Earnings Quality in Restating Firms: Empirical Evidence

Transcription

1 Earnings Quality in Restating Firms: Empirical Evidence Master Thesis Spring 2012 Marie Herly Stud. MSc. Finance and International Business Aarhus University Business and Social Sciences Department of Economics and Business Supervisors: Frank Thinggaard & Jan Bartholdy

2 Abstract The objective of this paper is to compare the earnings quality in US rms required to restate their nancial statements with the quality of a matched control group. More specically, I test whether the earnings quality diers in the ten years before the (last) restatement event for the two groups. To examine if the scrutiny of the Securities and Exchange Commission (SEC) improves the nancial reporting of restating rms, I also test whether the quality diers between the two groups after the last restatement. Using a wide portfolio of accounting quality metrics, I predict and nd that the earnings quality of restating rms is poorer than that of the control group, already in the years before the restatement. Using a dierence-in-dierence research design, I also nd that the restating rms improve the quality of their nancials statements, but surprisingly not signicantly more than the control group. It is therefore not possible to attribute the improvement to the restatement event alone.

3 Contents 1 Introduction Problem Statement Limitations Accounting Quality Denition Operationalising the Conceptual Framework The Conceptual Framework for Financial Reporting Operationalising the Conceptual Framework Determinants of Accounting Quality - Dependent Variable Consequences of Accounting Quality - Independent Variable Earnings Management Incentives to Manage Earnings Consequences of Earnings Management Proxies for Accounting Quality Abnormal Accruals Models The Jones Model The Modied Jones Model The Performance-matched Modied Jones Model Accruals Quality Models Dechow-Dichev Model Modied Dechow-Dichev Model Other Accruals Models Magnitude of Accruals Change in Total Accruals Avoiding Earnings Decreases & Small Losses Avoiding Small Earnings Decreases Loss Avoidance Asymmetric Timeliness Timeliness Timely Loss Recognition Smoothness Variability of Earnings Correlations between Accruals and Cash Flows Persistence Predictability Restatements Previous Literature Identifying Restatements SEC's AAERs

4 4.2.2 Other Sources Issues when Working with Restatements Hypotheses Development 42 6 Research Design and Descriptive Statistics Sample Selection Matching Procedure Accounting Quality Metrics Descriptive Statistics Empirical Results Before Restatement Event After Restatement Event Dierence-in-Dierence Test Robustness Tests 66 9 Conclusion 68 References 69 2

5 List of Figures 1 Research Design Research Design and Matching Procedure List of Tables 1 Denitions of Accounting Quality Overview of Earnings Quality Metrics Compare original/restated values for Restating Firms Calculation of Earnings Quality Metrics Restatements divided in Years Comparison of Restaters and All Compustat Firms Descriptive Statistics of Key Variables Restaters and Non-Restaters Before the Restatement Event Restaters and Non-Restaters After the Restatement Event Dierence-in-Dierence Robustness Tests

6 1 Introduction This paper examines earnings quality in rms before and after they restate their nancial statements. So far, research within the earnings quality of restating rms has mainly focused on measuring the quality in the restatement year only, and has made no inferences on the quality before and after. Thus, little is known on how restating rms behave before and after they are detected by the Securities and Exchange Commission (SEC). This gap is problematic for at least two reasons. First, if earnings quality is dierent for restating and non-restating rms even before the actual restatement, then SEC may begin the surveillance and scrutiny of these rms even earlier. Investors can also determine which rms that are likely to restate based on the earnings quality of these rms. Second, given the enormous eorts SEC puts into detecting restaters, it is certainly in the interest of SEC, regulators, and investors to examine if the rms actually improve afterwards. This can also shed light on the educative role of SEC. Previous research shows that rms that restate do have lower earnings quality in the restatement year. Some studies have found that restaters have poor corporate governance, high growth, and extreme values of specic accounting fundamentals in the years before the restatement event. It is also evident that the nancial markets react very negatively to restatements, and that management and auditor turnover increases after a restatement. However, to my knowledge no studies examine the earnings quality before and after a restatement event. An examination of this will deepen our understanding of restating rms and how they generically dier from non-restating ones. 1.1 Problem Statement The purpose of this thesis is twofold: First, a broad portfolio of earnings quality metrics will be thoroughly described, and their advantages and drawbacks will be discussed. Second, these earnings quality metrics are used to examine the properties of rms with restated nancial statements and a matched control group. Thus, the paper is a joint test of the earnings quality of restating rms before and after the restatement event, and whether the accounting quality proxies actually work. The thesis outlines the following research questions: Research Question 1. Do restating rms have poor accounting quality compared to non-restating rms prior to a restatement? Research Question 2. Do restating rms have poor accounting quality compared to non-restating rms after a restatement? 4

7 Earnings Quality Before Earnings Quality After Restatement Year Quality Quality Restaters Quality Quality Non-Restaters Non-Restatement Year Figure 1: Research Design Research Question 3. Does a restatement event change accounting quality for restaters and non-restaters, respectively? The restating rms are identied through the Securities and Exchange Commission (SEC), and the sample contains prominent cases of earnings management, such as Xerox, Bristol-Myers Squibb, and QWest. I use a matched sample research design in which I match each restating rm with a control rm with similar size and protability, and within a similar industry. I rst test the earnings quality for both the treatment and the control group both before and after a restatement event, and then test the relative change in earnings quality for the two groups with a dierence-in-dierence design. The research design is depicted graphically in Figure Limitations This study is not intended to examine the properties of restating rms per se, such as their characteristics, which rms are more likely to restate, and the reasons for the restatements. I also do not focus on the detection mechanism of the SEC, i.e. which rms they are likely to scrutinise. Earnings quality metrics are often split in accounting-based or marketbased (Francis et al., 2004). Whereas the rst uses only accounting information, with for example cash or earnings as the reference construct, the latter has stock returns or prices as the reference construct. As this paper depicts whether the poor earnings quality in the restatement year can be detected both before and after the actual year, the market-based metrics are not of interest. These metrics are concerned with the investors' perception of the 5

8 quality of nancial statements which is not the area of interest in this paper. In addition, the eect after the restatement event is already partly described in the event studies conducted on restatement events. The remainder of this paper is structured as follows: Section 2 is a thorough literature review on the determinants and consequences of accounting quality. The next section continues with a description of the most commonly used metrics for measuring accounting quality. Section 4 gives an overview of the literature within restatements, the sources available for identifying them, and potential pitfalls in research designs.in Section 5 my three hypotheses are then developed based on the previous sections, and Section 6 describes the research design and data. Empirical results are presented in Section 7, followed by robustness tests and conclusion, as well as suggestions for further research. 6

9 2 Accounting Quality The notions of accounting and earnings quality have been used both synonymously and as two dierent concepts. Melumad and Nissim (2008) dene accounting quality as a generalised view of earnings quality which evaluates the impact of accounting choices. This view is shared with Francis et al. (2006), who note that nancial reporting quality is a special case of information quality and dene earnings quality as a summary indicator of nancial reporting quality. The authors caution against focusing only on earnings when evaluating nancial reporting quality, since a lack of focus on for example balance sheet information will mask true dierences in accounting quality. In the following, accounting quality and earnings quality will be used interchangeably. 2.1 Denition Many denitions of accounting quality exist in the literature. As such, earnings quality is a function of both the ability of the accounting system to measure the rm's fundamental performance and how the accounting system is implemented, both of which are unobservable. The task of disentangling these is therefore challenging to say the least. DeFond (2010) in consequence suggests that the inability to directly observe the two constructs of interest is unlikely to be solved by the literature. Former chairman of SEC, Arthur Levitt, mentions comparability and transparency as two main attributes of high quality nancial reporting. He determines the key objective of nancial information to be providing disclosure to investors on where the company has been, where it is and where it is going (Levitt, 1998a). Barth and Schipper (2008) and Bhattacharya et al. (2003) also propose transparency as a desired attribute of high quality earnings. The European Union attaches true and fair view as the main attribute of high quality nancial reports (EU, 1978). The idea behind this view is that nancial reports that truthfully reect the underlying reality in the rm are of high quality. This corresponds well to the idea of decision usefulness, since faithful representation of nancial reporting by nature must be more useful to users. The notion of decision usefulness as an indicator of high earnings quality is widely accepted and has been used by a number of researchers, e.g., Abdelghany (2005), Ball and Shivakumar (2005), and Dechow et al. (2009). The latter dene earnings quality broadly as decision usefulness, but they also stress the notion of faithful representation for accounting quality. Dechow and Schrand (2004) dene earnings quality broader than decision usefulness. They view the denition of high quality earnings as threefold: 7

10 rst, the reported earnings number should reect current performance, second, it should be a good indicator of future operating performance and, - nally it should accurately annuitise the intrinsic value of the company. Thus, the authors extend the concept of earnings quality to include reliable valuation, which highlights analysts as very important users. The perspective of Francis et al. (2006) complements that of Dechow and Schrand, and associates high quality earnings with precise information. They also emphasise the role of analysts and dene high quality information as giving a more precise valuation. According to Melumad and Nissim (2008), FASB appears to emphasise the valuation role of accounting information over the information needs of users, as the information required by FASB is precisely what is needed to calculate the value of an enterprise. In continuation of this, Penman and Zhang (2002) dene high quality earnings as being sustainable indicators of future earnings, and Comiskey and Mulford (2000) classify high quality nancial reporting as the ability to generate sustainable earnings. In connection with this, persistence has also been named a key attribute of high quality earnings by Abdelghany (2005) and Dechow and Dichev (2002). A branch of research sees precision as the main attribute of high quality earnings, implying that earnings should accurately reect the underlying reality of the rm. These include for example Francis et al. (2006). This denition corresponds well to the qualitative characteristic faithful representation dened by IASB (IASB CF Ÿ33). Visvanathan (2006) uses the notion of closeness-to-cash as a desirable property of earnings. Thus earnings that are closer to cash ows, i.e. earnings that contain relatively small amounts of accruals, are of higher quality. Hence, this view is closely related to for example Dechow et al. (2009) and Francis et al. (2006). Conservatism, in the meaning of prudence, has also been put forward as a characteristic of accounting quality (Basu, 1997). This implies that caution is exercised when estimation assets and income, and liabilities and expenses, such that the former are not overstated, and the latter are not understated. Whether unconditional conservatism increases or decreases decision usefulness is an unresolved issue 1 Barth et al. (2008) dene high quality earnings as those that exhibit less earnings management, implying that quality is not an innate characteristic, but rather the absence of manipulation and bias. This corresponds well to the discussion of Guay et al. (1996) who argue that managerial opportunism reduces information precision and accounting quality. A brief review of some of the denitions is shown in Table 1. 1 Distinction between conditional conservatism (more timely recognition of bad news than of good news in earnings) and unconditional conservatism (policy that results in lower book values of assets/higher book values of liabilities in the early periods of asset/liability life time). 8

