cash Flows in 1987. In Opinion No. 95, the Financial Accounting Standards Board (FASB) expressed the board's belief that cash flow information, when taken together with accrual information, should help users predict future cash flows and future firm insolvency. The board required that companies report three net cash flows by activities: cash flow from operating activities, cash flow from investing activities, and cash flows from financing activities.

Prior researchers have used various response variables as proxies for financial distress. The earliest researchers used a dichotomous nonfailed versus failed response variable, while subsequent researchers have primarily used a dichotomous nonbankrupt versus bankrupt response variable for financial distress. A few studies have used multi-state response variables for financial distress.

Researchers often compare results across these different studies, attempting to make conclusions concerning the usefulness of particular accounting information. However, comparing the results of prior financial distress studies is questionable considering researchers used different response variables (dependent variables) for financial distress. Comparisons between studies are valid only if the various response variables have construct validity; the different response variables must all measure the same intended construct, economic financial distress.

This study compares the predictive ability of cash flow and accrual information using different response variables for financial distress. The primary purpose of this paper is to determine the validity of various response variables of financial distress by observing the stability of results across three different response variables. Similar results across the different response variables would suggest that researchers can validly compare results of the various financial distress studies. However, results that vary depending on the response variable used would indicate that different response variables may actually measure different constructs, and that the accrual and cash flow results reported in previous studies may be dependent on the response variable used.

REVIEW OF THE LITERATURE

The author selected the comparison of cash flow and accrual information as the vehicle to address the issue of response measure validity because of the substantial amount of research found in accounting and business journals during the last three decades. This subject is also very important since firms spend substantial time and cost preparing a cash flows statement, while results of prior financial distress cash flow research has been mixed concerning the incremental predictive usefulness of cash flow information over accrual information. Early financial distress researchers found a naive measure of operating cash flow, net income plus depreciation and amortization, to be a significant predictor of financial distress. Bankruptcy studies since 1980 tested more refined measures of operating cash flow (researchers eliminated other allocations and the impact of current receivables and payables on operating cash flow) and tested other cash flows. These studies' results suggest that cash flows do not have incremental predictive power over accrual information, although cash flow from operations is sometimes significant, especially one year before financial distress. However, two multi-state studies in the 1990s found evidence suggesting the naive operating cash flow, cash flow from operating activities, and cash flow from investing activities may have incremental predictive ability certain periods before financial distress. More recent studies have used various response measures to measure the usefulness of related accounting information.

Ward (1999) contains a through review of the earlier research in this area. Table 1 contains brief summaries of some of the studies over this period of time.

Early cash flow studies compared the predictive usefulness of accrual and cash flow information. Beaver (1966), Deakin (1972), and Blum (1974) tested the predictive ability of a naive operating cash flow, net income plus depreciation and amortization, to predict financial distress. These three studies used a dichotomous nonfailed versus failed response for financial distress. All three studies found that the naive operating cash flow scaled by total debt was a strong predictor of financial distress.

Subsequent studies by Altman et al. (1977), Norton and Smith (1979), and Mensah (1983) tested the ability of financial ratios and cash flows to predict nonbankrupt and bankrupt firms. The authors used various stepwise linear and quadratic MDA models. Except for Altman et al., the naive cash flow, scaled by various measures, was a strong predictor of financial distress when included in models with accrual ratios.

Largay and Stickney (1980), Casey and Bartczak (1984; 1985), Gentry et al. (1985; 1987), Gombola et al. (1987), Aziz et al. (1988), and Aziz and Lawson (1989) subsequently tested more refined measures of operating cash flow (they eliminated additional accounting allocations and the timing differences in payables and receivables, etc.). Subsequent studies also tested the predictive ability of additional cash flows besides operating cash flow. For the studies that sampled more than one firm, the results provided little evidence suggesting that cash flows have incremental content above accrual information in predicting bankruptcy. These results are surprising, since a main stated benefit of cash flows is their incremental usefulness in helping creditors to predict insolvency (Staubus, 1989). The only study showing that cash flow based components have incremental predictive content (Gentry et al., 1987) actually found that certain changes in accounts that comprise working capital have incremental predictive content.

Gilbert et al. (1990) replicated the study of Casey and Bartczak (1985) and using two separate samples of firms, a sample of nonbankrupt versus bankrupt firms and a sample of distressed (defined as firms having consecutive losses) versus bankrupt firms. The authors found that operating cash flow could significantly distinguish between distressed and bankrupt firms. However, a model developed from the nonbankrupt versus bankrupt sample performed poorly when used to distinguish distressed firms from bankrupt firms. This result suggests that cash flow information may be more useful in distinguishing between events of financial distress other than bankruptcy. However, Gilbert et al. 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.

Ward et al. (2006) investigated whether bankrupt firms had greater articulation problems than nonbankrupt firms. The authors found that bankrupt firms where more likely to have articulation problems than nonbankrupt firms and that these articulation problems resulted in an overstated estimated operating cash flow measure. The authors concluded that this articulation problem for the bankrupt firms might explain why earlier distress studies found little significance in explaining financial distress for operating cash flows.

Lau (1982; 1987) extended the methodology of prior studies 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 the distressed firms to be on an ordinal scale, stating that "states one to four are states of increasing severity of financial distress" (pg. 128).

Lau compared the predictive ability of four funds flow measures, of which one was CFO/TL (cash flow from operations/total liabilities), in her 1982 dissertation. However, the results from this study were mixed. The CFO/TL model was the strongest model when classification was used to evaluate the predictive ability of each model, while working capital from operations scaled by total liabilities was the best measure when a rank score was used. Lau only reported the results for the working capital from operations model in the 1987 published study.

