International Research Journal of Finance and Economics ISSN 1450-2887 Issue 15 (2008) © EuroJournals Publishing, Inc. 2008 http://www. eurojournals. com/finance. htm Predicting Corporate Failure of Malaysia? s Listed Companies: Comparing Multiple Discriminant Analysis, Logistic Regression and the Hazard Model Nur Adiana Hiau Abdullah Universiti Utara Malaysia, Associate Professor Faculty of Finance and Banking, Universiti Utara Malaysia Sintok 06010, Kedah, MALAYSIA E-mail: [email protected] edu. my Tel: 00-604-9286464/006013 5306566; Fax: 00-604-9286406 Abd.
Halim Universiti Utara Malaysia Hamilton Ahmad Universiti Utara Malaysia Rohani Md. Rus Universiti Utara Malaysia Abstract This study compares three methodologies for identifying financially distressed companies, multiple discriminant analysis (MDA), logistic regression and hazard model. In a sample of 52 distressed and non-distressed companies with a holdout sample of 20 companies, the predictions of the hazard model were accurate in 94. 9 % of the cases examined. This was a higher accuracy rate than generated by the other two methodologies.
However, when the holdout sample is included in the sample analyzed, MDA had the highest accuracy rate at 85%. Among the ten determinants of corporate performance examined, the ratio of debt to total assets was a significant predictor of corporate distress regardless of the methodology used. In addition, net income growth was another significant predictor in MDA, whereas the return on assets was an important predictor when the logistic regression and hazard model methodologies were used. Keywords: Bankruptcy, Multiple Discriminant Analysis, Logistic Regression, Hazard Model JEL Classification Codes: G33, C51
I. Introduction The sudden currency crisis in 1997 has thrown many financially strong companies out of business. All because they were not able to face the challenges and the unexpected changes in the economy. The growing economy suddenly became an alien to them when depression took place in a split second. As a 202 International Research Journal of Finance and Economics – Issue 15 (2008) result, many companies were forced into bankruptcy or became a financially distressed company when they were not able to pay their financial obligations due to inadequate cash flows.
Looking at the above situation, it is important to understand the reasons behind the collapse of a company. Knowing these reasons might hinder a company from being financially distress and early actions could be taken as a precaution. Studies in Malaysia have looked into this area, and have used models such as the multiple discriminant analysis (MDA), the logit model or a combination of both models. However, this study takes a different approach where a comparison of three models? MDA, logistic regression and hazard model? is implemented.
The motivation for this study arises from the arguments made by several authors who claimed that MDA suffered from serious drawbacks. Some of these drawbacks were regarding the assumptions of similar variance covariance matrices and linear distributions of independent variables that might lead to invalid results. Logit, on the other hand, uses data averages where a healthy company is given a value of 0 and a distressed company is given a value of 1. Hence, the logit model treats bankrupt companies as if they were bankrupt ever since their inception. A hazard model is able to overcome this problem by examining all firm year observations.
These companies would only be assigned as a distressed company in the year they became problematic or distress; otherwise, they are being treated as a healthy company. In comparing the MDA, logistic and hazard models by using a US data set, Shumway (2000) claimed that the hazard model was more reliable and accurate in predicting distress or bankrupty. Based on the above arguments, we are trying to examine the outcome from these different techniques, and to determine which variables appear significant in all the three models. Furthermore, we want to determine the most rigorous model in detecting distressed companies.
The remaining part of this study is organised as follows. A discussion on previous studies would be covered in Section Two, which is subsequently followed by an explanation of the data and research design use to answer the objectives of this study in Section Three. Section Four is an analysis of the results coming from the MDA, logistic and hazard models. The conclusion forms Section Five of this study. II. Literature Review Previous bankruptcy research had identified many ratios that were important in predicting bankruptcy. However, there was no conclusive agreement of which ratios were most useful to assess the likelihood of failure.
Altman (1993) noted that ratios measuring profitability, liquidity, solvency and cash flow were the most significant indicators of bankruptcy. The priorities were not clear as most studies cited different ratios being the most effective indicator of bankruptcy. Moreover, most of the studies did not have absolute test for the importance of variables (Barnes, 1987; Altman, 1993; Mohamed, Li and Sanda, 2001). The lack of theoretical support in choosing the appropriate variable that could predict bankruptcy had led researchers to search for other guides in variable selection.
Therefore, most researchers have selected financial ratios as predictor variables based on their popularity and predictive ability in the previous bankruptcy research studies (Beaver, 1966; Altman, 1968; Ohlson, 1980; Frydman, Altman and Kao, 1985; Casey and Bartczak, 1985). Among the most popular financial ratios used by researchers were net income to total assets (Beaver, 1966; Deakin, 1972; Libby, 1975; Ohlson, 1980; Lennox, 1999), total liabilities to total assets (Beaver, 1966; Deakin, 1972; Ohlson, 1980; Zmijewski, 1984) and size (Ohlson, 1980; Lennox, 1999; Shumway, 2001).