11 Quality Construct Table 1: Denitions of Accounting Quality Selected References Decision usefulness Ball and Shivakumar (2005), Schipper and Vincent (2003) Closeness-to-cash Visvanathan (2006) Comparability IASB (2006) Faithful representation Francis et al. (2006), EU (1978), IASB (2006) Persistence Schipper and Vincent (2003), Dechow and Dichev (2002), Comiskey and Mulford (2000) Precision Francis et al. (2006) Prudence Basu (1997), Beekes et al. (2004), Watts (2003) Relevance IASB (2006) Transparency Levitt (1998a), Bhattacharya et al. (2003) Understandability IASB (2006) Valuation input Dechow and Schrand (2004), Melumad and Nissim (2008) As Melumad and Nissim (2008) note, some of the attributes of earnings quality have contradictory implications. As an example, earnings that are close to the underlying cash ows are not necessarily predictable or accurately reecting future performance. Schipper and Vincent (2003) describe a possible contradiction between the persistence and predictive ability of earnings: Highly persistent earnings will have low predictive ability if the variance of a typical to the series is large. Consequently, earnings that are of high quality on the persistence dimension may be of low quality on the predictive ability dimension. It also seems clear that earnings quality as a construct is contextspecic, partly because the users, to whom the denition is targeted, dier from situation to situation (Dechow et al., 2009). In a related vein, Schipper and Vincent (2003) argue that earnings quality diers according to the users of nancial statements; thus, standard setters and managers with compensation contracts tied to the earnings number may have dierent perceptions of accounting quality. 2.2 Operationalising the Conceptual Framework The Conceptual Framework for Financial Reporting A widely used denition of accounting quality is derived from the qualitative characteristics of nancial disclosure from IASB (IASC, 1989), FASB (FASB, 1980), and the convergence project between the two (Board, 2008). These characteristics can be seen as quality attributes, and thus high quality nancial reporting will comprise the characteristics identied by standard setters. 9

12 Common to these is the emphasis on usefulness to users, following the intuition that nancial reporting should eectively help users make informed decisions. As a result, nancial statements, which do aid in decision making, are of high quality. In IASB's and FASB's joined Conceptual Framework for Financial Reporting (Board, 2008), the core aim of nancial statements is decision usefulness. The two primary qualitative characteristics supporting this aim are relevance and representational faithfulness. Four enhancing characteristics distinguish more-useful information from lessuseful information. Comparability ensures consistency between units and time-periods (Board, 2008, QC16), whereas veriability assures users that information indeed faithfully represents the underlying reality in the rm (Board, 2008, QC20). Timeliness means having the information available at the right time (Board, 2008, QC22), and, nally, understandability presents information clearly and concisely (Board, 2008, QC24). According to FASB (1980), the characteristics are conventionally referred to as qualitative, but some are more quantitative. For example, understandability might be dicult to quantify, whereas timeliness and relevance have been measured empirically Operationalising the Conceptual Framework The characteristics in the Conceptual Framework can all be seen as quality attributes, and thus rms that full the purpose of decision usefulness must have high quality nancial reporting. As put by (Barth et al., 2001, p. 89), "...because the CF sets forth the FASB's objective criteria for evaluating accounting amounts, researchers need only to operationalise the criteria, and not determine them." This reasoning has resulted in a number of studies which have tried to operationalise the qualitative characteristics from the Conceptual Framework. Krisement (1997) attempts to operationalise the concept of comparability, and Velury and Jenkins (2006) use proxies for representational faithfulness, timeliness, and neutrality to measure earnings quality. Wagenhofer and Ducker (2007) propose proxies for conservatism, predictability, timeliness and relevance, and Schipper and Vincent (2003) quantify predictability. Holthausen and Watts (2001) note that a number of papers have used value relevance as a proxy for relevance and reliability. Finally, Barth et al. (2001) measure relevance and reliability combined. Operationalising the conceptual framework is easier said than done, though. The empirical measures of earnings quality evaluate only parts of the overall earnings quality. Hence, it is dicult to attach a single measure to a single qualitative characteristic. According to Francis et al. (2006), accounting researchers have not been able to devise a single measure that captures the qualitative characteristics of the FASB Conceptual Framework. It is also em- 10

13 pirically problematic to assess the characteristics separately, because there is a degree of trade-o between them. Schipper and Vincent (2003) provide the examples of relevance and reliability. These are not mutually exclusive, nor necessarily comparable. In addition, some of the attributes from the conceptual frameworks are dicult, if not impossible, to measure objectively. How does one, for example, measure understandability, neutrality and comparability? Barth et al. (2001) argue that another issue on operationalising the conceptual framework is the fact that neither relevance nor reliability are dichotomous attributes, and thus the conceptual frameworks do not specify how much relevance or reliability is sucient to meet FASB's criteria. Another obstacle against operationalising the Conceptual Framework is its subjective nature. One example of this is that the recognition and measurement criteria depend on probabilities and judgement. As evident, the operationalisation of the framework is not an easy task. FASB and IASB have in fact addressed this issue. And it seems that the boards have little belief that operationalisation is actually doable:.. the Boards are unaware of useful means of quantifying either the overall quality of faithful representation or its components and concluded that they should not attempt to develop such means in the proposed framework. In reaching that conclusion, the Boards noted that an inability to quantify characteristics identied as qualitative is not surprising (Board, 2008, BC2.24). 2.3 Determinants of Accounting Quality - Dependent Variable A very important question in relation to earnings quality is naturally which factors inuence the quality of nancial statements. Francis et al. (2006) separate innate and discretionary sources of earnings quality. The former arise from the business model and operating environment, whereas the latter arise from the nancial reporting process, such as managerial judgement, governance activities, and the reporting standards. It is widely accepted that the quality of the standards and the diligence of regulators are important determinants of earnings quality. Soderstrom and Sun (2007) mention the accounting standards, the tax system, and the legal and political systems as drivers of earnings quality, and Barth et al. (2008) suggest that regulators inuence earnings quality. Beuselinck et al. (2009) also nd positive impacts on earnings informativeness following IFRS adoption, supporting the view that high quality standards have positive eects on earnings quality. Ewert and Wagenhofer (2005) show that tighter accounting standards increase earnings quality but can do little to restrict real earnings management. Levitt (1998a) argue that strong accounting framework and outside auditing will improve earnings quality. Capital market forces also have an eect on earnings quality. Soderstrom and Sun (2007) argue that nancial market development inuences earnings 11

14 quality, and Ball and Shivakumar (2005) show that UK public companies have higher accounting quality than private ones, due to the market demand for information. Burgstahler et al. (2006) examine how capital market pressures and institutional factors inuence rms' incentives to report accurate earnings. Some factors of the rm itself also aect the quality of its nancial statements. These are for example the capital and ownership structure (Soderstrom and Sun, 2007), the accounting methods chosen by the rm (Altamuro et al., 2005), and rm performance (DeFond and Park, 1997). The internal control regulation and corporate governance mechanisms of the rm also affect the accounting quality (Altamuro and Beatty, 2010). Dechow and Schrand (2004) argue that the nature of the rm as such can also inuence the earnings quality. The authors suggest that high growth companies, companies with intangible assets or complex transactions, and companies in volatile business environments can provide earnings numbers that do not accurately reect rm performance or indicate future cash ows. In these cases, neither earnings management nor poor monitoring is to blame for the low earnings quality. Finally, managerial intent and purely discretionary decisions also inuences the quality of earnings. The case of earnings management is elaborated in Section Consequences of Accounting Quality - Independent Variable A large part of the literature on consequences of earnings quality addresses the market outcomes of rms' nancial statements, such as the cost of capital, the stock return, and the bid-ask spread. Another branch of literature addresses other non-market outcomes, e.g., audit opinions and analysts' forecasts. The cost of capital is a widely used proxy of market outcomes. In general, empirical evidence suggests a negative relation between earnings quality and the cost of equity capital. Francis et al. (2006) view the cost of capital as a summary indicator of investors' resource allocation decisions, which is related to earnings quality since high quality nancial reporting should assist users in their resource allocation decision. Easley and O'Hara (2004) show that accounting information of high quality reduces cost of capital by reducing information risk. In their much cited paper, the authors extend the capital asset pricing model to include information asymmetry, and show that information risk is non-diversiable and thus a priced risk factor. Chen et al. (2007) and Lambert et al. (2007) nd similar results. Francis et al. (2004) show that rms with the least favourable values of seven attributes of earnings quality generally experience larger cost of equity. Extending this research to the cost of debt as well, Francis et al. (2005) show that rms 12

15 with poor accruals quality have larger cost of debt and equity capital. They assign this to the fact that higher magnitudes of accruals indicate higher information risk, which demands higher cost of capital. There is, however, not completely consensus on this issue. Thus, Cohen (2008) suggests that rms providing nancial information of higher quality do not necessarily enjoy a lower cost of equity. He attributes this to classical asset pricing theory which shows that diversiable risk is not priced and argues that information is not diversiable. Another branch of research studies how earnings quality inuences stock prices. Dechow and Schrand (2004) argue that they would expect no response to low quality earnings if investors are rational in their response to earnings. However, empirical evidence suggests that the quality of information does inuence investors. For example, Dechow et al. (2009) report that markets show negative reactions following a decline in earnings quality. Francis et al. (2005) show that rms with low earnings quality exhibit higher price-earnings ratios and equity betas. In addition, the authors nd that innate earnings quality has larger expected returns eects than discretionary components. Other market outcomes include the bid-ask spread, which has been used to measure for example liquidity (Amihud and Mendelson, 1986), information asymmetry (Huang and Stoll, 1997). Research has linked high quality earnings with reduced bid-ask spread (Francis et al., 2006). Another branch of research is concerned with non-market outcomes; following Dechow et al. (2009), the stock price reaction to qualied audit opinions is either negative or non-existing. The connection between analysts' forecasts and earnings quality has also been examined, under the perception that high quality earnings will yield smaller forecast errors (Ashbaugh and Pincus, 2001). Francis et al. (2005) show that poor earnings quality results in worse credit ratings. 2.5 Earnings Management Earnings management literature is closely related to earnings quality literature since it is clear that earnings management decreases earnings quality (Dechow and Schrand, 2004). By denition, earnings management induces an intentional bias in nancial reports (Melumad and Nissim, 2008). Numerous denitions of earnings management exist, and most circle around the use of discretion in accounting to achieve a specic goal. Schipper (2003) describe earnings management as subversive implementation of the standard, not consistent with the intent of the standard. Former SEC chairman, Arthur Levitt describe it as the grey area between legitimacy and outright fraud (...) where earnings reports reect the desires of management rather than the underlying nancial performance of the company (Levitt, 13