The model used by Lau, however, did not incorporate the ordinal structure of the dependent variable; her statistical models were nominal, not ordinal. For an ordinal multi-state dependent variable, ordinal logistic regression provides many advantages over nominal logistic regression (Agresti, 1984; Kennedy, 1992).

Ward (1994) developed an ordinal four-state model similar to Lau's nominal five-state model to determine the reason why Beaver's naive cash flow measure is such a strong predictor of financial distress. Ward extended the methodology of Lau's study by using ordinal logistic regression to generate the prediction models, thus incorporating the ordinal scale of the dependent variable. Ward found that the naive measure of financial distress is an incrementally significant predictor variable two years before financial distress, while cash flow from operating activities is significant one and two years before financial distress.

Ward and Foster (1996) subsequently used various multi-state models in testing the usefulness of allocation free information in predicting financial distress. The authors found that accounting ratios free of deferred tax components and depreciation best explained future financial distress.

More recent studies have used various measures of financial distress to test the ability of neural network models to predict financial distress (Zurada et al., 1999; 2001a; 2001b; Agarwal, 2001). Results of these studies have been somewhat mixed, with various neural network models showing some success in predicting financial distress.

Sensitivity of Results to Response Variable Used

Bahnson and Bartley (1992) investigated the sensitivity of cash flow results to the response variable used. They compared prior cash flow models of Casey and Bartczak (1984; 1985) with their own models using three different definitions (responses) of financial distress (1) nonbankrupt (nonevent, technical default, default, or troubled debt restructures) versus bankrupt, (2) solvent (nonevent or technical default) versus insolvent (default, troubled debt restructured or bankrupt), and (3) three state response with nonevent, technical default, and insolvent (default, troubled debt restructured or bankrupt). Results showed that the usefulness of CFO (cash flow from operations) depended on the definition of failure. CFO was not significant in either model when a nonbankrupt versus response was used. However, when the response was measured in a broader sense as solvent versus insolvent, CFO is significant in the Bahnson and Bartley model as type 1 errors (i. e., incorrectly classify a failed firm as being nonfailed) are increased.

However, Bahnson and Bartley's study suffers from two limitations that limit the generalizability of their results. First, Bahnson and Bartley used response variables substantially different from those used in prior studies. The multi-state and broadly defined responses used by Bahnson and Bartley were based on their previous unpublished paper. Nonbankrupt firms included technical default, default, and troubled debt restructured firms. Although prior bankruptcy studies likely included some technical default, default, and troubled debt restructure firms in their nonbankrupt samples, the numbers were likely much smaller than used in Bahnson and Bartley's study.

The authors also included technical default firms in all of their measures. Since loan covenants are normally based on accounting information, using this event to measure a response variable (dependent variable) regressed on accounting ratios produces some statistical bias and can produce misleading results. The bias from including technical default firms in the samples could be great considering that technical default dominated the "event" sample (76 of the 119 event companies were technical default firms).

Neill et al. (1991) reviewed prior financial distress cash flow research and concluded that CFO (cash flow from operations) doesn't appear to be a consistent predictor of financial distress. Other cash flows, particularly investment and dividends paid cash flows, appear more useful. They conclude that the usefulness of CFO information appears to be affected by "(1) the definition of failure employed (CFO is more important when failure is defined broadly, (2) the condition of the economy (CFO is more important in an economic downturn) and (3) the condition of the firms (CFO is more important for extreme observations)" (pp. 143-144). The authors stress that "greater attentions should be paid to the definition of failure employed and to the independent variables used" and that "future research should present results using alternative definitions of failure" (pp. 144-145).

Ward and Foster (1997) tested whether a loan default/debt accommodation response variable produced different results than a bankruptcy response measure. The authors concluded that a loan default/accommodation response seems to be a better measure of economic distress than bankruptcy. However, the authors never investigated multi-state response measures and a dichotomous distress versus nondistress measure.

This paper extends the research of Neil et al. (1991), Bahnson and Bartley (1992), and Ward and Foster (1997) by comparing the predictive ability of cash flow and accrual information using different response variables for financial distress. Similar to Bahnson and Bartley, this paper tries to determine whether or not cash flow results differ across various response variables. This study extends prior research by: (1) using response variables (especially the multi-state and broadly defined dichotomous responses) more similar to those used in prior studies, (2) including investing and financing cash flow variables in all models, (3) looking at cash flows during strong economic times, and (4) developing a separate holdout sample to determine the predictive ability of each model.

METHODS

Sample Selection

This study uses two separate samples of firms, an original sample and a holdout sample. The original sample was used to generate the prediction models and is composed of healthy and financially distressed 1988 firms. A holdout sample of 1989 firms was used to validate the predictive ability of models generated. The author developed the samples from separate years to provide intertemporal validation of each model's predictive strength. Readers should consult Lau (1982; 1987) and Altman et al. (1981) for discussions concerning the need for intertemporal validation of prediction models.

Since the purpose of this paper is to compare whether results are consistent across the differing responses used in prior research, the author needed to select samples from periods similar to prior studies. This sample is the same initial sample used by Ward (1994) and Ward and Foster (1997) and the period is consistent with the other cash flow studies. For a complete description of the sampling procedures used to select these firms, see Ward and Foster (1997).

The 1988 sample contained 227 firms of which 164 were healthy, twenty-two reduced cash dividends, twenty-three experienced a loan default or debt accommodation, and eighteen filed bankruptcy. The 1989 sample included 158 firms of which 111 were healthy, seventeen reduced cash dividends; fourteen experienced a loan default or debt accommodation, and sixteen filed for bankruptcy.