Furthermore, Ohlson (1980) added changes in net income as a factor that represents growth. Lennox (1999) utilised cash flow ratios, specifically cash to current liabilities, debtor turnover ratio and gross cash flow ratio to explain bankruptcy in the UK. In Korea, Nam and Jinn (2000) stated that financial expenses to sales, debt coverage and receivables turnover were important to explain bankruptcy. In contrast to Nam and Jinn (2001) but consistent with Lennox (1999), Low et al. 2001) found that in Malaysia the cash flow ratios were significant in explaining bankruptcy during the period 1996-1998; while Mohamed, Li and Sanda, (2001) found that the leverage ratio and efficiency ratio (total asset turnover) were found to be significant during the period 1987 to 1997. Low et al. (2001) International Research Journal of Finance and Economics – Issue 15 (2008) 203 were using the logit model whereas Mohamed et al. (2001) combined both the MDA and logit model. In another study by Zulkarnain et al. 2001) that concentrated on the MDA model, it was found that total liabilities to total assets, sales to current assets, cash to current liabilities and market value to debt were significant in explaining financial distress in Malaysia during the period 1980 to 1996. Although there were quite a number of possible financial variables available to predict bankruptcy, researchers were neither guided nor constrained by the theory for the selection of ratios (Scott, 1981). Therefore, the important criterion would be to choose ratios based on their simplicity and relevancy to the local environment (Chen & Shimerda, 1981; Mohamed et al. 2001; Soo et al. , 2001). When we look at the development of bankruptcy prediction model, it started with the use of univariate analysis by Beaver (1966), followed by multivariate discriminant analysis (MDA) by Altman in 1968. Beaver? s (1966) univariate analysis used individual financial ratios to predict distress. By using 79 failed and non-failed companies that were matched by industry and assets size in 1954 to 1964, his results from the prediction error tests suggested that cash flow to total debt, net income to total asset and total debt to total assets have the strongest ability to predict failure.
These ratios differed from the MDA model proposed by Altman (1968). By utilizing 33 bankrupt companies and 33 nonbankrupt companies over the period 1946 to 1964, five variables were selected on the basis that they did the best overall job in predicting bankruptcy. These were working capital to total assets, retained earnings to total assets, earnings before interest and taxes to total assets, market value of equity to book value of total debt and sales to total assets. Z-Score was determined and those companies with a score greater than 2. 99 fall into the non-bankrupt group, while those companies having a Z-Score below 1. 1 were in the bankrupt group. The area between 1. 81 and 2. 99 is defined as the zone of ignorance or the gray area. The cut-off index that made the most accurate prediction of bankruptcy one year before filing for bankruptcy was 2. 675. The MDA model was able to provide a high predictive accuracy of 95 % one year prior to failure. For this reason, MDA model had been used extensively by researchers in bankruptcy research (Altman, Haldeman and Narayanan, 1977; Apetiti, 1984; Izan, 1984, Micha, 1984; Shirata, 1998; Ganesalingam and Kumar, 2001).
Some of the recent findings in Japan (Shirata, 1998) and Australia (Ganesalingam and Kumar, 2001) showed an accuracy rate of 86. 14% and 81. 7%. However, Eisenbeis (1977), Ohlson (1980), and Jones (1987) found that there were some inadequacies in MDA with respect to the assumptions of normality and group dispersion. The assumptions were often violated in MDA. This may biased the test of significance and estimated error rates. Logit analysis which did not have the same assumptions as MDA was made popular by Ohlson (1980). He used 105 bankrupt companies and 2058 non-bankrupt companies from 1970 to 1976.
The results showed that size, financial structure (total liabilities to total assets), performance and current liquidity were important determinants of bankruptcy. In the logit analysis, average data is normally used and it is considered as a single period model. Hence, for each non-distressed and distressed company, there is only one company-year observation. The dependent variable is categorized into one of two categories that is distressed or non-distressed. There are two econometric problems with the single period logit model (Hillegeist, 2004).
First, is the sample selection bias that arises from using only one, non-randomly selected observation for each bankrupt company, and second, the model fails to include time varying changes to reflect the underlying risk of bankruptcy. This will induce crosssectional dependence in the data. Shumway (2001) demonstrated that these problems could result in biased, inefficient, and inconsistent coefficient estimates. To overcome these econometric problems, Shumway (2001) predicted bankruptcy using the hazard model and found that it was superior to the logit and the MDA models.
As mentioned earlier on, several studies were also implemented in Malaysia. Zulkarnain et al. (2001) used twenty-four distressed and non-distressed companies from the period 1980-1996 matched according to the industry, failure year, closest asset size and age since incorporation. Distressed companies were defined as those companies that resorted protection under section 176 of the Companies Act 1965 for the purpose of obtaining court protection against their creditors. Utilising the stepwise multivariate discriminant analysis to determine the discriminating variable, they compared the results from using market based ariables and without market based variables. They found that total 204 International Research Journal of Finance and Economics – Issue 15 (2008) liabilities to total assets, sales to current assets, cash to current liabilities and market value to debts were important determinants of corporate failures in Malaysia. The original model with market based variables correctly classified 89. 7% of the sample whereas the other model only correctly classified 87. 9% of the sample. Using the same definition of failure, Low et al. (2001) analysed financial distress using the logit analysis.
They utilised 26 distressed companies and 42 non-distressed companies in 1988. The hold-out sample consisted of 10 companies. They found that sales to current assets, current assets to current liabilities, change in net income, cash and marketable securities to total assets were significant determinants of financial distress. However, the coefficients of the first three variables were not as expected when a significant positive coefficient prevail. Therefore they claimed that measures of liquidity and profitability may be misleading, and concluded that only the cash flow ratio served as an indicator to detect potential failure of a company.
The accuracy rate is 82. 4% in the estimation sample and 90% in the hold-out sample. Mohamed et al. (2001) compared the MDA and the logit model in the analysis of bankruptcy. Their sample consisted of 26 companies that have sought protection under section 176 of Companies Act 1965 and 79 non-distressed companies. Their results showed that when using MDA, debt ratio and total assets turnover were found to be significant but when logit analysis was used, an additional variable, interest coverage was also found to be significant. Thus, Mohamed et al. 2001) study emphasized the importance of leverage ratio as a predictor of failure. The logit model predicted 80. 7% of the companies in the estimation sample and 74. 4% in the hold-out sample, whereas the MDA model predicted 81. 1% of the companies in the estimation sample and 75. 4% in the hold-out sample. The accuracy rate of Mohamed et al. (2001) prediction model was lower than Low et al. (2001) and Zulkarnain et al. (2001). In these studies, none of them have used the hazard model which includes the time varying changes to reflect the underlying risk of distressed companies.