16 1998b). Healy and Wahlen (1999, p. 368) dene earnings management as:..when managers use judgment in nancial reporting and in structuring transactions to alter nancial reports to either mislead some stakeholders about the underlying economic performance of the company or to inuence contractual outcomes that depend on reported accounting numbers. Financial disclosures can be managed in several ways, for example by selecting accounting methods within GAAP or by applying methods in a way that supporst the picture management wishes to display to users (Schipper, 1989). Jiambalvo (1996) describes two ways of manipulating earnings, real decision and pure accounting decisions. The rst could be delays or accelerations of sales or sale of xed assets to aect gains and losses. The second includes for example changes in accounting principles and changes of estimate of residual value of xed assets. The line between earnings management and fraud is ne, and earnings management does not necessarily equal fraud. According to Levitt (1998b), earnings management is located in the gap between legitimate accounting and outright fraud, whereas Melumad and Nissim (2008) term fraud as an extreme case of earnings management. Some cases of earnings management clearly violate GAAP, whereas other cases are within the borders of GAAP. Following Dechow and Skinner (2000, p. 238), (...) while nancialreporting choices that clearly violate GAAP can clearly constitute of both fraud and earnings management, it also seems as if systematic choices made within GAAP can constitute earnings management. The authors suggest that within-gaap choices can be considered earnings management if they are used to obscure or mask true economic performance. Nelson et al. (2003) categorise earnings management in three categories: earnings management consistent with GAAP, earnings management dicult to distinguish from GAAP, and earnings management clearly violating GAAP. The discussion of whether earnings management violates GAAP is also carried out by Dechow and Schrand (2004) who like Jiambalvo (1996) distinguish between manipulation of either real transaction (e.g. cutting back on R&D expenditures or boosting sales by oering discounts) or accruals (e.g. reducing the allowance for doubtful accounts and avoiding write-os of assets). The former is not a violation of GAAP as long as the rm accounts for the transactions, and the latter is not necessarily a violation of GAAP since the exibility in nancial reporting allows for dierent accounting treatments. Dechow et al. (1996) argue that earnings management is generally within the boundaries of GAAP and expect that the rms violating GAAP with earnings manipulation also manage earnings within GAAP Dechow and Skinner (2000) observe that although numerous measures of earnings management have been devised by researchers, none of them have been very powerful in identifying and predicting the managing of earnings in practise. Leuz et al. (2003) acknowledge that earnings management is 14

17 dicult to measure because it manifests itself in very dierent forms Incentives to Manage Earnings Clearly, there are many dierent incentives to manage earnings, resulting in the many dierent forms of earnings management. One group of incentives is capital market expectations and valuation, a group in which rms manage earnings to inuence short-term stock price performance (Healy and Wahlen, 1999). Examples of this are increasing the share price prior to seasoned equity oerings or decreasing the stock price before management buyout (Dechow and Schrand (2004); Ecker et al. (2006)). Often, managers attempt to report prots (Burgstahler and Dichev, 1997) or sustain recent performance and meet analysts' expectations (Degeorge et al., 1999). However, the opposite could also be true, illustrated for example by Jones (1991), who proves that rms manage earnings to decrease reported income prior to import relief investigations. Hence, rms can have incentives to manage earnings both upwards and downwards. Dechow and Schrand (2004) suggest that earnings management especially occurs after a period of high growth and increasing performance in booming economies. When the economy slows down, managers nd it dicult to meet the expectations set during the boom and a decline in earnings can compel managers to use aggressive accounting, earnings management or even fraud. This view is supported by Richardson et al. (2002), who note that a history of previously reported positive earnings forces management to manage earnings, as they are unwilling to break a string of positive earnings. Dechow and Schrand (2004) also note that managers may create some of the problems themselves by guiding analysts about future results and thus creating unrealistic expectations. In addition to this, they also suggest that the corporate culture as such can aect the likelihood of earnings management. Another group of incentives is contracts written in terms of accounting numbers, such as lending and compensation contracts. These include avoiding debt covenants (Abdelghany (2005); Ecker et al. (2006)), bonus plans and compensation packages (Dechow and Schrand (2004); Richardson et al. (2002)). Healy and Wahlen (1999) also outline regulatory motivations. These can for example be industry-specic regulations for nancial institutions and utilities, or anti-trust regulation (Ecker et al., 2006). The standards also inuence the degree of earnings management, and even though tighter standards, that is, standards that leave less room for discretion, should impede earnings management, the opposite could also be true. According to Ewert and Wagenhofer (2005), tighter accounting standards may lead to a substitution eect, so that accounting earnings management is met with real earnings management. Obviously, some of the incentives are interrelated in some way, e.g., man- 15

18 agement could wish to aect share prices and thus increase the value of their own bonus plans, which is often attached to the market price of the rm's stock Consequences of Earnings Management Following Melumad and Nissim (2008) there are two types of costs associated with earnings management: the costs associated with undetected earnings management and costs following a detection of earnings management. In the rst case, an overstatement of earnings will generally lead to an understatement of future earnings 2. In the second case, costs incurred when earnings management is detected include negative eects on reputation, stock prices, and increased fees to auditors. Burgstahler and Dichev (1997) note that the extent of earnings management is likely to be a function of the ex ante costs of earnings management, such that earnings manipulators are likely to be rms faced with relatively lower ex ante costs of earnings management. Following the discussion of the impact from earnings quality on cost of capital, it seems clear that earnings management is expected to increase the cost of capital. This is documented by Dechow et al. (1996), who attribute the increase to investors' decreased estimates of rm value and credibility. They also show a large, negative stock price reaction after the detection of earnings management is made public, since investors believe that rm value has been overstated. According to Healy and Wahlen (1999), investors are not "fooled" by earnings management and thus nancial statements provide useful information to users. However, Healy and Wahlen (1999) do acknowledge that some studies reach opposite conclusions. For example, Teoh et al. (1998) show that rms with income-increasing abnormal accruals in the year of a seasoned equity oering signicantly underperform in the following years, suggesting that earnings management prior to equity issues aects share prices. 2 For example, if a rm has managed earnings by understating bad debts to increase net receivables and thereby overstate current earnings, it will most likely be forced to write-down in the next period, resulting in a large bad debt expense and lower earnings. 16

19 3 Proxies for Accounting Quality Following the various denitions of accounting quality, the natural next step is guring out how to measure the quality of the rm's nancial statements. This implies the objective measuring of how decision useful the statements are. Table 2 summarises the metrics described. Abnormal Accruals Models Accruals Quality Models Other Accruals Models Avoiding Earnings Decreases and Small Losses Asymmetric Timeliness Smoothness Table 2: Overview of Earnings Quality Metrics The Jones Model The Modied Jones Model The performance-matched Modied Jones Model Dechow-Dichev model Modied Dechow-Dichev Magnitude of Accruals Change in Accruals Avoiding Earnings Decreases Loss Avoidance Timeliness Timely Loss Recognition Variability of Earnings Correlations between Accruals and Cash Flows Large amounts of discretionary accruals not explained by accounting fundamentals indicate poor quality Earnings not mapping closely into cash ows are of low quality High levels of or changes in accruals indicate poor quality Articially avoiding small earnings decreases and losses indicate low quality Less timely recognition of losses implies poor quality Articially smooth earnings are of low quality Persistence Persistence Impersistent earnings indicate low quality Predictability Predictability Earnings not able to predict themselves indicate poor quality Francis et al. (2006) distinguish between two types of proxies for accounting quality: those that reect the innate factors or nature of the rm, and those that reect the surrounding business environment and accounting choices. Innate sources are the economic fundamentals of the rm, such as the operational environment, whereas discretionary sources include for instance managements reporting and accounting choices, auditing and the quality of reporting standards. The intrinsic factors are slow to change relative to the 17

20 discretionary ones. Dechow et al. (2009) separate the proxies in a slightly dierent manner, as they distinguish between a true component of earnings, reecting the real underlying cash ows, and an element of error induced by the accounting choices. As many of the proxies for quality are tailored to a specic study, the proxies are highly context specic. As a result, they can have somewhat contradictory implications, even though most of them are related. Melumad and Nissim (2008) provide the example of articially smooth earnings, which increase the persistence and predictability, but weaken the relation between cash ows and earnings. 3.1 Abnormal Accruals Models A large part of the accounting based metrics examines how accruals directly inuence earnings quality. Earnings consist of both cash ows and accruals and since accruals are discretionary and based on estimates, pure cash ows are generally considered more reliable than earnings (Dechow, 1994). Dechow and Dichev (2002) state that one role of accruals is to shift or adjust the recognition of cash ows over time, so the adjusted numbers better measure rm performance 3. Dechow and Schrand (2004) note that accruals used correctly mitigate irrelevant volatility in cash ows and thus improve the decision usefulness of earnings. Dechow and Skinner (2000) agree with this and argue that accrual accounting as such tend to dampen the uctuations of underlying cash ows, thereby creating a more useful earnings number, than currentperiod cash ows. In line with this view, Melumad and Nissim (2008) argue that earnings smoothed with accruals increase earnings quality, since they improve persistence. Sloan (1996) shows that the accruals component of earnings is less persistent than the cash ow component. However, this does not imply that accruals are not decision useful. As Dechow et al. (2010) note, research has shown that earnings are more persistent than cash ows and that earnings produce smaller forecasting errors than cash ows in valuation models. This suggests that accruals indeed can improve decision usefulness, even if they have lower persistence than cash ows. This also highlights the fact that accruals are useful, although they introduce measurement error and managerial discretion in nancial statements. A large body of literature hypothesises that earnings are primarily misstated via the accruals component (Dechow et al., 2009). Since accruals by 3 Following IASC (1989, Ÿ 22), accrual accounting implies that...the eects of transactions and other events are recognised when they occur (and not as cash or its equivalent is received or paid) and they are recorded in the accounting records and reported in the nancial statements in the periods to which they relate." 18

21 nature are subjective judgements, they do open the door for opportunistic, short-term earnings management, if accruals are used to hide value-relevant changes in cash ows (Dechow and Schrand, 2004). The introduction of estimates also decreases the predictability of earnings, since they are subjective and hard to predict. Francis et al. (2006) note that several proxies of earnings quality are based on the view that accruals, ceteris paribus, reduce earnings quality. Richardson et al. (2002) argue that accrual information is a key determinant of earnings manipulation and that high levels of accruals should be seen as a red ag indicating an increased likelihood of earnings manipulation. Caution is needed when accruals are used to measure earnings quality. Dechow and Schrand (2004) suggest that some rms can erroneously be classied as having low accruals quality and predictability, due to the nature of the business. These are for instance rms with high growth or a large proportion of intangible assets. As mentioned by Dechow and Skinner (2000), certain forms of earnings management, such as income smoothing, are hard to distinguish from appropriate accrual accounting choices. Jiambalvo (1996) shows that accrual models generate low power tests of earnings management even for fairly high levels of earnings management. The line between appropriate exercise of managerial discretion through accruals and earnings management is thus very ne. DeFond (2010) argue that the abnormal accrual models all suer from the inherent limitation that researchers are unable to determine whether the estimated discretionary component is a result of management's discretionary accounting choices or just an artifact of the model used. Thus, every test using abnormal accrual models is a joint test of the hypothesis tested in that specic study and the hypothesis that the proxy is a valid measure. Therefore it is also dicult to evaluate which accrual model is 'best' The Jones Model The abnormal accruals models originally assumed that expected normal accruals are identical to last period's total accruals and that they consist of both normal, non-discretionary accruals (NA) and abnormal, discretionary accruals (DA) (DeAngelo, 1986). This approach thus assumes that the change in total accruals from one period to the next is due to change in discretionary accruals. As noted by Dechow et al. (1995), the DeAngelo model and the somewhat similar Healy model (Healy, 1985) only measure discretionary accruals without error if non-discretionary accruals are constant and discretionary accruals have a mean of zero over the estimation period. The Jones Model (Jones, 1991) relaxes this rather strict assumption. The general view in her model is that accounting fundamentals, such as revenues or assets, should explain accruals. The objective is thus to divide accruals 19