For some firms, the bankruptcy announcement comes before financial reports for the preceding year are issued. Consequently, these financial reports include information about a firm's bankruptcy (Ohlson, 1980). This problem can also occur for firms experiencing a default or debt accommodation. Therefore, this study substitutes reports from the previous fiscal year as the most current year of interest for firms releasing financial reports after the date of financial distress.

Response Variables

This study tests the predictive ability of cash flow and accrual information using three response variables (dependent variables) for financial distress. These three response variables are similar to variables used in prior financial distress studies. The three dependent variables are as follows: (1) healthy versus distressed (nonfailed versus failed) response, (2) healthy versus bankrupt response, and (3) ordinal four-state response.

The nondistressed versus distressed dependent variable was coded as follows:

DIST = 0 if firm was healthy, and

 = 1 if firm experienced a greater than forty percent reduction in
 cash dividend per share after a history of successive dividends
 per share, a loan principal/interest default or debt
 accommodation, or filed (or was forced to file) for Chapter XI
 protection.

This binary response variable is similar to the nonfailed versus failed measure used in previous financial distress studies.

The healthy versus bankrupt dependent variable was coded as follows:

DIST = 0 if firm was healthy, and

 1 if firm filed, or was forced to file, for Chapter XI
 protection.

This binary response variable is similar to the nonbankrupt versus bankrupt response measure used in prior studies.

The ordinal four-state dependent variable was coded as follows:

DIST = 0 if firm was healthy (no event of financial distress),

 1 if firm experienced a greater than forty percent reduction in
 cash dividend per share after a history of successive dividends
 per share (Deciding on a criterion for selecting dividend
 reduction/default firms is somewhat arbitrary. The author chose
 a forty percent criterion because this criterion was used by Lau
 (1982 and 1987)),

 2 if firm experienced a loan principal/interest default or debt
 accommodation, and

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

This ordinal four-state response variable is the same response variable used by Ward (1992; 1994) and Ward and Foster (1997) and is similar to Lau's five-state response variable. The primary difference between the four-state response variable used in this study and Lau's five-state response variable is that the four-state response variable does not include liquidation firms as a fifth state because of the small number of liquidation firms in the original sample (four firms).

Advantages and Disadvantages of Each Response Variable

Each response measure has certain advantages and disadvantages over the other measures. One advantage of the dichotomous nondistressed versus distressed response variable is that this response measure includes firms that are marginally distressed as well as firms that are very distressed. Thus, the researcher can obtain a much larger sample of distressed firms. This mixture of distressed firms should result in a stronger test of the predictive ability of accounting information. The major criticism of this dichotomous response variable is that the nondistressed state is composed of heterogeneous firms. 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).

A major advantage of the healthy versus bankrupt response measure is that a researcher can easily find a sufficient sample of bankrupt firms. Researchers can identify the bankrupt firms from many sources (such as CD data bases or by identifying firms that have been transferred to the Compustat Research Tape). Bankruptcy also has the advantage of tradition. Users of financial accounting information, accountants, and researchers are comfortable with using bankruptcy as the traditional definition of financial distress.

However, a nonbankrupt versus bankrupt dichotomous response variable suffers from two weaknesses somewhat ignored in prior financial distress literature. First, the use of bankruptcy as the sole proxy for 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. Financial distress literature stresses the belief that many events indicate different degrees of financial distress (Giroux and Wiggins, 1984; Lau, 1987).

Second, one can also question 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 after the firm is economically insolvent, or occur even though the company is not economically insolvent. 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.

The ordinal four-state response variable's principal advantage is that it should provide a stronger test of the usefulness of accounting information, while not suffering from the limitation of having heterogeneous firms combined in one level. Predictor (independent) variables must distinguish between firms that are healthy and those marginally distressed, as well as distinguish between healthy firms and very distressed firms, thus providing a stronger test of the predictive usefulness of the variables tested.

The ordinal response measure does suffer from the complexity of using multi-states. Much time and effort is needed to obtain sufficient sample sizes using multiple states of financial distress. From a statistical perspective, the researcher must determine whether to use nominal (such as Lau) or ordinal regression (such as Ward) to generate the four-state model. If the response scale is ordinal, then ordinal logistic regression is the appropriate method to use (Kennedy, 1992). However, nominal logistic regression would be more appropriate if, after logistic transformation, the independent variables are not linearly related to the dependent variable (e.g., the relationship may be curvi-linear or the states may reverse).

Independent Variables

The independent variables examined consist of seven control variables and the three net cash flows required on a Statement of Cash Flows. The control variables are six accrual ratios found significant in prior financial distress studies (Casey and Bartczak, 1984; 1985; Gentry, et. al., 1987; Gilbert et. al., 1990; Ward, 1992; 1994; Ward and Foster, 1997) and a control variable to control for firm size. The control variables are as follows:

SIZE = log (total assets),
NITA = net income/total assets,
SALESCA = sales/current assets,
TLOE = total liabilities/owners' equity,
CACL = current assets/current liabilities,
CATA = current assets/total assets, and
CASTA = cash plus marketable securities/total assets.

The cash flow variables tested are as follows:

CFFO = cash flow from operating activities,
CFFI = cash flow from investing activities, and
CFFF = cash flow from financing activities.

The cash flow variables were computed from Compustat tapes. The author calculated the cash flows using the following formulas based on Compustat line items: CFFO = Income before extraordinary items + depreciation and amortization + deferred taxes + equity in net loss (earnings) + loss (gain) from sale of property, plant, and equipment and investments + funds from operations-others + accounts receivable-decrease (increase) + inventory-decrease (increase) + other current assets-decrease (increase) + current liabilities other than current debt-increase (decrease). CFFI = sale of property, plant and equipment--capital expenditures--acquisitions--increase in investments + sale of investments + short-term investments-change. CFFF = change in current debt- increase (decrease) + change in long-term debt-increase (decrease) + sale of common and preferred stock--purchase of common and preferred stock--cash dividends.