Hence, our studies were implemented to fill the gap of comparing the predictive accuracy of MDA, logit analysis and hazard model and to examine which among the variables were essential in predicting companies in distress. III. Research Design The sample consisted of both distressed and non-distressed listed companies in the Bursa Malaysia Berhad. The distressed status was indicated by the appointment of receivership, restraining order under section 176, winding up petition, special administrator under Bank Negara Malaysia and interim judicial management order as at December 2000.
A total of thirty-six distressed companies were identified from the Bursa Malaysia daily diary. For each distressed companies, a non-distressed match was identified during the period from 1990 to 2000. Companies were matched if they belonged to the same industry group and have the closest asset size. A one to one procedure is consistent with the previous studies documented in Beaver (1966), Altman (1968) and Blum (1974) and is an acceptable method in failure prediction studies. Financial data for both groups were collected from the annual reports in the Bursa Malaysia and Sultanah Bahiyah library.
Companies were excluded from the sample if they were classified under the financial and property industries. The reasons for these are that their ratios are highly volatile where their businesses rely heavily on the economy. In addition, the interpretation of the ratios is slightly different because the nature of the income and expenses for these companies is different from that of non-financial companies. Method For model building, multiple discriminant analysis (MDA) which takes the form of Z = 1Va + 2Vb + ? nVn based on a stepwise approach is adopted to select the best discriminating variables that could predict distressed and non-distressed companies. This model would then be compared to the logit analysis and hazard model to examine which among the three could provide a higher accuracy in International Research Journal of Finance and Economics – Issue 15 (2008) 205 predicting financially distressed companies. The logit prediction model is adopted from Ohlson (1980), Gujarati (1995, pp. 554) and Joo and Jinn (2000).
Zi ‘ xi u i (1) Where: Zi = non-distressed if Zi ; 0 Zi = distressed, otherwise xi = companies financial ratios ui = error term Zi ranges from – to + The probability and likelihood function for the non-distressed can be defined as follows: 1 Pi E (Y 2 | xi ) (2) 1 e ( ‘ xi ui ) For ease of exposition, it is written as 1 Pi 1 e Zi ‘ xi u i where Z i Equation (2) represents what is known as the (cumulative) logistic distribution function. In order to apply the prediction model, the weights of the financial ratios were estimated in equation (1) using the financial ratios of listed companies.
If Pi represents the probability of nondistressed which is given in equation (2), then (1-Pi), would be the probability of distressed. Hence, 1 1 Pi (3) 1 e Zi Optimal (weights) can be estimated where the likelihood value is maximized. The probability of distressed is obtained by substituting into the cumulative probability function. A company is classified as distressed if the calculated probability from the logit model is more than 0. 5, otherwise it would be non-distressed. Similar to the discriminant analysis, a forward stepwise method is adopted in logistic regression.
This procedure would enable the predictor variables to be entered based upon their contribution to the likelihood ratio statistics. Therefore, variables that do not contribute significantly to the statistics are not entered by the procedures. Soo et al. (2001) stressed that the lack of theoretical basis in selecting the independent variables was the main reason why a stepwise procedure is needed. A similar argument was made by Menard (1995) as he stated that stepwise methods were used when neither the theory nor knowledge correlates to the phenomenon.
In addition, the used of a stepwise procedure would at least reduce multicollinearity problems which make it difficult to make any statistical inferences. A discrete hazard model is applied to assess how well each variable explains the actual probability of bankruptcy in our sample. It is in the form: e i ,t (t ) (t ) X i,t X i ,t 1 e (4) where i,t is the hazard function, X represents a vector of explanatory variables used to forecast failure, (t ) is a time-varying covariates, and is the coefficient vector.
The discrete hazard model is well suited to analyse data that consists of binary, time series and cross-sectional observations such as bankruptcy data. The hazard model has the same likelihood function and the same asymptotic variancecovariance matrix as the logit model and therefore the estimation of the hazard model is similar to the estimation of the logit model. However, the hazard model uses time varying covariates and companyyear observations and consequently it is able to eliminate the sample selection bias. It will also results in more efficient coefficient estimates since all available data will be used in the estimation. 06 International Research Journal of Finance and Economics – Issue 15 (2008) In the discrete hazard model the dependent variable is coded as 1 if company i failed at time ti, and equal zero otherwise. For example if a company has been in existence for six years and was classified under section 176 in year 6, only year 6 will have the value of 1, the other 5 years will have the value of 0 indicating that the company is healthy during those years. Since the hazard model uses company-year data, adjustment has to be made to the test statistics from the logit model.
We divide the test statistics from the estimation by the average number of company-years per company because the correct value of n for test statistics in the logit analysis is the number of companies in the data. The discrete hazard model uses company-year data and has several advantages (Shumway, 2001). By using all company-year observations, it is able to eliminate the sample selection bias, produces more efficient out of sample forecasts and is able to adjust for risk automatically. It is also possible to track changes in bankruptcy probability since all data in each year are included in the analysis.
The independent variables used in this study can be classified according to the different set of ratios. They are leverage ratios (interest coverage and total debt to total assets), profitability ratios (net income to total assets), cash flow ratios (cash to total assets and cash to current liabilities), size (total assets employed), and growth (change in net income and change in sales). As noted by Scott (1981), many of the variables that appeared in most empirical work do not rest on any strong underlying theory, hence the use of these ratios in our study is acceptable.