22 into two components: Normal, non-discretionary accruals associated with the rm's fundamental earnings process; and abnormal, discretionary accruals which stem from intentional or unintentional accounting errors (Dechow et al., 2010). Higher levels of accruals which are not associated with the fundamental earnings process of the rm are assumed to reduce the quality of earnings. The accounting fundamentals are thus determinants of unmanipulated accruals. In order of separating normal from abnormal accruals, Jones develops a framework in which she controls for changes in property, plant and equipment, and revenue. Large amounts of accruals not explained by these fundamentals indicate lower earnings quality, since abnormal accruals are sensitive to managerial discretion. (Jones, 1991) calculated total accruals as change in net working capital adjusted for changes in all current operating accounts (adapted from DeAngelo (1986)). Later studies using the Jones Model often nd accruals directly as operational cash ows subtracted from earnings. Jones relaxes the assumption of changes in accruals being due to the discretionary component, by estimating an expectation model for total accruals to control for the economic circumstances of the rm. Thus, Jones species a linear relationship between total accruals and change in sales and property, plant and equipment: T A it = α i 1 A it 1 + β 1i REV it + β 2i P P E it + ε it (1) where T A i t = Total accruals in year t for rm i, scaled by lagged, total assets; A it 1 = Total assets in year t-1 for rm i; REV it = Revenues in year t less revenues in year t-1 for rm i, scaled by lagged, total assets; P P E it = Gross property, plant, and equipment in year t for rm i, scaled by lagged, total assets. Then, the prediction error from the OLS regression, dened as u ip = T A ip (α i 1 A ip 1 + β 1i REV ip + β 2i P P E ip ) represents the level of discretionary accruals for rm i at time t. To test the earnings management hypothesis, Jones tests if the average prediction error is greater than or equal to zero, using Patell (1976). While the prediction error from the regression is used in the original Jones Model, Schipper and Vincent (2003) argue that residuals can also be used as a proxy for discretionary accruals. Jones notes that management naturally cannot hide the true accrued amount in multiple accounting periods, since rm income must equal cash ows over all years. However, it is possible to hide the true nature 20

23 of a rm's earnings in the short run. Originally, the Jones Model used time-series data, but the estimation of abnormal accruals can be use both rm-specic, time-specic or cross-sectional data 4. Although widely used, the Jones Model has also been much criticised. Numerous studies have shown misspecications and low predictive ability in the Jones model (see e.g. McNichols (2000)). Francis et al. (2006) question whether the separation in abnormal and normal accruals reects the true dierence between discretionary and non-discretionary accruals. They show that accounting fundamentals such as rm size and standard deviation of sales revenue explain approximately 65 percent of the variation in the abnormal accruals. McNichols (2002) argues that the Jones Model might be misspecied since the lagged and future eect of change in sales is ignored in it. Bernard and Skinner (1996) suggest that the Jones Model treast some legitimate accruals as abnormal. Beneish (1997) also argues that the separation in discretionary and non-discretionary accruals is not convincing and that managers can exercise discretion, without intentionally inating earnings. He also provides evidence that accrual models have poor detective performance even among rms with extreme earnings management behaviour. Since the correlation between discretionary accruals and total accruals is more than 80 percent, Dechow et al. (2003) note that it might be just as useful to look at overall level of accruals, rather than dividing them into abnormal and abnormal. This is the case for all abnormal accruals model, which can be criticised for misclassifying non-discretionary accruals as discretionary. Finally, Dechow et al. (2009) show that discretionary accruals are generally less powerful than total accruals at detecting earnings management in SEC enforcement releases The Modied Jones Model In their 1995 article, Dechow, Sloan and Sweeney criticise the Jones Model for its implicit assumption on nondiscretionary revenues. They argue that managers can easily use discretionary revenues for instance to speed up revenues before nancial year-end. Such a situation would lead to an increase in revenues and accruals but also in receivables. The accruals attached to this form of earnings management will be classied as non-discretionary accruals in the setting of the Jones Model, causing the estimate of earnings management to be biased towards zero. Therefore, rms that manage revenue will not be detected in the Jones Model. 4 Subramanyam (1996) prefers the cross-sectional version rather than the time-series, since the rst generates a considerably larger sample and the long time-interval in the time-series could cause misspecication in the model due to non-stationarity. 21

24 As a response to this critique, Dechow et al. (1995) extend the model to control for managed revenues, by including accounts receivables when determining non-discretionary accruals, following the intuition from above that overstated revenues will lead to boosted receivables: T A it = α 1 1 A it 1 + α 2 ( REV it REC it ) + α 3 P P E it + ε it (2) where REC it = Net receivables in year t less net receivables in year t-1 for rm i, scaled by lagged total assets; Other variables are as previously dened. Again, the prediction error from the above regression is the estimate on discretionary accruals. The modication is designed to eliminate the tendency of measuring discretionary accruals with error when revenues are managed, and thus changes in revenues are adjusted for changes in receivables. The inclusion of net receivables changes the implicit assumption from the Jones Model dramatically. Hence, the Modied Jones Model assumes that all changes in credit sales result from earnings management. The empirical evidence on the Modied Jones Model versus the Jones Model is somewhat ambiguous. Francis et al. (2006) prefer the Modied Jones model over the Jones Model, whereas Subramanyam (1996) and Stubben (2010) nd that the Modied Jones Model shows no real improvement in performance compared to the original. The Modied Jones Model has like its predecessor been subject to much criticism. Francis et al. (2005) suggest that the measure of abnormal accruals contains a substantial amount of uncertainty and believe that the link to information risk is unclear. Guay et al. (1996) also show imprecision in the Modied Jones Model, but note that this might be the case for all the abnormal accrual models. Following Beneish (1997), the Modied Jones Model could be augmented with lagged total accruals and a measure of past price performance, to control for past rm performance. Dechow et al. (2009) suggest that the Modied Jones Model has higher explanatory power than the original version but suers from the same performance related problems as the Jones Model The Performance-matched Modied Jones Model From the intuition that rms with extreme performance are likely to engage in earnings management and thus have lower earnings quality, Kothari et al. (2005) propose a performance-matched discretionary accrual approach. They match each rm-year observation with another rm from the same industry and year and the closest possible ROA. The paper thus builds on the 22

25 work of Dechow et al. (1995), who showed that both the Jones Model and the Modied Jones model were misspecied when applied to samples with extreme performances. As a control, Kothari et al. (2005) also propose a modied version of the Modied Jones Model, in which they include ROA as an additional regressor besides sales and PPE, thereby controlling for rm performance on discretionary accruals. The authors describe a non-linear relation between accruals and performance, and justify the model with the argument that discretionary accrual models are misspecied under extreme performance. The regression approach imposes stationarity of the relation between accruals and performance through time or in the cross-section. The authors show that misspecication issues are attenuated, albeit not eliminated, when ROA is included. Using current year ROA rather than lagged ROA yields better results. Teh inclusion of ROA as an additional regressor does not impose a specic relation between accruals and performance. T A it = δ 0 + δ 1 1 A it 1 + δ 2 REV it + δ 3 P P E it + δ 4 ROA it + υ it (3) where ROA it = Return on assets in year t for rm i. Other variables are as previously dened. As with the other models, υ it from the above regression is the measure of abnormal accruals. As evident, Kothari et al. (2005) include a constant in the mode, which was not the case for neither the Jones nor the Modied Jones Model. Dechow et al. (2010) note that the approach of Kothari et al. (2005) is likely to add noise to the measure of discretionary accruals, and it is best applied when correlated performance is in fact an important concern. 3.2 Accruals Quality Models Another group of accrual models measure accruals quality as such. The quality of total accruals is of interest, and thus researchers do not attempt to distinguish between normal and abnormal accruals. Accruals quality is consistent with the view that low-variance rms have high earnings quality (Francis et al., 2006) Dechow-Dichev Model The rst model introducing accruals quality as a measure of earnings quality was proposed by Dechow and Dichev (2002). This model is based on the fact that accruals shift the recognition of cash ows over time to better measure rm performance. Since accruals are based on estimates, the incorrect estimates must be corrected in future accruals and earnings; as a consequence, 23

26 estimation errors are noise that reduces the benecial role of accruals 5. Dechow and Dichev predict that the quality of accruals and earnings decreases when the magnitude of estimation errors increases. They therefore propose an empirical measure of accruals quality which maps working capital accruals into operating cash ows. A poor match thus signies low accruals quality. More specically, the model is built up as a regression of the change in working capital accruals on last year, present, and future cash ows. Dechow and Dichev measure accruals as the change in working capital, which yields the following regression: W C it = β 0 + β 1 CF O it 1 + β 2 CF O it + β 3 CF O it+1 + ε it (4) where W C it is working capital in year t less revenues in year t-1 for rm i, scaled by lagged, total assets; CF O it is cash ows from operations in year t for rm i, scaled by lagged, total assets. Changes in working capital is the Dechow-Dichev measure of accruals. The error term shows the extent to which accruals map into realised cash ows, and the variance hereof is a proxy for accruals quality. High variance in the estimation errors yields non-persistent earnings, and it is an inverse measure of earnings quality. The idea is that systematically small or large estimation errors do not create problems for users since these still enable them to predict future earnings. This is intuitively appealing, since a persistent residual does not necessarily equal low accruals quality but can just be a result of a reality in the rm. On the other hand, volatile residuals impede investors' prediction of future earnings, creating an earnings number of low quality. It is expected that rms with low accrual quality will also have low earnings persistence. Dechow and Dichev do not distinguish intentional estimation errors from the unintentional ones, since all errors signify poor accruals quality, regardless the underlying intent. The model therefore deviates from the Jones model, which attempts to capture earnings management, whereas the Dechow- Dichev Model focuses on earnings quality per se. According to Francis et al. (2004), the Dechow-Dichev model is a powerful earnings quality measure. Schipper and Vincent (2003) nd that the Dechow- Dichev Model avoids several of the problems associated with the accounting 5 When cash ows are received after they are recognised, management must estimate the expected, received amount. If the amount is estimated incorrectly, this will naturally be corrected when the accrual is close. However, each period's accrual will contain an estimation error in the opening accrual and a realised error in the closing accrual. Since earnings feature an estimation error and its correction, their ability to measure rm performance is reduced. As a result, a minimisation of the estimation error is desirable. 24