To prevent heteroscedasticity, this study scaled the cash flow variables by total liabilities. The author selected total liabilities as the scaling measure because it yielded a better fit to the data than scaling by current assets, total assets, current liabilities, sales, or owners' equity. This result is consistent with prior result (Gilbert et al., 1990; Lau, 1987; Ward, 1992; 1994).

Statistical Models

This study uses financial data for 1984/85 (year three models), 1985/86 (year two models), and 1986/87 (year one models) to predict the financial distress of 1988 firms. The predictive accuracy of each model is then validated with a holdout sample of 1989 firms.

The ordinal four-state prediction model was constructed using ordinal logistic regression (OLGR), proportional odds variation. This procedure fits a parallel lines regression model based on transformed cumulative logits. OLGR assumes an ordinal relationship between the dependent and independent variables. However, OLGR does not make an assumption concerning the intervals between the levels of the dependent variable.

Binary logit regression (LG) was used to generate the dichotomous prediction models in this study. LG is similar to OLGR. LG fits a regression model based on a single transformed logit instead of cumulative logits. (The ordinal and binary logistic models used in this study were proportional odds models.) LG has been used extensively in prior financial distress research (e.g., Casey and Bartczak, 1985; Gentry et al., 1985; Aziz et al., 1988; Aziz and Lawson, 1989). For brevity's sake, this study doesn't illustrate the OLGR or LG models. Agresti (1984), Kennedy (1992), and Ward (1992; 1994) discuss ordinal logistic regression, while Hosmer and Lemeshow (1989) provide a thorough discussion of binary logit regression.

Since dichotomous financial distress studies use nonrandom techniques to sample distressed firms, parameter estimates can be biased (sample proportions are not similar to population proportions). Zmijewski (1984) demonstrated a weighted probit procedure to correct for this choice-base bias in a binary probit model, while Cosslett (1981) illustrated weighted binary conditional logit models. However, Maddala (1991) demonstrates that the binary logit model does not result in biased parameter estimates. According to Maddala, one does not need to use a weighting procedure for the logit model because the unequal sampling rates do not affect the coefficients of the predictor variables; only the intercept needs to be adjusted based on the proportion sampled from the population for each group.

Since the purpose of this study was to compare across the different models, and since it is almost impossible to determine the percentage of loan default/accommodation firms actually identified (the researcher is unable to identify all of these firms in a population) from the population, the author of this study didn't adjust the intercept of each binary model by the proportions. Failure to adjust the intercept should not affect comparisons between the models since all models were treated the same. However, this study did use the sample ratios (healthy to distressed) as cutoffs for classification. The author of this study is unaware of research addressing the existence of choice-base sampling bias in ordinal multi-state models.

RESULTS

Significance of Independent Variables--Original Sample

The author first developed four regression models, one for each response variable, to test the explanatory power of the independent variables. If all four response measures are measuring the same financial distress construct, then the results for the predictor variables should be similar across the various response variables. Table 2 contains the statistical results for the regression models.

Table 2 shows that the statistical results are very dependent on the response variable used; results are not consistent across the different responses. Only CFFO is significant in all of the four models one year before distress. The H vs. B (healthy vs. bankrupt) model's results generally fail to agree with the other models' results, especially two or three years before financial distress. However, the H vs. B model's results are consistent with prior bankruptcy cash flow studies. CFFO is the primary incremental explanatory cash flow variable of bankruptcy one year before financial distress. This result suggests that either the other three response variables are poor proxies of financial distress or a bankruptcy response variable is not a good proxy for financial distress.

The H vs. D (healthy vs. distressed) model shows significance for the operating cash flow variable two of the three years before financial distress. This result suggests that the main advantage of cash flows may be their ability to distinguish between firms with different levels of financial distress (marginally distressed firms). Cash flows may not add explanatory power to accrual information when asked to discriminate between firms that are healthy and firms that are very distressed. This finding is consistent with the findings of Gilbert et al. (The author reviewed the correlation matrices (not reported) of the estimated parameter estimates for all models to determine if multicollinearity was a problem. These correlation matrices did not indicate a multicollinearity problem in any model.)

Significance of Independent Variables--Combined Sample

To determine if the small sample sizes for the loan default/accommodation firms in the H vs. L model (twenty-three loan default/accommodation firms) and the bankrupt firms in the H vs. B model (eighteen bankrupt firms) could have affected the results, the author combined the original and holdout samples and reran the models. Table 3 contains the statistical results for the models using the combined sample.

Table 3 results show that more predictor variables are now significant. Still, the results primarily agree with the original sample results reported in Table 2. The significance of particular predictor variables are still dependent on the scaling measure used. Only CACL and SIZE are incrementally significant explanatory variables for all four models (one and three years before the event, respectively). The four-state and H vs. D models tend to agree more than any of the other models.

However, as for the original sample, the H vs. B model's results are different from the other models' results (except for CACL and SIZE), especially for the cash flow variables. The H vs. B model's results suggest that no cash flow variable is significant either year before bankruptcy. However, CFFO is incrementally significant for the other three models one year before financial distress, while CFFI is significant for the other three measures two years before financial distress.

The results using a combined sample also suggest that a binary bankruptcy proxy may be a poor proxy for financial distress. Thus, using a binary bankruptcy response as the sole proxy for financial distress could result in misleading conclusions concerning the incremental predictive ability of accounting information, especially cash flow information. Since prior financial distress studies primarily used a binary nonbankrupt versus bankrupt response variable for financial distress to test the predictive usefulness of cash flow information, their findings that cash flow information was not incrementally useful may have been affected by the response variable used.