These ratios are selected based on the popularity of their usage in the literatures and the predictive success stated in previous research. Financial leverage is related to bankruptcy to the extent to which a company relies on debt financing rather than equity. Measures of financial leverage are tools in determining the probability of a company defaulting on its debt contracts. Debt ratio which is calculated by dividing total debt to total assets provides information on a company? s insolvency and its ability to secure additional financing for good investment opportunities. This is to ensure that creditors are protected.
As for interest coverage which is measured by dividing earnings with interest, it emphasizes the ability of a company to generate enough income to cover interest expense. Beaver (1966), Deakin (1972), Ohlson (1980) Zmijewski (1984) and Mohamed et al. (2001) find that these ratios were significant determinants of corporate failure. Profitability ratio is represented by return on assets, computed as net income divided by total assets. This ratio is a common measure of managerial performance and is therefore vital in the study of financial distress. Libby (1985), Ohlson (1980), Lennox (1999) and Zulkarnain et al. 2001) showed that profitability is an important determinant of bankruptcy. It is expected that companies with large profits have a lower probability of bankruptcy, hence the relationship between them is negative. In addition to the above ratios, short term solvency is also an important element to be looked into as it measures the ability of a company to meet its short term financial obligations, thus avoiding corporate failures. Cash flow ratio, represented by cash to total assets and cash to current liabilities are used as a proxy to measure short-term solvency for distressed and non-distressed companies.
Lennox (1999), Low et al. (2001) and Zulkarnain et al. (2001) found that cash flow ratios were found to be significant in their studies. It is expected that the relationship between cash flow ratios and the probability of bankruptcy is negative, the higher the level of cash flow, the lower is the probability of bankruptcy. Another factor that seems to discriminate between distressed and non-distressed companies is size, which is measured by total assets employed. Big companies normally have large assets base if compared with smaller companies.
Ohlson (1980) found that size was significant in discriminating between distressed and non-distressed companies. It is expected that the relationship between these two variables is negative, the larger the size of a company, the lower the probability of bankruptcy. Other ratios that could probably discriminate between healthy and distressed companies are change in net income and change in sales. The rationale behind these ratios is that healthy company? s net income and sales grow rapidly as compared to distressed companies. Hence, it is expected that the greater the growth, the healthier is the company.
International Research Journal of Finance and Economics – Issue 15 (2008) 207 IV. Analysis of Results A descriptive statistic of the variables used to estimate the MDA, logit and hazard models is presented in Table 1. As expected, the mean for interest coverage is lower for the non-distressed companies for both the MDA and logit data set as well as the hazard model. Healthy companies can cover 278. 78 times of their interest as compared to 7. 26 times for distressed companies in the MDA and logit model whereas the hazard sample shows that healthy companies can cover 95. 1 times of their interest but distressed companies are not able to have that luxury as the figure drops to -. 52 times. It appears that distressed companies rely heavily on debt, which is approximately 84. 06%; whereas the build up of debt for healthy companies is only 41. 13% of the MDA and logit sample. This scenario is even worst for the hazard sample as distressed companies are having a debt ratio of 247. 31%, which is much higher in comparison to the healthy companies that carry approximately 60. 51% debt. Table 1: Descriptive statistics of independent variables
Panel A: MDA and Logit Interest Cover Debt/ Assets Net Income/ Total Asset Return On Equity Cash/ Total Asset Cash/ Current Liabilities Net Income Growth Sales Growth CA/CL LnTA Non-distressed Companies Mean Std. Dev 278. 77796 854. 084445 . 41136 . 171435 . 07289 . 041282 . 11989 . 056202 . 02088 . 020178 . 09950 . 118207 1. 10809 2. 380989 . 54082 1. 702728 2. 29603 2. 023123 18. 94557 1. 343271 Distressed Companies Mean Std. Dev 7. 26027 8. 466179 . 84065 . 340254 -. 12246 . 179765 -2. 50187 14. 160312 . 02092 . 015956 . 04601 . 032689 -2. 2079 3. 159702 . 50758 3. 552954 1. 28472 . 756635 19. 25582 . 951891 Panel B: Hazard Interest Cover Debt/ Assets Net Income/ Total Asset Return On Equity Cash/ Total Asset Cash/ Current Liabilities Net Income Growth Sales Growth CA/CL LnTA Non-distressed Companies Mean Std. Dev 95. 214457 1061. 91926 . 605168 . 5127353 -. 024829 . 3098230 -1. 274714 28. 8747013 . 019190 . 0258730 . 067988 . 1116086 -. 5477 7. 51348 . 8033 8. 46602 1. 804498 2. 7495946 19. 035820 1. 3861094 Distressed Companies Mean Std. Dev -. 520106 1. 9425379 2. 473130 2. 252016 -. 361832 . 5004877 . 306871 . 5974697 . 018389 . 0237851 . 012332 . 0148575 -. 2177 1. 44923 -. 2043 . 52036 . 337737 . 2787643 19. 094047 1. 3412095 If we were to look at the current ratio, for every RM1 of current liabilities, there is a support of RM1. 28 and RM0. 34 from current assets for distressed companies in the respective MDA/logit and hazard samples. This ratio is 0. 7 (MDA/logit) and 5 (hazard) times higher for healthy companies where every RM1 of current liabilities is covered with RM2. 30 and RM1. 80 of current assets.
Cash to total asset ratio for both groups are almost equivalent, which is about two percent; but when it comes to cash against current liabilities, healthy companies could cover RM0. 10 and RM0. 07 for every RM1 of current liabilities in the respective MDA/logit and hazard data set. This amount reduces to RM0. 07 and RM0. 01 for distressed companies, an indication that these companies would probably have difficulties to meet their short term financial obligations. During the period of study, the net income to total assets (ROA) is found to be negative for distressed companies with a figure of ? . 122 and ? 0. 361 for both data sets. This figure is slightly better for the healthy companies where it shows for every one ringgit of total asset, these companies are producing seven cents of net income for the MDA/logit sample whereas for the hazard sample, it 208 International Research Journal of Finance and Economics – Issue 15 (2008) shows a loss of 2. 5 cents for every ringgit of total asset. As for the return on equity, the MDA/logit data provide an 11. 99% and -250. 19% for the healthy and distress companies respectively.