27 fundamental approach by Jones (1991). However, the model requires that working capital accruals lag or lead cash receipts by no more than a year. In a discussion of Dechow and Dichev (2002), McNichols (2002) notes that the Dechow-Dichev Model is only applicable where the key element of accruals is current accruals. In settings where this assumption is not met, the model is not operational. As opposed to this, Francis et al. (2006) nd that total accruals make a legitimate proxy for current accruals. Francis et al. (2005) nd that Dechow-Dichev Model best captures the uncertainty in accruals. They also suggest that the Jones Model and the Dechow-Dichev Model work well together, since the rst does not suer from the limitations of the second, in terms of using current accruals rather than total accruals. According to Dechow et al. (2010), it is an important limitation to the Dechow-Dichev Model that it focuses on short-term working capital, since for instance impairments of goodwill and PPE are likely to reect earnings management. Finally, McNichols (2002) nds misspecication in the model. This is also the case with the Jones model, though Modied Dechow-Dichev Model In her discussion of the Dechow and Dichev (2002) paper, McNichols (2002) argues that measurement error in the Dechow-Dichev model may preclude the model from controlling for the fundamental factors inuencing accruals. She suggests that the model is misspecied; more specically, the residuals from the Dechow-Dichev model are correlated with change in sales, indicating that cash ow from operations is a noisy proxy for the cash ow recognised in the accruals. She thus extends the model to include additional explanatory variables, which are important in forming expectations about current accruals. The extent to which accruals map into cash ows, change in sales and PPE is thus an inverse measure of accruals quality. This yields the following regression (all variables scaled by average assets): W C it = β 0 + β 1 CF O it 1 + β 2 CF O it + β 3 CF O it+1 + β 4 REV it + β 5 P P E it + ε it (5) where all variables are as previously specied. As in the original Dechow-Dichev Model, the variance of ε it is an inverse measure of earnings quality. The Modied Dechow-Dichev Model has shown greater t than the Dechow- Dichev Model (Francis et al., 2005). Kent et al. (2010) nd that Dechow- Dichev and the modied Dechow-Dichev Model perform equally well, but the latter provides a signicantly larger coecient of determination. Barth et al. (2008) criticise the Modied Dechow-Dichev Model for not capturing the perceptions of investors and analysts and the fact that it focuses on current accruals instead of total accruals. 25

28 3.3 Other Accruals Models Magnitude of Accruals The sheer magnitude of accruals has been used as a measure of earnings quality. Bhattacharya et al. (2003) expect that the level of accruals increases with earnings aggressiveness if cash ow realisations are held equal. Following their intuition, aggressive accounting will lead to fewer negative and more positive accruals. This will lead to a higher level of accruals over all since rms are more likely to overstate than to understate earnings. Dechow et al. (2010) argue that extreme accruals are of low quality because they represent a less persistent component of earnings. In accordance with Dechow and Dichev (2002), their proxy for earnings quality and the level of accruals are both proxies for the same aspect of unobservable true accrual quality, as a measure of earnings persistence. They also note that the sheer level of accruals is simple and easy to use. However, the authors still nd that the Dechow-Dichev model does capture better the variation in earnings persistence. An important note made by Dechow and Dichev (2002) is that rms with large accruals tend to generate larger estimation errors, emphasising the fact that the magnitude of accruals and Dechow-Dichev model might be two sides of the same coin. Dechow et al. (2011) examine whether rms that misstate earnings have unusually high working capital accruals. They predict and nd that accruals are larger in misstating years. Dechow and Schrand (2004) also note that the magnitude of accruals can reveal the quality of earnings, but they note that the level cannot be seen independently. Thus, high accruals in companies with low cash ow volatility should "ring the bells", whereas high accruals in for example a rm in the high growth industry, do not necessarily indicate poor earnings quality. Bhattacharya et al. (2003) and Leuz et al. (2003) measure earnings aggressiveness as the level of accruals: where all variables are as previously dened. T A it A it 1 (6) The higher the ratio, the higher the earnings aggressiveness. This measure of earnings quality is certainly simplistic, and thus it is rarely used as a single metric for earnings quality. As Dechow and Schrand (2004) note, some rms will have higher levels of accruals due to the very nature of their business, despite of otherwise honest management. 26

29 3.3.2 Change in Total Accruals The general idea behind this metric is that accruals should be constant over time, and thus a signicant change in accruals could indicate managerial manipulation. Schipper and Vincent (2003) suggest that as long as some portion of accruals is both non-manipulated and approximately constant over time, changes in total accruals could stem from managerial manipulations, and they may provide an inverse measure of earnings quality. Hence, the more accruals change over time, the poorer is the earnings quality. DeAngelo (1986) argues that large accruals as such do not necessarily indicate low earnings quality or earnings management behaviour, but a jump in accruals might indicate that managers have deliberately over- or understated earnings. She therefore assumes that a signicant change in accruals from one period to the next indicates low earnings quality. This rests on the following logic: Given that accruals consist of both normal and discretionary accruals, DeAngelo assumes that average change in normal accruals is zero. Therefore, a signicant change in total accruals reects a change in abnormal accruals. More specically, DeAngelo (1986) calculates the following fraction: where Where all are variables as previously dened. T A it A it 1 (7) A fraction signicantly dierent from 0 indicates poor earnings quality. Sloan (1996) shows that the accrual component of earnings is less persistent than cash ows. One interpretation of this is that current overand understatements of accruals are adjusted via accruals in future periods (Dechow and Schrand, 2004). Thus, the recording and reversal of accrual misstatements result in accruals, that are more volatile than cash ows. As Dechow and Schrand note, these misstatements may simply be due to the fact that managers have to make judgements and forecasts when determining accruals. Therefore, accruals that change from year to year do not necessarily indicate poor EQ. As with the magnitude of accruals, this approach is somewhat simple, but it has often been included in earnings quality studies on par with with other accrual models. In general, it seems to correlate signicantly with other accruals models, indicating that it is useful as an additional robustness test. 3.4 Avoiding Earnings Decreases & Small Losses The discussion on avoidance of small losses and earnings decreases is linked to the reported kink in the distribution of reported earnings, as discussed by Hayn (1995). She argues that losses are not expected to perpetuate 27

30 since rms have liquidation options. Losses are thus less informative about future rm performance than prots are. She shows that even though the overall distribution of earnings is not signicantly dierent from a normal distribution, there is a point of discontinuity around zero. More specically, there is a concentration of cases just above zero and fewer than expected of small losses just below zero. She suggests...that rms whose earnings are expected to fall just below the zero earnings point engage in earnings manipulation to help them cross the `red line' for the year" (Hayn, 1995, p. 132). Thus, it is suggested that the kink in earnings is due to an unwillingness of rms to report losses or earnings decreases. Likewise, Burgstahler and Dichev (1997) make two predictions on earnings management. They posit rst, that earnings are managed to avoid earnings decreases and second, earnings are managed to avoid losses. Avoidance of small losses and earnings decreases have been used extensively in the research, under many dierent terms, e.g, small loss avoidance (Leuz et al., 2003), loss avoidance (Bhattacharya et al., 2003), frequency of small positive earnings (Lang et al., 2003), and managing towards positive earnings Barth et al. (2008). There are a number of incentives for managers to produce steadily increasing earnings numbers. Barth et al. (1995) show that rms with increasing earnings during a longer period enjoy higher price-earnings ratios and higher premiums for the long series of earnings increases. These benets vanish immediately when the line of increasing earnings is broken. These ndings have been supported in other studies later, and thus there seems to be strong incentives to keep earnings positive and steadily increasing. According to, for example, Degeorge et al. (1999) and Dechow et al. (2003), just exactly meeting or beating analysts' forecasts also points towards earnings management. Most of the research is connected with examining the frequency of small increases in earnings or small prots barely over zero. Leuz et al. (2003) note that while one might argue that managers are interested in avoiding all losses and earnings decreases, but, at the same time, they only have limited reporting discretion, it is not possible to conceal larger losses or earnings decreases. Therefore, managers may manage earnings to make them seem increasing or above zero only when possible Avoiding Small Earnings Decreases Burgstahler and Dichev (1997) use small increases in earnings in as a proxy for earnings management. They test the hypothesis of avoiding earnings decreases as: 28

31 where NI it = Net income in year t for rm i; MV E it 1 = market value of equity in year t-1 for rm i. NI it MV E it 1 (8) Under the hypothesis of no earnings management, the expected distribution of earnings change would be approximately symmetric and normal, following Hayn (1995) Loss Avoidance Burgstahler et al. (2006) estimate the small loss avoidance as the frequency of small prots compared to small losses: SP NI t (9) SNNI t where SP NI t = small positive net income, dened as NIt A t 1 between 0 and 1 %; SNNI t = small negative net income, dened as NIt A t 1 between 0 and -1 %. The higher this ratio is, the higher is the loss avoidance. The interpretation that small positive earnings indicate earnings management is somewhat controversial, and several researchers have questioned whether earnings management actually explains the kink in earnings. Dechow et al. (2003) argue that the increase in cash ows around the zero reference earnings point stems from the positive relation between cash ows and earnings. therefore an increase in cash ow is actually expected around the kink. They therefore reject the notion that the increase in cash ows is due to earnings management. Dechow et al. (2003) also argue that since there is a positive relation between working capital accruals and earnings, Burgstahler and Dichev's ndings concerning an increase around zero do not necessarily indicate earnings management. Since Dechow et al. (2003) nd that other accruals decrease for the small prot group relative to the small loss group, the conclusions of Burgstahler and Dichev (1997) on earnings management is - according to Dechow et al. (2003) - questionable. The main criticism of Dechow et al. (2003) pertains to the fact that rms can take real actions to avoid reporting losses or earnings decreases, and thus the overrepresentation of these two incidents is not necessarily evidence of earnings management. One can easily imagine that employees are more motivated when facing a loss and managers may make decisions that increase cash ows and hence earnings, absent of earnings management. Coulton et al. (2005) provide similar criticism in their examination of rms that beat a simple benchmark, such as achieving increasing earnings. 29

32 They argue that the kink in earnings is a poor proxy for earnings management and that the kink could just as well be attributable to the scaling of earnings by for instance lagged assets or price. However, a group of researchers have shown correlations between small prots and other earnings management proxies, indicating that small positive earnings could possibly stem from earnings management. 3.5 Asymmetric Timeliness The research of timeliness hypothesises that prots and losses are recognised in an asymmetric manner. More specically, losses are recognised on a more timely basis than prots, leading to less persistent and more reverting negative earnings changes. For asymmetric timeliness to be an indicator of earnings quality, it is assumed that the more timely recognition of losses in fact increases decision usefulness. According to Dechow et al. (2010), previous literature suggest that equity markets perceive asymmetric timeliness as increasing earnings quality. However, research has so far failed to prove that asymmetric timeliness in fact improves decision making. Ball and Shivakumar (2005) note that timely loss recognition increases the value relevance of nancial reporting. Asymmetric timeliness is related to conditional conservatism, and a more timely recognition of losses is often associated with a conservative accounting system (Basu, 1997) Timeliness A frequently used measure of asymmetric timeliness is Basu (1997)'s reverse earnings-returns regression also used to detect conservatism. Basu (1997) proposes a second measure of timeliness which is not based on returns: NI t = α 0 + α 1 NEGDUM t 1 + α 2 NI t 1 + α 3 (NEGDUM t 1 NI t 1 ) + υ t (10) where NI t = change in income from year t-1 to t, scaled by lagged total assets; NEGDUM t 1 = an indicator variable equal to 1 if NI t is negative. Timely recognition of economic losses implies that they are less persistent and tend to reverse faster than prots. This predicts that α 3 < 0. Basu (1997) nds support for this prediction Timely Loss Recognition The main intuition behind timely loss recognition is the fact that rms with high nancial reporting quality recognise losses as they occur, rather than 30

