The endogeneity bias in the relation between cost-of-debt capital and corporate disclosure policy

Abstract

The purpose of this paper is twofold. First, we provide a discussion of the problems associated with endogeneity in empirical accounting research. We emphasize problems arising when endogeneity is caused by (1) unobservable firm-specific factors and (2) omitted variables, and discuss the merits and drawbacks of using panel data techniques to address these causes. Second, we investigate the magnitude of endogeneity bias in Ordinary Least Squares (OLS) regressions of cost-of-debt capital on firm disclosure policy. We document how including a set of variables which theory suggests to be related with both cost-of-debt capital and disclosure and using fixed effects estimation in a panel data-set reduces the endogeneity bias and produces consistent results. This analysis reveals that the effect of disclosure policy on cost-of-debt capital is 200% higher than what is found in OLS estimation. Finally, we provide direct evidence that disclosure is impacted by unobservable firm-specific factors that are also correlated with cost of capital.

Acknowledgements

We thank Christine Botosan and Marlene Plumlee for providing us with the AIMR disclosure scores. Parts of this study were written while Valeri Nikolaev visited the Sloan School at MIT. We gratefully acknowledge helpful comments from Jan Bouwens, Peter Easton, Stephan Hollander, David Larcker, Christian Leuz, Matt Pinnuck, Maarten Pronk, Konstantin Rozanov, Tjomme Rusticus, Jeroen Suijs, Paula van Veen, Hylke Vandenbussche, Heidi Vander Bauwhede, Sofie Van der Meulen, Marleen Willekens, Peter Wysocki, and from workshop participants at the universities of Amsterdam (VU), Leuven, Melbourne and Tilburg. We also appreciate the constructive feedback from the Editor and two anonymous reviewers of European Accounting Review.

Notes

¹ Other potential explanations for these conflicting results are the current high standards of mandatory disclosure (rendering voluntary disclosure choices of second-order importance) and measurement problems in the somewhat elusive key constructs of ‘information problems’ and ‘disclosure quality’ (Leuz and Verrecchia, 2000; Healy and Palepu, 2001; Zhang, 2001).

² This definition is consistent with the econometrics literature (Greene, 2000; Wooldridge, 2002) and with the proposal in Chenhall and Moers (2004).

³ Often these costs of disclosure are defined to include the costs of collecting, processing, reporting and verifying information and the cost due to loss of competitiveness (see, e.g. Wagenhofer, 1990; Guo et al., 2004). Potentially interesting definitions also refer to the costs associated with uncertainty about investor reactions to a certain disclosure (Verrecchia, 2001; Fishman and Hagerty, 2003) or litigation costs (Skinner, 1997).

⁴ Within standard asset pricing models, such as the CAPM, only undiversifiable risk is priced on the market, and therefore we have to assume that the proposed joint determinants of ‘cost-of-debt capital’, such as the firm's default risk, are at least partly correlated across firms. Indeed, an often-heard critique on studies that relate disclosure to cost of capital is that differences in disclosure quality are idiosyncratic and therefore should not ‘survive the forces of diversification’ (Leuz and Verrecchia, 2005, p. 1) nor impact on the cost of capital. Leuz and Verrecchia (2005), in contrast, argue that disclosure improves the coordination between the firm and its investors with respect to capital investment decisions. As such, poor disclosure quality can lead to misaligned investments and higher cost of capital. Other studies have suggested that disclosure may impact on cost of capital, even if it is idiosyncratic, because it improves market liquidity (Leuz and Verrecchia, 2000; Verrecchia, 2001), reduces estimation risk (Barry and Brown, 1985) or increases the investor base (Merton, 1987).

⁵ Sengupta's model provides a convenient vehicle to illustrate the effect of endogeneity bias in disclosure research. It is also to some extent an arbitrary choice since endogeneity bias is present in many contexts in (financial) accounting research and many potential candidates exist for similar analysis as is conducted in this paper. Ittner and Larcker (2001), Chenhall and Moers (2004), and Larcker and Rusticus (2005) provide helpful discussions of endogeneity in accounting research.

⁶ We recognize that causal statements cannot be made based on statistical considerations, but only on theory. When we refer to a causal relation, we use this as shorthand for ‘a causal relation as suggested by theory and underpinned by empirical evidence’.

⁷ One test is that the choices made should be palatable to the researcher's peers.

⁸ While the disturbance term then includes variables that are unobservable to the researcher, these factors may very well be observable to the economic agent under study. Indeed, endogeneity arises when the explanatory variables represent decisions made by the agent on the basis of such factors (Hayashi, 2000).