Validation of Models--Using Classification Accuracy

To validate the statistical results reported in Tables 2 and 3, the author checked the ability of the models to classify firms correctly one, two, and three years before financial distress. As stated before, the author used prior probabilities equaling the sample sizes for classification purposes to eliminate the effects of choice-base bias on classification rates. The author also reports the classification rates for the four-state model. Comparing the classification rates of the four-state model with the binary models' classification rates is questionable, considering the four-state model is scaled differently. Since the four-state model must select classification among four different states, it naturally would have a lower overall percentage of firms classified correctly. A rank score that considers the ordinal scale of the response variable would be a better measure of the predictive ability of an ordinal regression model. Because this study is primarily interested in comparing the change in prediction ability after adding the cash flow variables to the accrual variables, the author had to select a validation method that could be applied to all of the response variables. Classification accuracy has been used extensively in the binary financial distress studies. Thus, the author of this study believes it appropriate to also calculate the classification rates for the four-state model.

Table 4 contains the classification rates for the models using the original sample, while Table 5 contains the results for the holdout sample.

If cash flow information has practical incremental predictive usefulness above accrual ratios, then the combined models with the cash flows added to the accrual ratios should out-predict an accrual model. Table 4 results primarily validate the statistical results reported earlier. Results vary depending on the response variable used. For example, the cash flow variables improve predictions all three years for the four-state and H vs. D models, while the H vs. B and H vs. L (healthy vs. loan default) models show an improvement in classification in only one of the three years.

The holdout sample results also suggest that results vary depending on the response variable used. However, results differ for particular response variables. For the holdout sample, the H vs. D model failed to show an improvement for the cash flows either year, while the four-state and H vs. B responses showed improvement two of the three years.

Additional Analysis Using a Nonbankrupt versus Bankrupt Response

The H vs. B response used in this study differs somewhat from the nonbankrupt versus bankrupt response used by prior researchers. Because of sampling techniques used in earlier studies, prior researchers likely included some dividend default and loan default/accommodation firms as nonbankrupt firms. To determine the effect of this difference on results, the author reran the models using a bankrupt versus nonbankrupt (dividend reduction, loan default/accommodation and healthy firms combined) sample of firms. Table 6 contains the results using a nonbankrupt versus bankrupt (NB vs. B) response variable.

Table 6 results show that the results for the NB vs. B (nonbankrupt vs. bankrupt) model are very similar to the results for the H vs. B model reported in Table 2. The primary difference between the two models' results is that CFFF is significant one and two years before financial distress for the NB vs. B response model. Since CFFF was never significant either year using the other three response variables (four-state, H vs. D, and H vs. L), this result further suggests that a binary bankruptcy response variable may measure a different construct than the other response variables.

Summation of Results

Combined, the results of this study suggest that results concerning the predictive usefulness of accrual and cash flow information are very dependent on the response variable used for financial distress. Thus, one cannot validly compare the results of prior financial distress studies that used different measures of financial distress.

The results of this study suggest that the various response variables are not equal measures of financial distress. Thus, one is left wondering which financial distress variable best measures financial distress. Although this study did not specifically attempt to determine which variable best measures financial distress, the results of this study do offer some insights. If theory indicating that cash flow information should have short-term predictive content (in predicting financial distress) is valid, then the results of this study suggest that the four-state response measure is the better response variable. The four-state models consistently show that various cash flows are incrementally important predictors of financial distress in the short-term.

Loan default/accommodation firms appear to be as financially distressed as bankrupt firms two and three years before the event. In fact, predictive models tended to distinguish healthy firms from loan default/accommodation firms easier than they did healthy firms from bankrupt firms. (Classification rates for H vs. L models were normally higher than rates for H vs. B models.) This finding suggests that future applied binary prediction models developed for creditors should be based on a healthy versus loan default/accommodation response instead of a bankrupt response such as Altman's Z-score model. Creditors would benefit more from H vs. L models since loan default/accommodation normally occurs before bankruptcy (Giroux and Wiggins, 1984); thus, prediction models based on loan default/accommodation should provide creditors more time to take action concerning future losses.

REFERENCES

Agarwal, A., J. T. Davis & T. J. Ward 2001). Supporting ordinal four-state classification decisions using neural networks. Journal of Information Technology and Management, 2(1), 5-26.

Agresti, A. (1984). Analysis of Ordinal Categorical Data. New York: John Wiley and Sons, Inc. Altman, E. I., R. G. Haldeman & P. Narayanan (1977). ZETA ANALYSIS: A new model to identify bankruptcy risk of corporations. Journal of Banking and Finance, (June), 37-72.

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

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.

Bahnson, P. R. & J. W. Bartley (1992). The sensitivity of failure prediction models to alternative definitions of failure. Advances in Accounting, 10, 255-278.

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

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, 167-179.

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.

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.

Giroux, G. & C. Wiggins (1984). An events approach to corporate bankruptcy. Journal of Bank Research, 15 179-187.

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

Hosmer, D. W. & S. Lemeshow (1989). Applied Logistic Regression. New York: AICPA.

Kennedy, D. B. (1992). Classification techniques in accounting research: Empirical evidence of comparative performance. Contemporary Accounting Research, 8(spring), 419-442.

Largay, J. & C. 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 (1982). On the prediction of firms in financial distress, with an evaluation of alternative funds-flow concepts. Ph.D. dissertation, Washington University, Saint Louis, MO.

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

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.

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

Neill, J. D., T. F. Schaefer, P. R. Bahnson & M. E. Bradbury (1991). The Usefulness of Cash Flow Data: A Review and Synthesis. Journal of Accounting Literature, 10, 117-149.