33 deferring them (Lang et al., 2006). This will lead to a higher frequency of large losses. Such a high frequency also indicates that earnings have not been articially smoothed. If earnings had been smoothed, large losses should be relatively rare. The opposite can, however, also be true, since a high frequency of large losses could indicate big bath earnings management. There is thus a conict between the two. Ball and Shivakumar (2005) argue that timely loss recognition increases nancial statement usefulness, in particular in corporate governance and debt agreements. The rst is aected because managers are less likely to make NPV-negative investments. The second is aected because timely loss recognition provides more accurate information for loan pricing. They also note that the demand for timely gain recognition is smaller since mangers have natural incentives to report prots. Timely loss recognition is linked to some of the other metrics for earnings quality. Barth et al. (2008) suggest that one characteristic of high quality earnings is that large losses are recognised as they occur, rather than being deferred to future periods. This characteristic is closely related to smoothing since large losses should be relatively rare if earnings are smoothed. Therefore, rms reporting (large) losses on a regular basis have higher quality earnings than those that do not, from the intuition that the latter are expected to have managed earnings. According to Beekes et al. (2004), timely loss recognition is also linked to the asymmetric timeliness of earnings and therefore conservatism. Lang et al. (2006) and Barth et al. (2008) use the frequency of large losses as an indicator of earnings quality: where all variables are as previously dened. NI it AT it (11) If the above ratio is less than -0.2, it is dened as a large loss. A high frequency of those is juxtaposed with high accounting quality, since losses are recognised as they occur. They use an indicator variable equalling 1 if annual net income scaled by total assets is less than -0.2, and 0 otherwise. They use this in a regression where they regress a cross-listing variable on the negative NI variable -a negative coecient indicates that cross-listing rms are less likely to report large losses. 3.6 Smoothness Two conicting views on smooth earnings as an indicator of earnings quality exist in the literature. One view reects the idea that managers articially smooth out relevant uctuations. This leads to a less timely and informative earnings number. In this view, smooth earnings indicate poor quality 31

34 earnings. The opposite view reects the idea that management uses private information to smooth out transitory, value irrelevant uctuations in earnings, thereby achieving a more useful earnings number. In this view, smooth earnings indicate high quality earnings (Francis et al., 2006). The rst view, that smoothing decreases earnings quality, stems from the hypothesis that management responds to a negative (positive) cash ow stream by increasing (decreasing) accruals (Barth et al., 2008). According to Kirschenheiter and Melumad (2002), managers have several incentives to report smooth earnings. First, the authors argue that the market rewards smooth earnings since they are assumed to have higher precision. Therefore, if the rm reports a large, positive earnings surprise, the positive eect on stock prices might be dampened since investors prefer smooth, unsurprising earnings. Second, consistently positive earnings may raise the expectations of cash ows to investors, thereby increasing share prices. Managers can also wish to appear less risky and attract inexpensive capital. Francis et al. (2004) suggest that capital market participants reward smoother earnings streams with reduced costs of equity and debt. Bhattacharya et al. (2003) use earnings smoothing as a measure of earnings opacity. They argue that articially smooth earnings fail to depict the true swings in underlying rm performance. As a consequence, smoothing increases earnings opacity. Supporters of the second view argue that very volatile earnings variability could indicate poor accounting quality since it could arise due to big bath earnings management (Healy, 1985). Big baths refers to an earnings management technique, where management understate losses to be able to report future prots. Barth et al. (2008) also note that high earnings variability could indicate low accounting quality since it could be due to errors in estimating accruals. Dechow and Skinner (2000) give appealing examples to show that the line between smoothing as a way to include only material changes and uctuations and smoothing as opportunistic earnings management is subtle. As a consequence, they note that detecting earnings management via smoothing in large samples is extremely dicult. They also note that to be able to characterise income smoothing as earnings management, one needs to dene the point at which managers' accrual decisions result in too much earnings smoothing. This is naturally not an easy task. In a related vein, Subramanyam (1996) notes that while some smoothing has an opportunistic connotation, not all smoothing is necessarily opportunistic, since managers might smooth earnings to create more persistent earnings. Thus, smoothing can enhance the value relevance of earnings. The question thus remains whether or not smoothness per se in fact is a quality indicator. Smooth earnings are not necessarily desirable attributes following the concepts statements. But smoothness is an outcome of an accrual-based system assumed to improve decision usefulness, and as such 32

35 not the ultimate goal of the system. Researchers thus need to dierentiate fundamental smoothness from smoothness as an outcome of discretionary accounting choices. Following Dechow et al. (2010) the smoothing measures known so far only measure if the earnings stream is smooth, not why it is Variability of Earnings Earnings variability is often measured by the standard deviation or the variance of earnings, either as a stand-alone measure or relative to the underlying cash ows. Leuz et al. (2003), Lang et al. (2003) and Francis et al. (2004) measure smoothing as the ratio of standard deviation of earnings to the standard deviation of cash ows: σ(ni i /A it 1 ) (12) σ(cf O i /A it 1 ) where all variables are as previously dened. If rms use accruals to manage earnings, the variability of change in operating income should be lower than that of cash ows. Low ratios hence indicate that insiders exercise accounting discretion to smooth reported earnings. Francis et al. (2004) use net income before extraordinary items in the numerator, whereas Leuz et al. (2003) use operating income. Barth et al. (2008) test the level of earnings smoothing by testing the variability of earnings directly: where all variables are as previously dened. σ( NI it ) (13) Higher values indicate less smoothing and thus higher accounting quality after the rst view described in section Correlations between Accruals and Cash Flows Dechow (1994) argues that cash ows and accruals are expected to be negatively correlated over time as a natural result of accrual accounting, because accruals reverse over time. However, she also suggests that a large negative correlation between accruals and cash ows could indicate that accruals are used to smooth uctuations in cash ows, suggesting lower accounting quality. Following this intuition, management responds to low (high) cash ows by increasing (decreasing) accruals, thus boosting (lowering) income. Leuz et al. (2003) and Bhattacharya et al. (2003) use the measure of negative correlation between change in accruals and change in cash ows: ρ( T A it, CF O it ) (14) 33

36 where all variables are as previously dened. Barth et al. (2008) assume that high quality rms will exhibit a less negative correlation between accruals and cash ows than low quality rms. Like Dechow (1994), they acknowledge that the proper rule of accruals is to smooth variability in cash ows, and that the correlation by denition is expected to be positive, because accruals reverse over time. 3.7 Persistence Persistent earnings can be viewed as desirable since earnings that are able to predict themselves are more valuable for users, e.g., for valuation purposes. Persistence or sustainability has therefore been used often as a measure of high earnings quality (Francis et al., 2006). Regardless of the sign and magnitude of earnings, persistence captures the extent to which the current period innovation becomes a permanent part of the earnings series (Schipper and Vincent, 2003). The intuition behind persistence as an earnings quality metric is that persistent earnings will make current earnings a more useful measure of future performance in perpetuity. Thus, higher earnings persistence is of higher quality when the earnings is also value relevant (Dechow et al., 2010). Following Schipper and Vincent (2003) persistent earnings have been associated with larger investor responses to earnings, which supports the hypothesis that persistent earnings are more useful for users, in particular for valuation purposes (Dechow and Schrand, 2004). Dechow et al. (2009) note that studies of earnings persistence and cash ow predictability are motivated by an assumption that persistence improves decision usefulness in an equity valuation context. A common measure for persistence is the autocorrelation of earnings where high autocorrelation between current and past income is desirable A stationary AR1 model with φ close to 1 is thus considered persistent (Heij et al., 2004). Francis et al. (2004) use an autoregressive model on earnings per share to measure persisitence: X j,t = φ 0,j + φ 1,j X j,t 1 + ε j,t (15) where X j,t = Net income before extraordinary items in year t for rm i, scaled by the weighted average number of outstanding shares during year t; Large values of φ 1,j indicate more persistent earnings. In continuation of his discussion of conservatism, Basu (1997) argues that negative earnings changes are less persistent that than positive earnings changes. As a consequence, negative earnings changes do generally not 34

37 become a permanent part of future earnings, whereas good earnings changes will. The accruals quality measure proposed by Dechow and Dichev (2002) is related to earnings persistence, since rms with low accrual quality have a larger amount of accruals that are unrelated to cash ows, which induces noise and less persistency in earnings. This is natural, since more accrual estimation errors will lead to less persistent earnings. However, it has been documented several times (e.g. Sloan (1996)) that accruals are less persistent than cash ows. The authors also note that it is dicult to distinguish empirically between the eects of accruals quality and the level of accruals on earnings persistence. Schipper and Vincent (2003) argue that highly impersistent earnings can be the outcome of neutral application of accounting standards in volatile economic environments. Thus, they do not necessarily indicate poor accounting quality. 3.8 Predictability The concept of earnings predictability as a desirable attribute is closely connected with persistence. According to Schipper and Vincent (2003), predictability is the ability of the nancial statements to improve users' abilities to forecast items of interest, i.e., the ability of past earnings to predict future earnings. Following this denition, variability decreases predictability, and the term is therefore connected to both the sustainable and smoothing literature. Like these two attributes, some ambiguity still exists on whether predictability is actually a desirable attribute of earnings since it is not directly consistent with representational faithfulness. It is evident, though, that predictable earnings are valuable inputs for valuation purposes, such as DCF analysis. Schipper and Vincent (2003) note a contradiction between predictability and persistence; in cases where the variance of the typical shock to the series i large, highly persistent earnings (a random walk) will have low predictability. Hence, in this situation earnings that are of high quality under the persistence view are of poor quality under the predictability view. Predictability can be measured using the same model used for measuring persistence. Another measure often used is the forecast error of analysts' earnings forecasts. Dichev and Tang (2009) note that high volatility decreases earnings predictability. They calculate absolute predictability from autoregressive regressions of current on 1-year lagged earnings, i.e., the same AR1 process as above: where all variables are as previously dened. X j,t = φ 0,j + φ 1,j X j,t 1 + ε j,t (16) 35

38 Consequently, the variance of ε is Dichev and Tang's inverse measure of predictability, since the variance of the error term captures the variation in earnings remaining after accounting for the eect of the autoregressive coecient. Schipper and Vincent (2003) mention some empirical diculties when operationalising predictability. The choice of time period is not agreed upon in the literature, even though many researchers have used one-year-ahead predictions. They also note that no consensus exists on what to predict; researchers have used reported net income, cash ows and various subsets of net income. Finally, they criticise predictability for suering the same issues as income smoothing since it has not been claried whether predictable earnings indicate high quality earnings or opportunistic earnings smoothing. 36