There is a sign change between the two groups of companies for the hazard data where healthy companies are providing -127. 47% return on equity, whereas distressed companies are giving 30. 69% return to their shareholders. As expected, net income growth for distressed companies is -242. 08% and -21. 77% for the respective MDA/logit and hazard data sets; whereas healthy companies net income growth is 110. 81% for the MDA/logit data but -54. 77% for the hazard data. Surprisingly, the negative growth shown by distressed companies is smaller than the growth shown by healthy companies for the hazard data.
If we are to observed the sales growth for this data set, healthy companies are having better growth of 80. 33% than distressed companies that are having a growth of -20. 43%. As for the MDA/logit data set, the sales growth is 54. 08% and 50. 76% for the respective healthy and distressed companies. There was not much difference in the size of those companies that are healthy or distress. Some of the unexpected descriptive statistics show us that the use of an average figure in the MDA/logit might not represent a true picture of companies? haracteristics as observed in the hazard data set which take into consideration of the time varying covariates and company-year observations. Table 2 shows the univariate analysis to identify ratios that have the highest ability to differentiate between financially distressed and non-distressed companies for the MDA /logit and hazard data sets. The results show that variables with a mean difference that is significant at the 5 % level are debt to total assets, net income to total assets, cash to current liabilities, net income growth and current ratio for the MDA/logit sample.
These ratios with the exception of net income growth are also found to have significant mean differences between the healthy and distressed companies for the hazard sample. Table 2: Mean differences between distressed and non-distressed companies Panel A: MDA and Logit Variables Interest coverage Debt to total assets Net Income to total asset Return on equity Cash to total assets Cash to current liabilities Net income growth Sales growth CA/CL LnTA Distress 7. 2603 0. 8406 -0. 1225 -2. 5019 0. 0209 0. 0460 -2. 4208 0. 076 1. 2847 19. 2558 Means Non-Distress 278. 7780 0. 4114 0. 0729 0. 1199 0. 0209 0. 0995 1. 1081 0. 5408 2. 2960 18. 9456 Mean Differences t- statistics sig 1. 621 0. 111 -5. 745 0. 000* 5. 400 0. 000* 0. 944 0. 350 -0. 008 0. 993 2. 224 0. 031* 4. 548 0. 000* 0. 043 0. 966 2. 387 0. 021* -0. 961 0. 341 Panel B: Hazard Variables Interest coverage Debt to total assets Net Income to total asset Return on equity Cash to total assets Cash to current liabilities Net income growth Sales growth CA/CL LnTA * significant at = 0. 05
Means Distress -. 520106 2. 473130 -. 361832 . 306871 . 018389 . 012332 -. 2177 -. 2043 . 337737 19. 094047 Non-distress 95. 214457 . 605168 -. 024829 -1. 274714 . 019190 . 067988 -. 5477 . 8033 1. 804498 19. 035820 Mean Differences t-statistics sig . 459 . 646 -12. 860 . 000* 5. 134 . 000* -. 279 . 780 . 153 . 878 2. 539 . 011* -. 223 . 823 . 606 . 545 2. 716 . 007* -. 208 . 835 International Research Journal of Finance and Economics – Issue 15 (2008) 209 Table 3 and Table 4 present the correlation matrix among the variables.
It is shown that the pairwise correlations among the variables are uniformly low and insignificant except for several ratios: interest cover against debt/asset, interest cover against cash/current liabilities, interest cover against CACL, net income/total asset against return on equity (ROE), cash/total asset against net income growth for the MDA/logit sample; debt/asset against net income/total asset, debt/asset against cash/current liabilities, debt/asset against net income growth, debt/asset against current ratio (CACL), net income/total asset against net income growth, cash/total asset against cash/current liabilities, cash/current liabilities against CACL, net income growth against sales growth, net income growth against CACL for both the MDA/logit and hazard data set; net income/total asset against cash/current liabilities, net income/total asset against CACL, cash/total asset against size (lnTA), cash/current liabilities against lnTA, sales growth against lnTA and CACL against lnTA for the hazard sample. 212 International Research Journal of Finance and Economics – Issue 15 (2008) Four of the independent variables from the MDA/logit and hazard samples are highly correlated, such as shown by the value of -0. 832 (MDA/logit) and -0. 706 (hazard) for debt to asset against net income to total asset and 0. 766 (MDA/logit) and 0. 741 (hazard) for cash to total asset against cash to current liabilities. In addition to this, the MDA/logit sample shows that interest cover and current ratio are also highly collinear with a value of 0. 72. In general, the correlation coefficients of the MDA/logit sample are higher than those shown by the hazard sample.
The correlation coefficients would probably support the existence of multicollinearity problem between these variables. It is noted that the identification of these ratios is not related to any theoretical base except for the popularity of their usage and the predictive success that came from previous research. We could simply drop these ratios, but it is likely that this remedy could probably be worse than the problem of collinearity itself. We re-examine the independent variables to check on the seriousness of the multicollinearity problem in our data by looking at the Variance Inflating Factors (VIF). It is the ratio of a variable? s actual variance to the perfect variance of zero collinearity.