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

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

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

Ward, T. J. (1992). The incremental predictive ability of net and gross cash flows using four-state ordinal models of financial distress. Southern Business and Economic Journal, 16(October), 26-44.

Ward, T. (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, 21(June), 547-561.

Ward, T. J. & B. Foster (1996). An empirical analysis of Thomas's financial accounting allocation fallacy theory in a financial distress context. Accounting and Business Research, (spring), 137-152.

Ward, T.J. & B. Foster (1997). A note on selecting a response measure for financial distress. Journal of Business Finance and Accounting, (July), 869-879

Ward, T.J. & B. Foster (1999). A review of financial distress literature and recommendations for future research. Academy of Accounting and Financial Studies Journal, 3(winter), 136-151.

Ward, T. J., B. Foster & J. Woodroof (2006). Estimated operating cash flow, reported cash flow from operating activities, and financial distress. Advances in Quantitative Analysis of Finance and Accounting, 4(summer), 97-120.

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

Zurada, J. M., B. Foster, T. J. Ward & R. M. Baker (1999). Neural networks versus logit regression models for predicting financial distress response

variables. Journal of Applied Business Research, 15(winter), 21-30.

Zurada, J. M., B. Foster & T. J. Ward (2001a). Artificial neural networks in predicting a dichotomous level of financial distress for uneven training and testing samples. In Contemporary Trends in System Development, edited by M. Sein, B. Munkvold, T. Orvik, W. Wojtleowski, G. Wojtleowski, J. Zupancic, and S. Wrycza, (New York, NY: Kluwer Academic/Plenum Publishers, 2001), 175-182.

Zurada, J. M., B. Foster & T. J. Ward (2001b). Investigation of artificial neural networks for classifying levels of financial distress of firms: The case of an unbalanced training sample. In Knowledge and Discovery for Business Information Systems, edited by Abramowicz and Zurada, Kluwer Academic Publishers, Norwell, (Chapter 17), 397-424.

Terry J. Ward, Middle Tennessee State University

Table 1: Summary of Prior Financial Distress Cash Flow Studies

 Cash Flow
 Variables
Study Response/Sample Tested Findings

Used distressed versus non-distressed measure:

Beaver (1966). 79 failed and 79 Naive operating Cash flow/total
 nonfailed firms cash flow (net debt (CF/TD) is
 (failed = income + best single
 bankrupt bond depreciation predictor.
 default and
 overdrawn bank amortization
 account, or scaled by
 nonpayment of various balance
 preferred sheet totals).
 dividends).

Deakin (1972) 32 failed and CF/TD. CF/TD most
 non-failed significant in
 (failed = all models.
 bankrupt,
 insolvent, or
 liquidated).

Blum 115 failed and CF/TD. CF/TD variable
(1974) 115 non-failed generally
 industrial firms received high
 (failed = rankings.
 failure to pay
 debts when due,
 debt
 accommodation
 agreement with
 creditors, or
 bankrupt).

Used bankrupt versus non-bankrupt measure:

Altman et al. 53 bankrupt and Naive cash flow Out of 27
(1977). 58 non-bankrupt scaled by fixed variables, the
 firms from charges and naive cash flow
 manufacturing CF/TD. variables were
 and retailing. not found to be
 a part of the
 best model.

Norton & 30 bankrupt and CF/TD and naive CF/TA and CF/TD
Smith (1979). 30 non-bankrupt cash flow were part of
 publicly traded scaled by sales best
 firms. (CF/S), total discriminant
 assets (CF/TA), model 3 years
 and net worth before
 (CF/TW). bankruptcy.

Mensah (1983) For ex ante CF/S, CF/TA, CF/NW was most
 prediction CF/NW, CF/TD, important ratio
 purposes, 11 and naive cash in discriminant
 bankrupt and 35 flow scaled by model.
 non-bankrupt curent
 firms were liabilities
 randomly (CF/CL).
 selected.

Largay & One bankrupt Working capital CFO provided a
Stickney firm. from operations more accurate &
(1980). (WCFO), cash timely signal
 flow from of W.T. Grant's
 operations eventual
 (CFO, more bankruptcy.
 refined
 operating cash
 flow measure).

Casey & 60 bankrupt and CFO and CFO Cash flow
Bartczak (1984 230 non-bankrupt scaled by ratios are
& 1985) firms. Matched current significant
 by industry liabilities during certain
 Holdout sample. (CFO/CL) and years. However,
 total neither cash
 liabilities flow variable
 (CFO/TL). had higher
 classification
 accuracy than 6
 combined
 accrual ratios.
 Addition of
 each cash flow
 variable did
 not increase
 classification
 accuracy.

Gentry et al. 33 bankrupt and 7 cash-based Funds flow
(1985). loss firms and funds flows components have
 33 non-bankrupt (each divided predictive
 firms. No by total net content but the
 holdout sample, cash flow). cash flow
 but 2nd sample Never tested components of
 of weak versus CFP, but tested CFO do not
 non-weak firms. components of improve
 CFO. classification
 accuracy.

Gentry et al. Same as before. 11 funds flow Investment,
(1987). variables. dividend, and
 receivable
 funds flow
 variables are
 significant,
 and some have
 incremental
 predictive
 power.

Gombola et al. 77 bankrupt and CFO/TA and CFO variable
(1987).. 77 non-bankrupt WCFO/TA. not
 firms. Two significant,
 separate models: and CFO
 early (1967- variable not
 1972) & late more useful in
 (1973-1981). late-year
 model.

