The effect of financial leverage on real and accrual-based earnings management

Abstract

Past research has documented a substitution effect between real earnings management (RM) and accrual-based earnings management (AM), depending on relative costs. This study contributes to this research by examining whether levels of (and changes in) financial leverage have an impact on this empirically documented trade-off. We hypothesise that in the presence of high leverage, firms that engage in earnings manipulation tactics will exhibit a preference for RM due to a lower possibility – and subsequent costs – of getting caught. We show that leverage levels and increases positively and significantly affect upward RM, with no significant effect on income-increasing AM, while our findings point towards a complementarity effect between unexpected levels of RM and AM for firms with very high leverage levels and changes. This is interpreted as an indication that high leverage could attract heavy outsider scrutiny, making it necessary for firms to use both forms of earnings management in order to achieve earnings targets. Furthermore, we document that equity investors exhibit a significantly stronger penalising reaction to AM vs. RM, indicating that leverage-induced RM is not as easily detectable by market participants as debt-induced AM, despite the fact that the former could imply deviation from optimal business practices.

Keywords:

• leverage
• debt
• real earnings management
• accrual-based earnings management
• market performance

JEL Classification:

• M40
• M41

Acknowledgements

We are grateful to the Editor, Professor Edward Lee, and to an anonymous reviewer for helpful comments and suggestions, which greatly improved the paper. We would also like to thank seminar participants at the Essex Business School and the Queen Mary University of London, the participants of the American Accounting Association (AAA) 2012 Annual Meeting, held in Washington, DC, USA, and the 36th Annual Congress of the European Accounting Association (EAA) 2013, held in Paris, France, for helpful comments and suggestions.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1. We are grateful to an anonymous reviewer for suggesting this argument regarding agency costs in debt providers (e.g. banks).

2. At this point, it should be mentioned that past research has also associated high leverage and specifically debt covenant violation with downward earnings management, in an effort on the part of firms to achieve better concessions during debt renegotiations (Becker et al. 1998, Mohd Saleh and Ahmed 2005, Rodríguez-Pérez and van Hemmen 2010). Reference is made to this research in order to provide a complete account of different motivations for EM. However, EM in the course of debt renegotiations is considered to represent a more specialised event context.

3. As our models for the examination of AM (and RM that follows) are estimated according to year and industry, we require a minimum of 8 observations for every 2-digit SIC code in every year (Cohen et al. 2008, Cohen and Zarowin 2010).

4. In the sections that follow, the actual (usable) number of suspect firm-year observations decreases, as we further require that (a) at least one of our and metrics (outlined in Sections 3.1–3.2) can be computed for at least one year for each suspect firm, and (b) data for a number of control variables are available for our suspect firm-years. Section 5 provides a complete description of the sample selection procedure and detailed descriptive statistics.

5. As Vella (1998) points out, the omitted variables bias is probably the most commonly encountered problem in social and behavioural sciences, with self-selection one of the main common sources of this bias (see Wooldridge 2002).

6. For detailed analysis and justification for the inclusion of these variables in the first step of the Heckman (1979) approach, see Zang (2012), Chan et al. (2015) and the references therein.