If we were to refer to Table 5, the results show that the R2 is rather low for most of the variables except for cash to current liabilities that shows a figure of 0. 68 for the hazard data; but the R2 for the MDA/logit data is quite high for some variables such as those shown by debt to asset, net income to total asset, cash to total asset, cash to current liabilities, net income growth and current ratio. Nevertheless, when the VIF is calculated, all the variables present a figure below 10. For the MDA/logit sample, the VIF ranges in between 1. 127 to 6. 369 whereas for the hazard sample, it ranges in between 1. 011 to 3. 086. Hence, we can conclude that the degree of multicollinearity problem is not a threat to this study. Table 5: Variance inflating factors Panel A: MDA and Logit
Variables Interest cover against other independent variables Debt/asset against other independent variables Net income/total asset against other independent variables Return on equity against other independent variables Cash/total asset against other independent variables Cash/current liabilities against other independent variables Net income growth against other independent variables Sales growth against other independent variables CACL against other independent variables lnTA against other independent variables R2 0. 550 0. 780 0. 748 0. 113 0. 810 0. 843 0. 641 0. 194 0. 673 0. 119 VIF 1 (1 R 2 ) j 2. 222 4. 546 3. 968 1. 127 5. 263 6. 369 2. 786 1. 241 3. 058 1. 135 Panel A: Hazard
Variables Interest cover against other independent variables Debt/asset against other independent variables Net income/total asset against other independent variables Return on equity against other independent variables Cash/total asset against other independent variables Cash/current liabilities against other independent variables Net income growth against other independent variables Sales growth against other independent variables CACL against other independent variables lnTA against other independent variables R2 0. 011 0. 532 0. 537 0. 002 0. 587 0. 676 0. 156 0. 050 0. 381 0. 040 VIF 1 (1 R 2 ) j 1. 011 2. 137 2. 160 1. 002 2. 421 3. 086 1. 185 1. 053 1. 616 1. 042 Table 6 reports the results of the multiple discriminant analysis. It appears that debt to total assets is more important than net income growth in predicting financially distressed companies. The latter has the least discriminating power. Panel B of Table 4 shows that 73. 1 % of distressed companies and 88. 5 % of non-distressed companies were correctly predicted in the estimation sample. This implies an overall prediction accuracy of 80. 8 % in the estimation sample.
The model is then used to International Research Journal of Finance and Economics – Issue 15 (2008) 213 predict distress in the holdout sample. The result shows that the model could correctly classified 90 % of distressed companies and 80 % of non-distressed companies. This has brought to an overall classification accuracy rate of 85 % in the holdout sample. Table 6: Panel A Discriminant analysis Coefficient 2. 764 -0. 168 Significant 0. 000 0. 000 Variable Debt to Total Assets Net Income Growth * significant at = 0. 05 Panel B Percentage correctly predicted Distressed Non-Distressed Overall Estimation sample 73. 1 88. 5 80. 8 Holdout Sample 90 80 85
The discriminant analysis model accuracy prediction is slightly higher than the model of Ganesalingam and Kumar (2001), which is conducted in Australia. With regard to the Malaysian study, the predictive accuracy is lower than the predictive accuracy of Zulkarnain et al. (2001) model. This would probably be due to the different sample and variables utilized by Zulkarnain et al. (2001). Moreover, Zulkarnain et al. (2001) is focusing on the industrial sector whereas this study covered seven different industrial sectors in Malaysia. As compared to a recent study by Zulkarnain et al. (2002), their model could accurately classified 88. 1 % of companies in istress whereas the predictive accuracy of the discriminant model in this study is 80. 8 %. The results of the stepwise logistic regression are presented in Panel A and Panel B of Table 7. The log likelihood statistic tests the null hypothesis where the coefficients of independent variables in the model are zero. Panel A shows that among the ten variables, only two ratios? debt to total assets and net income to total assets? are found to be significant based on the Wald statistic. However, according to Menard (1995, pp. 38) a likelihood ratio (LR) test is more accurate in evaluating the statistical significance of the contribution of an independent variable to the explanation of a dependent variable.
Wald statistic normally gives an inflated standard error, which could result in a failure to reject the null hypothesis when the null hypothesis is false. Therefore, likelihood ratio test is adopted in this study as it is more reliable. Table 7: Stepwise logistic regression Panel A: Variables entering the model: Wald statistic Variable Debt to Total Assets Net Income to Total Assets Constant * significant at = 0. 05 Coefficient 10. 539 -21. 450 -5. 789 Wald 5. 745 5. 331 4. 824 Significant 0. 017* 0. 021* 0. 028* Panel B: Variables entering the model: likelihood ratio test Variable Debt to Total Assets Net Income to Total Assets Coefficient 10. 539 -21. 450 Change in ? 2 Log Likelihood 8. 963 9. 264 Significant 0. 003* 0. 002* 2 Log Likelihood 45. 850 with 2 degrees of freedom (p=0. 000) * significant at = 0. 05 214 International Research Journal of Finance and Economics – Issue 15 (2008) Estimation sample 80. 8 84. 6 82. 7 Holdout Sample 90 70 80 Panel C: Classification results Percentage correctly predicted Distressed Non-Distressed Overall Panel B shows that the p-values of the two variables (debt to total assets and net income to total assets) are less than 0. 05, which indicate that these variables are significant in contributing to the model and in predicting financial distress. In comparison to the MDA, only debt to total assets entered the logistic regression.