Aziz et al. 49 bankrupt and 6 cash flow Taxes paid,
(1988) 49 non-bankrupt variables, each operating cash
 firms. No sclaed by book flow, & lender
 holdout sample, value. cash flow most
 jackknife significant.
 technique.

Aziz & Same as before, Same cash flow Cash flow
Lawson except also used variables as variables do
(1989). a holdout sample before and the not improve on
 of 26 bankrupt 5 accrual existing
 and 67 non- ratios in models' overall
 bankrupt firms. Altman's (1968) accuracy.
 Z-score model.

Used bankrupt versus non-bankrupt and bankrupt versus distressed
samples:

Gilbert et al. Two main Replicated CFO/TL has
(1990) samples: (1) Casey and incremental
 sample of 76 Bartczak's predictive
 bankrupt and 304 study (1985) power for
 non-bankrupt and Altman's bankrupt versus
 firms and (2) study (1968). distressed
 sample of 76 models.
 bankrupt and 304 Bankruptcy
 distressed firms models
 (distressed performed
 firms = those poorly
 that had distinguishing
 negative bankrupt from
 cumulative distressed
 earnings over a firms.
 consecutive 3
 year period.
 Holdout (above
 samples split
 into two groups)
 sample.

Ward et 50 distressed Net income/ Distressed
al. (2006). and 50 non- total assets firms have
 distressed Sales/current greater
 firms. Primary assets, current nonarticulation
 holdout and assets/current than non-
 combined liabilities, distress firms.
 samples. total
 liabilities/
 owner's equity,
 current assets/
 total assets,
 cash + mk
 securities, log
 (total assets),
 estimated
 operating cash
 flow, and
 reported
 operating cash
 flow.

Used Multi-state measures of financial distress:

Lau (1982) 350, 10, 15, 10, Attempted to Results mixed,
related to & 5 firms in 5 test 4 funds CFO/TL model
1987 published states: healthy, flow variables, strongest using
article omitting or of which CFO/TL classification
 reducing was one. accuracy, WCFO/
 dividends, TL model
 default of loan strongest using
 interest &/or rank scores.
 principal
 payments,
 protection under
 Chapter X or XI,
 and bankruptcy &
 liquidation for
 1976. Nominal
 statistical
 model. Holdout
 sample of 1977
 firms.

Lau (1987) Same as above 10 variables, Multi-state
 of which WCFO/ model somewhat
 TL was the strong.
 funds flow
 variable
 tested.

Ward (1994). 164, 22, 23, & 9 variables, of CF/TL is
 18 firms in 4 which one was measuring an
 states: healthy CF/TL and one economic income
 omitting or was cash flow construct.
 reducing from operating CF/TL
 dividends, loan activities incrementally
 principal/ scaled by total significant two
 interest default liabilities years before
 or debt (CFFO/TL). financial
 accommodation, Purpose was to distress.
 and protection determine what CFFO/TL
 under Chapter X. construct CF/TL incrementally
 Ordinal statical was measuring. significant one
 model. Holdout year before
 sample of 1989 financial
 firms. distress.

Ward (1992) Same as above. 12 variables, LFF/TL has
 of which one incremental
 was cash flow significan
 from financing explanatory
 actitivies power over
 scaled by total CFFF/Tl one
 liabilitis year before
 (CFFF/TL). financial
 Author also distess. SFF/TL
 tested three has
 gross cash increamentail
 flows: significant
 long-term explanatroy
 financing flow power over
 (LFF/TL), CFF/TL two
 short--term years before
 financing flow) finacial
 SFF/TL), and distress.
 equity
 financing flow
 (EFF/TL).
 Purpose was to
 detemine
 whether the
 gross financing
 cash flows had
 incremental
 predictive
 ability over
 the net cash
 flow from
 financing
 actitivies.

Ward et al. Same as above, Same basic Depreciation
(1996) except, authors variables as and deferred
 also collapsed 1994 as control tax have no
 states into variables. useful
 various reduced Authors added information
 models. various content in
 allocation free predicting
 variables (11 financial
 variables) into distress.
 the models.

Table 2: Predictor Variables Significant at P-value [less than or equal
to] .05, Using Various Response Variables for Financial Distress

 Predictor Variables
 Response (1)
Year Variable SIZE (2) NITA SALESCA CACL TLOE

Year 1:
 Four-state .021 .001
 H vs. D .013
 H vs. B .012
 H vs. L

Year 2:
 Four-state .034
 H vs. D .006
 H vs. B .007
 H vs. L .050 .002

Year 3:
 Four-state
 H vs. D
 H vs. B
 H vs. L .004

 Predictor Variables
 Response (1)
Year Variable CATA CASHTA CFFO CFFI CFFF

Year 1:
 Four-state .001
 H vs. D .006 .023
 H vs. B .025
 H vs. L .008 .030

Year 2:
 Four-state .021 .040
 H vs. D .018 .026
 H vs. B
 H vs. L .004

Year 3:
 Four-state .043
 H vs. D .049
 H vs. B
 H vs. L .034 .047

(1) Four-state = ordinal four-state response. H vs. D = healthy versus
distressed response. H vs. B = healthy versus bankrupt response. H vs.
L = healthy versus loan default/accommodation response.

(2) SIZE = log (total assets). NITA = net income/total assets.
SALESCA = sales/current assets. CACL = current assets/current
liabilities. TLOE = total liabilities/owners' equity. CATA = current
assets/total assets. CASHTA = cash + marketable securities/total
assets. CFFO = cash flow from operating activities. CFFI = cash flow
from investing activities. CFFF = cash flow from financing activities.