⁹ Self-selection bias will also arise when the sample is truncated or censored, or sampling is on the dependent variable. When sampling is on one of the exogenous variables, the sample will not be random but estimation of the structural model is unaffected (Wooldridge, 2002). See also Shehata (1991) for a discussion of selection bias issues in an accounting context.

¹⁰ This discussion is geared towards one panel data technique in particular: fixed effect estimation.

¹¹ If fixed effects and IV estimation do not agree, the implication is that the model is misspecified (e.g. the instruments are invalid or endogeneity is not alleviated by fixed effects estimation. A Hausman-type test may be used to discriminate between the estimators.

¹² It is not immediate which estimator will be more efficient asymptotically. This will depend on the number and quality of instruments and the amount of within-variation.

¹³ It is often not immediate whether including more than one instrumental variable is beneficial in finite sample settings. See, e.g. Kennedy (2003) for a discussion. A Sargan (1958)–Hansen (1982) test is available to evaluate whether extra instruments should be used.

¹⁴ The tradeoff between single equation and system methods is that the latter are more susceptible to misspecification since they require the correct specification of all equations in the system. As an equivalent alternative one may estimate the reduced form of the structural model and then solve for the structural parameters in terms of reduced form parameters.

¹⁵ We choose a research design that allows us to investigate endogeneity caused by omitted variables in relative isolation from endogeneity caused by simultaneity. We provide more details on this in Section 4. In short, we rely on the pre-determinedness of most of our right-hand side variables to argue that simultaneity is less likely to be severe. Nevertheless, we cannot exclude the possibility that simultaneity bias is present and our results should be interpreted with this caution in mind. One possible explanation why these earlier studies have not found that OLS is inconsistent might be that the instrument variables that were used in prior work were weak (see also, Larcker and Rusticus, 2005).

¹⁶ Recent studies have pointed explicitly to the failure of many disclosure studies to take between-firm differences in costs of disclosure into account (Fields et al., 2001; Cohen, 2003).

¹⁷ We would like to stress that these are indeed examples and many other reasonable theories exist. Agency costs are a clear alternative illustration. These costs are unobservable but likely differ among firms. Agency costs are likely to affect both the disclosure decision and the cost of capital. Yet another alternative is firm (as opposed to management) reputation. We do not aim at providing an exhaustive list of firm heterogeneity.

¹⁸ See Hirshleifer and Teoh (2003) for a model in which pro forma disclosures are used to misdirect the attention of investors with limited cognitive abilities. To the extent that cognitive abilities among investors vary we expect different optimal levels of disclosure.

¹⁹ Lang and Lundholm (2000) on the other hand provide evidence that increasing disclosure prior to a seasoned equity offering may be interpreted as ‘hyping’ the stock and firm's experience continued negative returns subsequent to the offering announcement. This effect is probably difficult to witness in our sample since we do not have a continuous measure of disclosure policy, but instead rely on annual assessments of disclosure. See also, Mak (1996) and Jog and McConomy (2003).

²⁰ These ratings have been frequently used in earlier disclosure studies and are discussed in some detail elsewhere (Lang and Lundholm, 1996; Core, 2001; Healy and Palepu, 2001).

²¹ Sengupta (1998) includes two variables as control variables in his regression that would otherwise have been included in this category. These variables (current income and interest coverage) are therefore part of the specification of our equation (1) as ROS and COVER, respectively.

²² Sengupta (1998) includes the logarithm of total assets as a control variable in his regression. This variable (LASSETS) was therefore included as control in our equation (1). Otherwise, it would have been included in the category of structure variables to proxy for the economies of scale in producing information.

²³ The inclusion of MOODRNK as a determinant of cost-of-debt capital is contentious. While some prior studies have added credit ratings as a control variable (Bagnani et al., 1994; Campbell and Taksler, 2003; Mansi et al., 2003), others have not. Sengupta (1998) argues that credit rating agencies consider the quality of disclosure when deciding on a firm's credit rating. Including the rating alongside a measure of disclosure may therefore create multicollinearity problems and it might become difficult to separate out the effects of disclosure and of credit ratings. We decided to include MOODRNK not only because it is an established proxy of information asymmetry, but also because we believe it is important to try to establish if the market reacts to disclosure directly or to credit ratings which (indirectly) reflect disclosure quality. We have also conducted the empirical analyses without MOODRNK and we report these results in note 32. If MOODRNK is construed as a proxy for information asymmetry then a more appropriate measurement is before the firm discloses its information. Since MOODRNK is an issue-specific rating, it is not straightforward to implement this in the regressions. We check the robustness of our results to the timing of the measurement of information asymmetry by replacing MOODRNK by S&P long-term debt rating (Compustat item 280), which is available for all firm-years in the sample. We use a lagged (t − 1) value of this rating to ensure that it is measured before the disclosure at t. We report the results for this specification in note 32 as well.