7. Let denote the estimates from the probit selection Equation (1). Then the inverse Mill's ratio is estimated as for each suspect firm and is included in the second step equation of the Heckman (1979) procedure. is the density of the standard normal distribution. Our sample period begins in 1990 rather than in 1987 as in Zang (2012) and Roychowdhury (2006) because of data availability issues. In relation to how the numbers of observations for suspect firms reported by our study compare to relevant numbers reported by previous studies making use of approximately the same data set, our numbers are in line and closer to the ones reported by Roychowdhury (2006) rather than Zang (2012), when taking into account the sample period used by the respective studies. At the same time, all variables included in our analyses performed have been truncated at 1–99% on a variable-by-variable and year-by-year, while other studies make use of winsorising at 1–99% (Zang 2012, p. 689). Regarding numbers of observations reported for suspect firms by past research, in more detail, altogether, Roychowdhury (2006) identifies 503 suspect firm-years during 1987–2001 by defining suspect firm years where they have Net Income/Total Assets ≥0 but <0.005, while Zang (2012), according to her first suspect-firm definition, defines suspects as firms just beating/meeting the zero benchmark, or firm-years with earnings before extraordinary items over lagged total assets between 0% and 0.5%, and finds 3428 such firm-year observations (see Zang 2012, p. 690, Table 2, Panel A, and also Table description) for a sample period extending from 1987 to 2008. We make use of three (out of Zang's four) definitions of ‘Suspect’ firm-years. Finally, our study follows a different approach when defining the ‘Suspect’ firm dependent variable in the probit model of Equation (1) in the first step of the Heckman procedure: Equation (1) is estimated on the union of two samples: the sample of suspect firm-years and a sample of non-suspect firm-years that is defined as in Zang (2012), while Zang (2012) reports that ‘The dependent variable in the selection model is Suspect_t, which equals 1 if a firm just beats/meets one of the earnings benchmarks discussed above, and 0 otherwise’ (Zang 2012, p. 684).

8. Our maintained hypothesis that the decisions of suspect firms to engage in AM and RM are sequential in nature is confirmed by the results of a Hausman test (Hausman 1978) we conduct for the endogeneity of and . To conserve space, we make the details and results of the test available upon request.

9. We follow the approach by Graham et al. (2008) for calculation and do not include the ratio of market value of equity to book value of total debt, since a similar term, market-to-book, enters the system in Equation (2) as a separate variable. is calculated following Graham et al. (2008) as: (1.2 * working capital + 1.4 * retained earnings + 3.3 * EBIT + 0.999 * sales)/total assets, or (1.2 * #179 + 1.4 * #36 + 3.3 * #170 + 0.999 * #12)/#6. However, when Altman’s (1968) Z-score is employed out of the regression estimation context (e.g. with respect to definitions of distressed firms using cut-off points in subsequent analyses), this term is included for the estimation of the Score, and in these cases the Score is denoted as ‘Altman's ’.

10. We would like to thank Professor John Graham (Graham 1996a, b, Graham and Mills 2008, http://faculty.fuqua.duke.edu/~jgraham/) for providing us with the data on MTR.

11. is actually truncated at 1% and 98% to eliminate a few observations (about 20) with a leverage ratio in excess of 10.

12. For a detailed economic rationale of the correlations between AM and RM metrics, see Cohen and Zarowin (2010, p. 9).

13. Findings from Table 3 (both panels) remain qualitatively similar when means instead of medians are used, although results (not reported here) are more sensitive to the existence of extreme values. Neither does experimenting with different numbers (3 or 10, etc.) of leverage/leverage change portfolios seem to affect the nature of the results in Table 3.

14. We would like to thank an anonymous reviewer for pinpointing this subtle point and making a detailed and specific suggestion for the type of analysis to follow in order to address this concern.

15. Untabulated results for the analysis reported on Table 4 for changes in leverage are weaker when contrasting AM and RM patterns for distressed vs. healthier firms, indicating that the relative financial health of firms in relation to changes experienced in leverage does not have such a strong impact on the choice of suspect firms to engage in RM vs. AM. Results from Table 4 are qualitatively similar when defining healthy firms as the ones with an Altman's Z-score value over 2.99.

16. Multicollinearity is considered to be low when the VIFs are below a rule-of-thumb threshold value of 10 (see e.g. Chatterjee and Hadi 2012).

17. The first-stage Heckman equation is the same as in Table 5.

18. As Barber and Lyon (1997, p. 370) point out, matching sample firms to control firms on specified firm characteristics such as size and book-to-market alleviates known biases in abnormal return calculation (such as the new listing bias, the rebalancing bias and the skewness bias), especially for longer time horizons. We would like to thank an anonymous reviewer for underlining this last point.

19. Untabulated results are qualitatively the same with the use of means, for both Tables 8 and 9. (Untabulated) results when using a six-month window are qualitatively similar (although weaker) in comparison to results for the three-month window, for both Tables 8 and 9.