Table 3: Predictor Variables Significant at P-value [less than or equal
to] .05, Samples Combined

 Predictor Variables

 Response (1)
Year Variable SIZE (2) NITA SALESCA CACL TLOE

Year 1:
 Four-state .003 .001
 H vs. D .002 .017
 H vs. B .002 .006
 H vs. L .047 .001
Year 2:
 Four-state .001 .013
 H vs. D .001
 H vs. B .016 .008
 H vs. L .006
Year 3:
 Four-state .004 .017
 H vs. D .022
 H vs. B .015 .057
 H vs. L .007

 Predictor Variables

 Response (1)
Year Variable CATA CASHTA CFFO CFFI CFFF

Year 1:
 Four-state .007 .001
 H vs. D .005 .002
 H vs. B
 H vs. L .014
Year 2:
 Four-state .012 .014
 H vs. D .014 .016 .031
 H vs. B
 H vs. L .013 .032
Year 3:
 Four-state .003 .043
 H vs. D .007
 H vs. B
 H vs. L .025

(1) Four-state = ordinal four-state response. H vs. D = healthy versus
distressed response. H vs. B = healthy versus bankrupt response. H vs.
L = healthy versus loan default/accommodation response.

(2) SIZE = log (total assets). NITA = net income/total assets.
SALESCA = sales/current assets. CACL = current assets/current
liabilities. TLOE = total liabilities/owners' equity. CATA = current
assets/total assets. CASHTA = cash + marketable securities/total
assets. CFFO = cash flow from operating activities. CFFI = cash flow
from investing activities. CFFF = cash flow from financing activities.

Table 4: Classification Rates of Accrual and Mixed Models Using Various
Responses Variables for Financial Distress--Original Sample

 Response Scale

Year Model 4-State H vs. D H vs. B H vs. L

Year-1:
 Accrual Model:
 Total (1) 76.2 76.7 90.7 87.7
 H 98.8 78.7 91.5 89.0
 D, B, or L 17.5 71.4 83.3 78.3
 Mixed Model:
 Total 78.9 77.1 88.5 90.4
 H 98.8 78.0 89.6 91.5
 D, B, or L 27.0 74.6 77.8 82.6

Year-2:
 Accrual Model:
 Total 73.6 71.4 81.9 87.2
 H 98.2 72.6 83.5 87.2
 D, B, or L 9.5 68.3 66.7 87.0
 Mixed Model:
 Total 75.8 74.0 86.8 86.6
 H 100.0 75.0 87.2 88.4
 D, B, or L 12.7 71.4 83.3 73.9

Year-3:
 Accrual Model:
 Total 71.8 64.3 71.4 78.6
 H 98.8 65.2 72.6 78.7
 D, B, or L 1.6 61.9 61.1 78.3
 Mixed Model:
 Total 72.7 64.8 68.7 76.5
 H 97.0 65.9 70.1 78.0
 D, B, or L 9.5 61.9 55.6 65.2

(1) Total = total percentage of firms classified correctly by each
model for the different response scales. H = number of healthy,
nonbankrupt, or nonloan default/accommodation firms classified
correctly. D, B, or L = percentage of distressed, bankrupt, and/or loan
default/accommodation firms classified correctly.

Table 5: Classification Rates of Accrual and Mixed Models Using Various
Responses Variables for Financial Distress--Holdout Sample

 Response Scale

Year Model 4-State H vs. D H vs. B H vs. L

Year-1:
 Accrual Model:
 Total (1) 70.9 81.0 88.2 85.6
 H 96.4 82.9 92.8 88.3
 D, B, or L 10.6 76.6 56.3 64.3
 Mixed Model:
 Total 72.2 76.6 87.4 86.4
 H 96.4 81.0 93.7 89.2
 D, B, or L 14.9 63.8 43.8 64.3

Year-2:
 Accrual Model:
 Total 71.5 70.9 81.1 83.2
 H 98.2 75.7 82.9 87.4
 D, B, or L 8.5 59.6 67.8 50.0
 Mixed Model:
 Total 70.9 67.1 84.3 76.0
 H 96.4 70.3 87.4 81.1
 D, B, or L 10.6 59.6 62.5 35.7

Year-3:
 Accrual Model:
 Total 70.3 60.8 67.7 72.0
 H 100.0 63.1 68.5 73.9
 D, B, or L 0.0 55.3 62.5 57.1
 Mixed Model:
 Total 70.9 58.2 69.3 71.2
 H 99.1 61.3 69.4 73.0
 D, B, or L 4.3 51.1 68.8 57.1

(1) Total = total percentage of firms classified correctly by each
model for the different response scales. H = number of healthy,
nonbankrupt, or nonloan default/accommodation firms classified
correctly. D, B, or L = percentage of distressed, bankrupt, and/or loan
default/accommodation firms classified correctly.

Table 6: Predictor Variables Significant at P-value [less than or equal
to] .05, using a Nonbankrupt versus Bankrupt Response Measure for
Financial Distress

 Predictor Variables
 Response (1)
Year Variable SIZE (2) NITA SALESCA CACL TLOE

Year 1:
 NB vs. B .001 .001
Year 2:
 NB vs. B .001
Year 3:
 NB vs. B

 Predictor Variables
 Response (1)
Year Variable CATA CASHTA CFFO CFFI CFFF

Year 1:
 NB vs. B .012 .017
Year 2:
 NB vs. B .035 .047
Year 3:
 NB vs. B

(1) NB vs. B = nonbankrupt versus bankrupt response.

2 SIZE = log (total assets). NITA = net income/total assets.
SALESCA = sales/current assets. CACL = current assets/current
liabilities. TLOE = total liabilities/owners' equity. CATA = current
assets/total assets. CASHTA = cash + marketable securities/total
assets. CFFO = cash flow from operating activities. CFFI = cash flow
from investing activities. CFFF = cash flow from financing activities.