39 4 Restatements A restatement event means a correction of errors or irregularities in the - nancial statements. A broader denition of restatements also exists, e.g., providing restated results after adoption of new accounting standards or an M&A (Scholz, 2008). This study, however, only focuses on correction of misstated nancial statements. Extensive research has been done on rms restating their nancial statements. The properties of restatements in an earnings quality context are indeed compelling. First of all, an outside source has identied problems with the quality of the nancial statements of restating rms (Dechow et al., 2010). Thus, a perceived advantage of restatements is that they are a direct proxy for poor earnings quality (DeFond, 2010). Violations could include nancial and accounting fraud, insider trading, market manipulation, providing false or misleading information, and selling securities without proper registration (GAO, 2009). Thus, irregularities cover both fraud and earnings management within/outside GAAP. Following Scholz (2008), the major reasons for restatements include revenue (e.g. improper or questionable recognition of revenue), expenses (e.g. improper capitalisation of expenditures), and reclassication and disclosure (e.g. categorisation of debt payments as investments). The number of fraud cases has been fairly stable during the past 15 years, namely approximately 5 %. However, the number of rms that restate has increased dramatically in the last ten years - from 90 in 1997 to 1,577 in This rise is among other things traceable back the downturn of the American economy in the beginning of the new millennium, and the enaction of SOX in 2002, as well as various accounting issues in the mid 2000s. 4.1 Previous Literature Extensive research exists on both the determinants and consequences of restatements. The research on which rms that will misstate earnings is extensive. Some common characteristics of restating rms include high growth prior to restatements (Beneish, 1999a), executive compensation contracts (Burns and Kedia, 2006), high leverage and high likelihood to violate debt covenants (Dechow et al., 1996) and poor corporate governance (Farber, 2005), Dechow et al. (1996). Overall, the evidence on opportunistic reporting incentives is mixed, and the results across researchers are far from unambiguous. As an example, Beneish (1999a) do not nd debt covenants to be signicant and Dechow et al. (1996) only nd signicant inuence from some corporate governance characteristics, not all. 37

40 A range of literature also explores the consequences of SEC restatements. One would clearly expect market reactions to be negative, as restatements will aect investors' condence negatively. (GAO, 2002) and (GAO, 2006b) use a standard event study to determine the impact on stock prices following restatements. They nd that stock prices decline signicantly in a two-day window, whereas the long-term impact is more dicult to determine, but it is assumed to be negative as well. Investors respond dierently to dierent reasons for restating and thus reacted more negatively to the category restructuring, assets, or inventory than to the category cost or expense. Palmrose et al. (2004) nd similar results. While it is clear that investors react negatively to misstatements, empirical research has not been able to clearly determine if restatements equal poor earnings quality. Other outcomes of earnings restatements include increased management turnover (Feroz et al., 1991) and cost of capital (Dechow et al., 1996), Hribar and Jenkins (2004), signicant negative stock returns (Dechow et al., 1996) and obviously considerable costs attached to auditors, lawyers etc. Again, there exists some ambiguity as to the consequences of restatements. For example Beneish (1999a) examines the incentives and consequences of earnings overstatements and nds that revelation in SEC does not impose serious enough consequences on managers, and thus the presence of SEC alone will not prevent managers from engaging in earnings management. In contrast, Desai et al. (2006) nd that management turnover increases signicantly following a restatement. This suggests that the "risk of getting caught" might prevent managers from earnings overstatement. 4.2 Identifying Restatements Several sources identify restatement events. This study exploits SEC's restatement database which is described in detail below. Some other sources are described as well, along with my reason for not using these resources SEC's AAERs The US Securities and Exchange Commission (SEC) is the primary federal agency involved in accounting requirements for publicly traded companies. It recognises FASB's US GAAP as the general accepted standards which rms should comply with. The role of SEC is to protect users of nancial statements, in particular investors. It is therefore in the interest of SEC to ensure and maintain a high accounting quality among American corporations, and hence they monitor American rms for signs of reporting violations. If such a sign occurs, SEC publishes an Accounting and Auditing Enforcement Re- 38

41 lease (AAER) 6 on that particular company to inform investors that some kind of nancial reporting violation has taken place. The AAERs have been issued since April 1982 and provide information on the nancial statement quality as a whole, not just earnings. SEC identies the rms that allegedly misconduct nancial statements through anonymous tips, public criticism and news reports, voluntary restatements and random sampling among listed companies. Since SEC is concerned mainly with protecting investors, it is more likely to scrutinise rms that are either large, raising debt or equity, or IPO rms. Since SEC only has limited funds available, they are likely to pursuit only the worst cases of earnings manipulation (Hennes et al., 2008), and thus it is agreed among most researchers that the AAERs include mostly fraudulent or intentional misstating behaviour (Dechow et al. (2010) and Eilifsen and Messier Jr. (2000)). Indeed, in most AAERs, SEC accuses managers of intentional misstating nancial statements, i.e., fraud. However, in some cases SEC acknowledges that management was negligent, i.e., reckless in not knowing (Dechow et al., 2010). Dechow et al. (1996) assume that SEC correctly identies overstating rms and that the rms have knowingly engaged themselves in earnings overstating. Under this assumption, the AAERs are a powerful tool for examining earnings management and accounting quality hypotheses. The dataset excludes all restatements for example due to M&As and many unintentional errors. Since AAERs are likely to include only the cases of intentional misstatements, the sample has a low expected type I error rate (rms that do not manage earnings are incorrectly classied as managers) (Dechow et al., 2011). However, given the limited funds of the SEC, the expected type II error rate might be high as well (many rms are likely to remain undetected) Other Sources GAO Database 7 The Government Accountability Oce (GAO) is the audit, evaluation and investigative arm of the American Congress. GAO has issued lists of restating rms in 2002 and 2006, respectively. The restatements were identied searching the database Lexis-Nexis for variations of restate and other relevant words. GAO also searched SEC's ling, company web sites and compared qualitative features of the rms. Excluded from the lists are rms not registered with SEC, routine reporting issues such as mergers or stock splits, simple presentation issues and restatements following accounting policy changes. Thus, the list consists of restatements 6 A list of the AAERs is available through the commercial database Audit Analytics and from 7 Available from 39

42 following accounting fraud and accounting errors, by GAO termed aggressive accounting issues (GAO, 2006a). It is clear that there is an overlap between the databases of SEC and GAO, because SEC often requires restatements and restatements often trigger SEC's investigation. The GAO sample is larger than the SEC sample each year, but it only goes back to 2002 (compared to 1982 for SEC). The GAO list also includes a wider range of misstatements. The heterogeneity of the sample has been subject to some criticism. Dechow et al. (2010) argue that the GAO dataset includes too many unintentional misstatements and non-material errors which have nothing to do with earnings management. Hennes et al. (2008) suggest that researchers should split the GAO dataset into intentional earnings management and errors when doing accounting research. Another major drawback when working with the GAO dataset is the the time-lag from the restatement is detected, until it becomes public. This span varies greatly from case to case. This makes the GAO database well-suited for research on the consequences of restatements but less appropriate for research on the determinants of restatements (Dechow et al., 2009). Standard Law Database on Shareholder Lawsuit The cases in the shareholder lawsuit database have been used as an indicator of poor earnings quality. The database includes intentional misstatements, but lawsuits could also arise due to other issues, for example after a major decline of stock prices. The sample is thus very heterogeneous, and it has to be carefully cleaned before it is used to hypothesise about the quality of the rms involved. Study Specic Identications Some researchers have also created data sets specically to the study or research question at hand. Richardson et al. (2002) create their own dataset (like GAO) from 1971 to 2000, excluding unintentional errors and misapplications. Many of the rms identied were SEC targets as well. Palmrose et al. (2004) search for words like restate and restatements (like GAO) and assume that SEC targets are a part of this sample. Abbott et al. (2004) use rms that have restated but have not received an allegation of fraud by SEC, also similar to the GAO dataset. Beneish (1997), Beneish (1999a), and Beneish (1999b) combine SEC's rms with search in press releases to prevent the time-lag in the SEC database to inuence results. 4.3 Issues when Working with Restatements One issue when testing accounting quality hypotheses on restating dataset is that most of the rms identied there have in fact violated GAAP, i.e. the 40

43 cases consist of more fraud than earnings management. However, there are many degrees of earnings management that exercise managerial discretion within the borders of GAAP, and these rms are not likely to be caught by the dataset (Dechow and Skinner, 2000). The fact that only the more spectacular cases of earnings management or fraud are included limits the generalisability of the results (Dechow et al., 1996). The view is supported by Jiambalvo (1996), who mentions that the SEC sample is only suitable for research on GAAP violating earnings management. Therefore restatements and AAERs may not be appropriate for studies that wish to capture errors or earnings management within the boarders of GAAP. However, since Dechow et al. (1996) expect that rms violating GAAP also manage earnings within GAAP. This certies the use of restatements as a proxy for earnings management, because those rms identied by SEC have most likely also been engaged in the less severe earnings management. Another issue when working with restatements is highlighted by Dechow et al. (2010): The misstatement samples are often small, whereas the number of potential sources of incentives is large. So the tests may not be powerful enough to detect a true relation. In other words, it can be dicult to show empirically the exact causality between an incentive to manage earnings and an actual restatement. Palmrose et al. (2004) suggest that one of the most important limitations posed by the dierent restatement samples is the selection bias, namely that only the rms detected and judged are in the sample. The selection bias depends on the specic accounting irregularity and the decision of the rm to report it. This corresponds to a high Type II error rate as described earlier. 41

44 5 Hypotheses Development Previous research that links restatements and earnings quality (e.g. Dechow et al. (2011); Jones et al. (2008); Richardson et al. (2002)) focuses mainly on the actual restatement years. In particular, in a time-series analysis of misstating rms, Dechow et al. (2011) compare each rm's non-misstating years with misstating ones and thus assume that misstating rms have poor accounting quality in the restatement year alone. An underlying assumption is thus that accounting quality is only poorer in the misstatement years, when the rm was actually detected. I hypothesise that the poor quality of earnings is not limited to the actual restatement year, but that rms identied by SEC have continuously poor earnings quality in the years prior to the restatement(s). I hence test if the restating rms are generically dierent from otherwise similar, non-restating rms in the years before the restatement. In consequence, the rst hypothesis to be tested is the following: Hypothesis 1. Restating rms have poor accounting quality compared to non-restating rms prior to a restatement event. I compare earnings quality before the (last) restatement year for both restaters and non-restaters. I predict that the poor earnings quality detected by previous research in the restatement years is evident already before the event, and thus restaters are expected to have lower accounting quality than non-restaters. The opposite could, however, also be true. As assumed by for instance Dechow et al. (2011), rms may only exercise improper discretion in the restatement years. It could also be the case that rms manage earnings within the boundaries of GAAP before they are detected by the SEC, but the earnings quality metrics I apply cannot actually detect it. Given the severe market reactions following a restatement outlined in Section 4 (e.g. signicant drops in stock prices and increased management turnover), I expect that restating rms conform following a restatement event, and thus improve accounting quality after the actual restatement: Hypothesis 2. Restating and non-restating rms have similar accounting quality after a restatement event. The earnings quality of restaters and non-restaters is compared after the last restatement year. Under the assumption that the SEC enforcement has an educative role, it is expected that the accounting quality is improved relatively more for the restaters than the non-restaters, and thus I predict no dierence in quality between the two groups in the after-period. Similarly, I compare the change in earnings quality over the periods, for both groups: 42

