Can corporate governance save distressed firms from bankruptcy? An empirical analysis

See Federal Register, Vol. 68, No. 218, Wednesday, Nov. 12, 2003, Release No. 34-48745, P. 64154.

For example, Richardson and Waegelein (2002) argue that certain compensation structures may trigger aggressive earnings management.

See, for example, the Z score of Altman (originated in 1968 and updated in 1993) and the credit scoring method of Ohlson (1980). In addition, see Turetsky and McEwen (2001) for a study on the predictors of distress.

For example, Altman’s Z score is the following mapping: Z = 1.2 (working capital/total assets) + 1.4 (retained earnings/total assets) + 3.3 (earnings before interest and taxes/total assets) + 0.6 (market value of equity/book value of total liabilities) + (sales/total assets).

Other examples of instances in which financial/accounting data was misleading can be found in both the academic and business press. In the academic literature, Teoh et al. (1998a) indicate that some seasoned equity issuers may raise reported earnings by altering discretionary accounting accruals; Teoh et al. (1998b) report similar results for initial public offerings. Also, DeAngelo (1988) documents that following successful proxy fights managers adjust discretionary accounts in order to improve apparent performance. The business press reports that Waste Management, Cendant, Xerox, and Enron have all lied in their financial statements in order to mislead investors about their true financial position, while General Electric, the world’s largest firm by market capitalization, is suspected of manipulating its accounts in order to disguise risk. (See The Economist, May 4th, 2002).

During her February 20, 2004 speech, SEC Commissioner Cynthia A. Glassman indicated that the SEC notes that director independence is a common missing element in those firms embroiled in recent financial scandals. See http://​www.​sec.​gov/​news/​speech/​spch022004cag.​htm

Using a medical analogy, the current methods of estimating bankruptcy risk are akin to methods that predict whether or not a sick patient will die on the basis of current measurements of vital statistics, without any regard for the quality (or the incentives) of the doctor treating the patient.

Hazard models are widely employed in the physical and social sciences to estimate the conditional probability that an individual entity (e.g., a person, a firm, a molecule) will transition from one state of being to another as a function of the amount of time in the initial state and other conditioning variables (referred to as covariates). For excellent technical treatments of hazard models and estimation techniques see Cox (1972) and Kalbfleisch and Prentice (1980).

Static Probit techniques essentially force the researcher to collapse any differences in the time-series paths across firms into a set of static cross-sectional variables. While, in principle, this could be done hazard methods naturally allow for time-varying conditioning information.

From the corporations’ SEC proxy statements, 14-DEF, 14, 14D, 10-K, and 8-K filings we obtain data on compensation, board size, board composition, and firm ownership.

The interest coverage ratio (ICR) is defined as operating income/interest expense. See footnote 3 for the definition of the Z score.

Recent studies using these measures are Dichev (1998) (Z score) and Kahl (2003) (ICR).

Left-censoring refers to the situation in which a starting value is only known to be within a range. In this case, if a firm starts off the sample with a Z score or ICR below the critical values, we would only know that τ₀ < 1991. As such, given a bankruptcy date of say 1995, we would only know that T ≥ 1995–1991 = 4. Thus T is censored. However, this time it is not because we don’t know when it transitioned from distressed to bankrupt (which happens on the right-hand-side of a time line), but because we don’t know when it transitioned from fully solvent to distressed (which happens on the left-hand-side of the time line).

Kahl (2003) uses a similar selection criterion for the ICR.

As Lundstrum (2003) indicates, it is possible that the firm’s Z-score or its ICR are not helpful in diagnosing the firm’s finacial condition if the company suffers from information problems.

Smith and Watts (1992), Yermack (1996), and Fich and Shivdasani (2007) utilize a similar variable to proxy for growth opportunities. In addition, in robustness tests we use alternative proxies for growth opportunities such as the market-to-book ratio, and the capital expenditures to sales ratio. The results of these tests generate inferences similar to those tabulated.

Since we are concerned about the cases where bankruptcy is more likely to occur, we focus on the high Z-score sample. Nonetheless, summary statistics for the low Z-score sample are available from the authors upon request.

This estimate is obtained by estimating the natural log of the firm’s sales (Compustat item 12). Using the natural log of firm assets (Compustat item 6) as a size proxy yields similar hazard estimation results.

Low insider ownership has been reported by other authors in similar samples; see, for example, Shivdasani and Yermack (1999), who analyze a sample of unconditionally healthy Fortune 500 firms during 1994–1996.

Very similar correlations exist between the variables for the low-threshold Z-score sample and the ICR sample.

The low-threshold Z-score sample produces results that are very similar to those obtained with the high-threshold Z-score sample.

In the tables that follow we only report results for the high-threshold Z-score sample and the ICR sample. The two samples differ the most in terms of the firms that are included. As such, they should provide a good check regarding the errors-in-distress-measurement issue. However, we also estimate all the hazard models using the low-threshold Z-score sample and all of the results are very similar to those reported for the high-threshold Z-score sample.

Because the raw coefficient estimates of hazard models have no direct economic interpretation, throughout the paper and in the tables we provide and refer to risk ratios, which have an intuitive interpretation. A covariate’s risk ratio minus 1 quantifies the percentage change in the probability of bankruptcy given a one unit increase in the value of the covariate. We also provide and refer to the model’s generalized R². This measure is analogous to the R-squared in a standard OLS regression; it measures the percentage of the variation in the occurrence of bankruptcy explained by the variation in the covariates used in the model.

Of course, one could argue that the size of a firm is dictated by the governance structure of the firm. As Table 5 shows, however, firm size is not significant when no other governance-related variables are included.

For example, Cotter et al. (1997) report that outside directors of targeted firms enhance their stockholders’ gains during tender offers.

Vancil (1987) and Fich (2005) provide discussions on the market for directorships and the cost to an outside director from a loss in reputation as an independent monitor.

Gilson (1990) supports this view by documenting that directors of failed firms are less likely to obtain similar positions at other firms.

The idea that CEOs and other top managers would make self-serving decisions that might be detrimental to the firm’s shareholders is not new. In the context of merger proposals, Yen (1987) provides a clear rationale to explain why incumbent top managers, concerned about their own job security, might engage in opportunistic behavior.

Other commonly used proxies for firm size, such as the natural log of total capital or the natural log of assets, yield similar results.

Other proxies for growth opportunities such as the market-to-book ratio generate qualitatively similar results.

Shumway (2001) indicates that the firm’s return as a time-varying covariate significantly improves the fit of the hazard model. The focus in Shumway (2001) is on predictability and, as a result, there is a greater concern about the amount of cross-sectional variation in the occurrence of bankruptcy that can be explained by the model. Our approach differs in that we are primarily interested in isolating the impact of governance characteristics and obtaining precise estimates of their effect on the hazard. Thus, if the time-varying return is correlated with the decisions generated by the governance structure, resulting in the governance effects being masked, these returns should not be included as time-varying covariates.

A possible exception exists in the law literature. See LoPucki and Whitford (1993).