Furthermore, instead of net income growth, the logit model identifies return on asset as an important predictor of distressed companies. A significant negative coefficient for the return on asset variable suggests that companies with a higher proportion of net income to total assets are likely to experience financial distress. The unexpected sign contradicts to those appear in Beaver (1966), Deakin (1972), Libby (1975), Ohlson (1980) and Ward and Foster (1997). An explanation on this might be that the average data used in the logit model might have been affected by the financial crisis in 1997-1999 because during this period most companies have an unstable or volatile income. Panel C of Table 7 shows the classification results. 0. 8 % of distressed companies and 84. 6 % of the non-distressed companies were correctly classified in the estimation sample. When the coefficients of the estimated model are used to classify the holdout sample, 90 % of the distressed and 70 % of non-distressed companies are correctly classified. The overall accuracy rate for the estimation and the holdout sample is 82. 7 % and 80 % respectively. The predictive accuracy of the logistic regression model in this study is slightly better as compared to Mohamed et al. (2001). The findings of Mohamed et al. (2001) show 80. 7 % accuracy in their estimation sample and 74. 4 % in the holdout sample.
Soo et al. (2001) reported 82. 4 % accuracy for the estimation sample, which in this case about 0. 3 % lower than our accuracy rate, but their holdout sample is providing a higher accuracy rate of 90 % as compared to our result which shows an overall accuracy rate of 80 %. The third analysis is the hazard model. Panel A of Table 8 reports the determinants of financial distress. The results suggest that distressed companies in Malaysia could be determined by debt to total assets and return on asset ratios. These variables are similar to those found in the logit model. However, the positive coefficient of 2. 966 for net income to total asset is n contrast to the negative coefficient found in the logit model but similar to those reported in the Western world by Beaver (1966), Deakin (1972), Libby (1975), Ohlson (1980) and Ward and Foster (1997). This would mean that an analysis using the hazard model would probably be better as it takes into consideration of both the time varying factor and company year observations rather than taking an average in the logit analysis. Table 8: Hazard Model Panel A: Variables entering the model Variable Debt to Total Assets Net Income to Total Assets LnTA Constant * significant at = 0. 05 Coefficient 2. 907 2. 966 0. 516 -15. 384 Significant 0. 000* 0. 004* 0. 022 0. 001* Panel B: Classification results Percentage correctly predicted Distress Non-Distress Overall Estimation sample 38. 5 99. 1 94. 9 Holdout Sample 30. 0 97. 63. 9 International Research Journal of Finance and Economics – Issue 15 (2008) 215 The only variable that has appeared as a consistent indicator of financially distressed companies in all the models is leverage ratio that is debt to total assets. Its positive coefficient of 2. 764, 10. 539 and 2. 907 for the respective MDA, logit and hazard model shows that financially distressed companies carry a high level of debt. The probability of defaulting on debt contracts would likely be elevated if there is a sudden down turn of income in the companies. This is consistent to the results reported by Mohamed et al. (2001) and Zulkarnain et al. (2001).
In terms of the classification on the accuracy rate, the hazard model is able to correctly classify 38. 5 % of distressed companies and 99. 1 % of non-distressed companies. Overall, the model correctly classifies 94. 9 % of the estimation sample. Nevertheless, when it is applied to the holdout sample, the accuracy rate reduces to 30 % and 97. 7 % for the respective distress and non-distress groups. The overall accuracy rate decreases approximately 31 % to 63. 9 %. The finding in this study shows that the estimation sample of the hazard model has given an overall accuracy rate of 94. 9 % which is higher than the MDA and logit model which reported 80. 8 % and 82. 7 % respectively.
However, when the estimated equation of every model is applied to the holdout sample, the result changes where the MDA provides an overall accuracy rate of 85 % which is greater than the logit and hazard models that give a respective 80 % and 63. 9 % accuracies. Shumway (2001) claims that the hazard model is more reliable and accurate in predicting distress as compared to the MDA and logit model does not stand in the Malaysian context as there is a contradiction in the overall accuracy rate of the estimation and holdout samples. Another explanation would probably be that our own analysis might need to be refined in terms of using the enter or stepwise approach in analyzing the data.
Most of the studies implemented in Malaysia and some in the Western country argue that a stepwise approach is more appropriate as the variables included in the model are identified based on their significant and not through any underlying theories. As mentioned earlier on, in terms of the underlying assumptions of each model, MDA is quite restricted to its similar variance covariance matrices and linear distributions of the variables; and logit treatment of taking averages of the data as if these companies were in distress ever since their inception. The hazard model is supposed to be more robust in overcoming all the shortcomings of MDA and logit.
Hence, our future research would be focusing into running the analysis with the enter approach and comparing the results with the existing study. Furthermore, an analysis of the Type I and Type II errors for the enter and stepwise approaches for the MDA, logit and hazard model could also be executed to check on the reliability of the financial distress model. V. Conclusion Previous studies on Malaysian financially distressed companies have been emphasizing the use of MDA and logit analysis. However, studies done in the Western country have highlighted the inadequacies of both analysis and suggested that the hazard model gives a more accurate result than the MDA and logit model that suffer a few drawbacks with respect to the underlying assumptions.
This argument has provide an opportunity for us to investigate whether such claim is true for the Malaysian companies. Our study employs 26 distressed companies with a matched sample of another 26 nondistressed companies listed in Bursa Malaysia. Twenty companies which consist of ten distressed and another ten non-distressed companies matched based on the industry and size are used as the holdout sample. We find that the hazard model could accurately predict 94. 9 % and 63. 9 % of the respective estimation and holdout sample; whereas the MDA model provides an overall accuracy rate of 80. 8 % and 85 % for the estimation and the holdout sample respectively. For the logit model, it could correctly predict 82. % and 80 % of the respective estimation and holdout sample. The hazard model provides a higher overall accuracy rate in the estimation model, but when the estimated equation is applied in the holdout sample, MDA gives a higher accuracy rate. This would probably have been caused by the stepwise rather than enter approach adopted in the hazard model. 216 International Research Journal of Finance and Economics – Issue 15 (2008) Leverage ratio is found to be an important predictor of distressed companies in all the models. Its positive coefficient shows that the higher the debt, the higher is the probability of defaulting among financially distressed companies.