²⁴ In principle, equation (3) could be estimated using fixed and random effects, respectively. The appropriateness of each estimator depends on assumptions about the correlation between α _i and the included independent variables. If the firm-specific characteristics captured in α _i are independent of the regressors, random effects estimation is consistent and efficient. However, if the firm-specific characteristics are correlated with any of the regressors this estimation procedure is inconsistent and fixed effects are preferred. Since we have strong theoretical reasons to believe that firm-specific characteristics are correlated with the disclosure variable, our priors are that fixed effects estimation is the most appropriate when estimating equation (3). In fact, unreported results of a Hausman test of the consistency of random and fixed effects estimation support the choice for fixed effects. This is further evidence that firm heterogeneity is important in the current setting and should be taken into account (using fixed effects) when estimating the relation between disclosure and cost-of-debt capital.

²⁵ Seminal studies include Mundlak (1961, 1978), Hoch (1962), Ben-Porath (1973), Griliches (1977), Ashenfelter (1978), Chamberlain (1978), Hausman (1978) and Hausman and Taylor (1981). More recent applications in finance include Ashenfelter and Kruger (1994), Himmelberg et al. (1999), Campbell and Taksler (2003) and Doidge (2004). In accounting, Francis et al. (2004) and Hail and Leuz (2004) provide fixed effect results.

²⁶ Griliches and Hausman (1986), Himmelberg et al. (1999) and Zhou (2001) note that the fixed effect estimator may suffer from bias, which is associated with measurement error. Griliches and Hausman (1986) point out that measurement error will have a different impact on the fixed effects estimator and the first-differences estimator. Since we report fixed effects and first-differences results that are very close, it is unlikely that measurement error is a major issue here.

²⁷ Sengupta's (1998) sample consists of 103 observations (and as many firms, since he only retains one observation per firm). We have, due to our design, multiple observations for each firm, and consequently cannot claim that our observations are independent. To ascertain the extent of this problem we have compiled a sample in which each firm enters only once, and ran the benchmark model on this sample. Our results remained qualitatively unchanged and we conclude that any potential downward bias of the standard errors, due to dependent observations, is likely to be minor.

²⁸ Note that the magnitudes of our coefficients are not directly comparable to those in Sengupta (1998) because our variable definitions are sometimes different.

²⁹ Standard errors throughout the paper are White (1980) heteroskedasticity consistent.

³⁰ The simple correlations in Table 2 between each of the ‘joint determinants’ (and their best linear combination) and our disclosure variables are low and there is little reason to be concerned about multicollinearity being an issue in our subsequent analyses (see also Griffiths et al., 1993).

³¹ The (unreported) results for the other three disclosure measures are similar to those for PCTRNK.

³² We also estimated the model without MOODRNK. Unreported results show that in the augmented OLS regressions Disclosure remains significant, but the size of the coefficient is smaller than in a model without any control variables included. Replacing MOODRNK by the lagged value of S&P's long-term debt rating did not affect the main findings and our conclusions remained unchanged.

³³ Random effects estimates for PCTREL, PCTANL and PCTOPB are available from the authors upon request.

³⁴ It should be noted, however, that the incremental explanatory power of the Disclosure variable is small (and below 1%). This is not unexpected though, since our model already explains almost 90% of the variation in cost-of-debt capital. What is more, the incremental explanatory power of Disclosure is of similar magnitude as our leverage variable, which is always very significant. Therefore, we believe that adding Disclosure to the model is meaningful regardless of its low incremental explanatory power.

³⁵ We also used feasible generalized least squares to estimate the relation between unobservable firm-specific factors and disclosure and our results (not reported, but available on request) were qualitatively similar and did not change our conclusions.

³⁶ Indeed, signs and significance remain similar in all cases except for the regressions of PCTANL.

³⁷ Indeed, this is precisely why we use fixed effects estimation. The decision to commit to a disclosure policy is likely to be part of a portfolio of simultaneous firm choices on strategy, business profile, risk and environmental segments, compensation and customer/supplier related policies (Core, 2001). As such, the systematic component is likely to be endogenous and should be eliminated from the analysis.

³⁸ One alternative explanation for our findings could be that our Disclosure measure captures mostly random noise or performance-related variation in disclosure quality (either because good performance leads to better disclosure or because its leads to better perceived disclosure). Noise will attenuate the regression coefficient, but the performance part can induce a negative relation between disclosure and cost-of-debt capital. While the performance control variables should control for this, the net effect could still be a negative observed relation between disclosure and cost of capital.