Other determinants that are identified to be essential in detecting distressed companies are net income growth for the MDA model and return on asset for the logit and hazard model. For the latter variable, the sign of the coefficient differs between the logit and hazard model. Logit furnishes a positive coefficient and hazard the opposite of it. The reason for this might be due to the treatment of data by both models. The contradictory results in the accuracy rate and the different sign on the coefficient of net income to total asset variable would be investigated in future research. References                    Altman, E I. 1968) “Financial Ratios, Discriminant Analysis and the Prediction of Corporate Bankruptcy”, Journal of Finance, 23, pp. 589-609. Altman, E I. (1993) Corporate financial distress and bankruptcy, 2nd. Ed. , Wiley, New York. Altman, E I. , Haldeman, R. G. , and Narayanan, P. (1977) ? ZETA Analysis? , Journal of Banking and Finance, pp. 29-54. Appetiti, S. (1984) ? Identifying unsound firms in Italy? , Journal of Banking and Finance, 8, pp 269-279. Barnes, P. (1987) ? The Analysis and Use of Financial Ratios: A Review Article? , Journal of Business Finance and Accounting, 14(4), pp. 449-461. Beaver, W. (1966) “Financial ratios as predictors of failure”, Journal of Accounting Research (Supplement), 4, pp. 1-102. Blum, M. (1974) ? Failing company discriminant analysis? , Journal of Accounting Research, 12, pp. 1-25. Casey, C. and Bartczak, N. , (1984) ? Cash flow: it? s not the bottom line? , Harvard Business Review, Vol. 4, pp. 60- 66. Chen, K. , and Shimerda, T. (1981) “An empirical analysis of useful financial ratios”, Financial Management, 10, pp. 51-60. Deakin, E. (1972) “A discriminant analysis of predictors of business failure”, Journal of Accounting Research, 10, pp. 167-179. Eisenbeis, R. (1977) ? Pitfalls in the application of discriminant analysis in business, finance and economics? , Journal of Finance, 32, pp. 875-900. Frydman, H. Altman, E. , and Kao, K. (1985) ? Introducing recursive partitioning for financial classification: The case of financial distress? , The Journal of Finance, XL (1), pp. 269-291. Ganesalingam, S. and Kumar, K. (2001) ? Detection of financial distress via Multivariate Statistical Analysis? , Managerial Finance, 27(4), pp. 45-55. Gujarati, Damodar N. (1995) Basic Econometrics. McGraw-Hill International Editions, 3rd Edition. Singapore. Hillegeist S. A. , Keating E. K. , Cram D. P. and Lundstedt K. G. (2004) ? Assessing the probability of Bankruptcy? , Review of Accounting Studies, 9, pp. 5-34. Izan, H. Y. (1984) ? Corporate distress in Australia? Journal of Banking and Finance, 8, pp. 320-330. Jones, F. L. (1987) “Current techniques in bankruptcy prediction”, Journal of Accounting Literature, 6, pp. 131-164. Joo, Ha Nam and Jinn Taehong (2000) ? Bankruptcy Prediction: Evidence from Korea Listed Companies During the IMF Crisis? , Journal of International Financial Management and Accounting, 11(3), 178-197. Libby, R. (1975) “Accounting ratios and the prediction of failure: some behavioral evidence”, Journal of Accounting Research, 13, pp. 150-161. International Research Journal of Finance and Economics – Issue 15 (2008)              217 Lennox, C. 1999) “Identifying failing companies: A re-evaluation of the logit, probit and MDA approaches”, Journal of Economics and Business, 51(4), pp. 347-364. Low, S. , Fauzias M. N. and Zainal Ariffin, A. (2001) ? Predicting corporate distress using logit model: The case of Malaysia? , Asian Academy of Management Journal, 6(1), pp. 49-62. Menard, S. (1995) Applied Logistic Regression Analysis. Sage University Paper series on Quantitative Applications in the Social Sciences, 07-106. Thousand Oaks, CA: Sage. Micha, B. (1984) ? Analysis of business failures in France? , Journal of Banking and Finance, 8, pp. 281-291. Mohamed, S. , Li, A. J. and Sanda A. U. (2001) ? Predicting corporate failure in Malaysia: An application of the Logit Model to financial ratio analysis?
Asian Academy of Management Journal, 6(1), pp. 99-118. Nam, J and Jinn T. (2000) ? Bankruptcy prediction: Evidence from Korean listed companies during the IMF crisis? , Journal of International Financial Management and Accounting, 11(3), pp. 178-197. Ohlson, J. A. (1980) “Financial ratios and the probabilistic prediction of bankruptcy”, Journal of Accounting Research, 18, pp. 109-131. Scott, J. (1981) ? The Probability of Bankruptcy: A Comparison of Empirical Predictions and Theoretical Models? , Journal of Banking and Finance, 5, 317-344. Shirata, Cindy Yoshiko (1998) Financial ratios as Predictors of Bankruptcy in Japan: An Empirical Research, Tsukuba College of Technology Japan: 1-17.
Shumway, T. (2001) “Forecasting bankruptcy more accurately: A simple hazard model”, Journal of Business, 74 (1), pp. 101-124. Ward, T. J. and Foster, B. P. (1997) ? A Note on Selecting A Response Measure for Financial Distress? , Journal of Business Finance and Accounting, 24, pp. 869-879. Zmijewski, M. E. (1984) “Methodological issues related to the estimation of financial distress prediction models”, Journal of Accounting Research, 22, pp. 59-86. Zulkarnain, M. S. , Mohamad Ali, A. H. , Annuar, M. N. and Zainal Abidin, M. (2001) ? Forecasting corporate failure in Malaysian industrial sector firms? Asian Academy of Management Journal, 6(1), pp. 15-30.