45 Hypothesis 3. Restating rms improve accounting quality relatively more than non-restating rms following a restatement event. It is possible to nd reverse results, though. As evident from Section 6, some rms restate more than once which could indicate that an immediate improvement fails to take place. Again, it is also possible that the metrics I use are not able to measure a possible improvement. 43

46 6 Research Design and Descriptive Statistics An overview of the research design and matching procedure is depicted in Figure 2. Each restating rm is identied through Audit Analytics' (AA) Non-Reliance Database, after which the nancial statement information from these rms is found in Compustat Unrestated Quarterly. Each restating rm is then matched in the restatement year with a similar rm that has not restated, based on size, ROA, and industry. Identify restating rms in AA Match on name Accounting information in Compustat Restating Events Industry Size ROA All non-restating rms from Compustat Figure 2: Research Design and Matching Procedure All data coding and empirical tests are programmed in SAS, except panel data regressions which are conducted in Stata. 6.1 Sample Selection The restatement events are identied from Audit Analytics, a commercial database available through WRDS. The restatement data set from Audit Analytics covers all SEC registrants (US rms only) that have restated since January 1st The sample contains Form 8-K and 8-K/A lings under the title "4.02: Non-Reliance on Previously Issued Financial Statements or a Related Audit Report or Completed Interim Review". Audit Analytics contains detailed information on which year(s) each rm has restated, the date the restatement became publicly known, and the rea- 44

47 son for the restatement. Since only limited accounting information is available in AA, the information has to be downloaded from Compustat or a similar source. There is one major caveat when using restatement information from Audit Analytics, namely that there is no direct link between AA and for instance Compustat and CRSP. This complicates the process of nding the nancial statement numbers from the rms identied in Audit Analytics. Whereas AA uses CIK as rm identier, Compustat uses GVKEY, and CRSP uses CUSIP. Even though all data bases report the Ticker symbol for each company, this cannot be used to match, since it can change over time and be reused by more than one company. As suggested by WRDS, 8 I match each restating rm in Audit Analytics to its nancial statement information in Compustat by the company name. More specically, I perform a fuzzy merge using the SAS function COMPGED(), which returns the generalised edit distance between two distances, i.e., the dissimilarity between the two. 9 To be certain that the rm from Compustat is in fact identical to the restating rm from Audit Analytics, I accept the match if the distance between the two strings is less than Of the 7,325 unique restating rms present in Audit Analytics' restatement database, the above matching procedure yields 3,817 rms matched with their GVKEY in Compustat. The nancial statement information is taken from the add-on database to Compustat, Unrestated Quarterly. This contains both the original, unrestated data value and the backlled, restated data value for each item. 11 Even though some researchers (e.g. Dechow et al. (2011)) have used restated data values, I prefer using the original values for several reasons: First, according to Standard & Poor's, 12 the number of rms in Compustat with restated data values is as high as 35%. Second, the sign and magnitude of the dierence between the restated and unrestated data values varies enormously and independently on each of the three dimensions: from rm to rm, from time period to time period, and from data item to data item. Thus, it is not possible a priori to predict which data values that will dier and in which direction. 8 WRDS\%20Knowledge\%20Base\%20with\%20FAQs.cfm?folder_id=645&article_id= viewer.htm#a htm 10 A COMPGED score of 0 indicates perfect match (e.g. BIODEL INC = BIODEL INC), whereas a COMPGED score of 90 indicates a nearly perfect match (e.g. RVB HOLDINGS LTD = R.V.B. HOLDING, LTD.) 11 I gratefully acknowledge the nancial support of FSRs Studie- & Understøttelsesfond to buy access to the Compustat Unrestated Quarterly Database

48 Third, the dierence between the two values is material for some data items, which can inuence results. Table 3: Compare original/restated values for Restating Firms Line Item Original Value Restated Value Dierence Assets 2, , *** COGS 1, , Debt, long-term Debt, short-term ** Earnings Per Share Net Income Operational Cash Flows *** R & D Expense ** Revenue 1, , *** The table depicts the matched restaters in all available years. In millions USD, except earnings per share. All items have been winsorised at the 1st and 99th percentile. *, **, *** denote signicant dierence from 0 at the 10%, 5%, and 1% levels, respectively. Using paired t-tests. Table 3 compares specic line items that previous literature has identied as being the common reasons for restatements. More specically, Scholz (2008) states that the majority of restatements are due to revenue recognition, core expenses (e.g. R& D expenses), and disclosure issues (such as reclassication of debt). It is clear that material dierences exist between restated and original data values for some of the line items of interest. In particular, both assets and revenue show signicant dierences. This is in line with the argument of Scholz (2008) that many restatements contain revenue recognition issues, as described in Section 4. Since Compustat Unrestated Quarterly is given in quarterly data, I convert it to annual data by using the fourth quarter amount for balance sheet and cash ow amounts and by summarising the four quarters for each year to one year for the income statement amounts. Furthermore, all data values are winsorised at the 1st and the 99th percentile. 6.2 Matching Procedure For each restating rm, I nd a match among all Compustat rms, that are not identied by SEC. The Compustat rm must meet the following requirements: Financial statement information in the restatement year 46

49 Same 1-digit SIC code Total assets within ± 40% ROA within ± 40% If a restater has more than one match, I minimise the joined absolute dierence between assets and ROA. Of the 3,817 restating rms, 2,028 obtain a unique match. Each restating rm has one exact match in the control group of non-restating rms, with correspondent restatement and non-restatement years, respectively. I split the sample in before and after the restatement and non-restatement year, respectively. The before sample consists of ten years before the actual restatement event. If a rm has restated multiple times, the last restatement year ends the before sample. The after sample lasts a minimum of three and a maximum of ten year after the last restatement/non-restatement event. The sample thus consists of rm years from 1990 (ten years before the rst recorded restatement event in 1990) to The matching procedure is consistent with previous research; Dechow et al. (1996) match on industry, year, and rm size, and Beneish (1999b) use industry, year, and rm age. The underlying assumption behind this research design is naturally that the matched, non-restating control group on average has higher earnings quality and/or manages earnings less than the restating rms. If the rms in the control group also have poor earnings quality but are simply not detected by the SEC, it will seriously inuence the conclusions. But as mentioned in Section 5.2, since SEC mostly pursuits larger rms the matching on size partly deals with this problem, as the control group is likely to have been scrutinised and approved by the SEC. Kothari et al. (2005) and Dechow et al. (2010) discuss another serious concern related to a matched sample in accounting quality studies, namely that rms in the control group have similar incentives to manage earnings. In particular, Dechow et al. (2010) argue that earnings manipulation and earnings quality issues appear to cluster by industry. The fact that I match very broad on industry (1-digit SIC code) partly mitigates this issue, even though it remains a concern that has to be taken into account when interpreting the results. If both of the above concerns impact the results, it will increase the likelihood of maintaining the hypothesis of no dierence in earnings quality between the two groups, i.e., Type II errors A failure to reject a false null hypothesis 47

50 The Jones Model The Modied Jones Model Table 4: Calculation of Earnings Quality Metrics T Ait = αi + β1i REVit +β2ip P Eit + ρ(cf O)it + T Ait 1 + εit T Ait = αi + β1i REVit RECit +β2ip P Eit + ρ(cf O)it + T Ait 1 + εit TA = (Income bef. Ex. Items (123)-(Cash Flows from Operating Activites (308)-Ex. Items (124)))/Lagged Assets (6) REV = Sales (12)/Lagged Assets (6) PPE = Gross Property, Plant, and Equipment (7)/Lagged Assets CFO = Cash Flows from Operating Activites (308)/Lagged Assets (6) TA = (Income bef. Ex. Items (123)-(Cash Flows from Operating Activites (308)-Ex. Items (124)))/Lagged Assets (6) REV = Sales (12)/Lagged Assets (6) REC = Receivables (2)/Lagged Assets (6) PPE = Gross Property, Plant, and Equipment (7)/Lagged Assets CFO = Cash Flows from Operating Activites (308)/Lagged Assets (6) Residual represents discretionary accruals. High levels indicate poor earnings quality Residual represents discretionary accruals. High levels indicate poor earnings quality 48

51 The Performancematched Modied Jones Model T Ait = δ0 + δ1 + δ2 REVit + δ3p P Eit +δ4roait + ρ(cf O)it + T Ait 1 + εit Dechow-Dichev Model W Cit = β0 + β1cf Oit 1 +β2cf Oit + β3cf Oit+1 + εit Modied Dechow-Dichev Model W Cit = β0 + β1cf Oit 1 +β2cf Oit + βcf Oit+1 +β4 RECit + β5p P Eit + εit TA = (Income bef. Ex. Items (123)-(Cash Flows from Operating Activites (308)-Ex. Items (124)))/Lagged Assets (6) REV = Sales (12)/Lagged Assets (6) PPE = Gross Property, Plant, and Equipment (7)/Lagged Assets ROA = Net income (172)/Total Assets (6) CFO = Cash Flows from Operating Activites (308)/Lagged Assets (6) WC = -( Acc. Receivables (302) + Inventory (303) + Acc. Payables (304) + Taxes Accrued (305) + Other Assets and Liabilities (307))/Lagged Assets (6) CFO = Cash Flows from Operating Activites (308)/Lagged Assets (6) WC = -( Acc. Receivables (302) + Inventory (303) + Acc. Payables (304) + Taxes Accrued (305) + Other Assets and Liabilities (307))/Lagged Assets (6) CFO = Cash Flows from Operating Activites (308)/Lagged Assets (6) PPE = Gross Property, Plant, and Equipment (7)/Lagged Assets (6) REC = Sales (12)/Lagged Assets (6) Residual represents discretionary accruals. High levels indicate poor earnings quality High variance in the residual indicates poor earnings quality High variance in the residual indicates poor earnings quality 49

52 Magnitude of Accruals Change in Accruals Avoiding Earnings Decreases Small Loss Avoidance Asymmetric Timeliness T Ait Ai,t 1 T Ait Ai,t 1 NIit Ait 1 SP NIit SNNIit NIit = α0 + α1negdumit 1+ α2 NIit 1 + α3 (NEGDUMit 1 NIit 1) + υit TA = Income bef. Ex. Items (123)-(Cash Flows from Operating Activites (308)-Ex. Items (124)) A = Total Assets (6) TA = Income bef. Ex. Items (123)-(Cash Flows from Operating Activites (308)-Ex. Items (124)) A = Total Assets (6) NI = Net Income (172) A= Total Assets (6) SPNI = Indicator variable equalling 1 if (Net Income (172)/Lagged Assets (6)) is between 0 and 0.01 SNNI = Indicator variable equalling 1 if (Net Income (172)/Lagged Assets (6)) is between 0 and NI = Net Income (172)/Lagged Assets (6) NEGDUM = Indicator variable equalling 1 if NI is less than 0 High levels of accruals indicate poor earnings quality High levels of change in accruals indicate poor earnings quality Distribution dierent from 0 signify poor earnings quality High ratios indicate poor earnings quality α3 < 0 implies high accounting quality 50

53 Timely Loss Recognition NIit Ait Variability of Earnings to Cash Flows σ(n Iit) σcf Oit Variability of Earnings σ( N Iit) Correlation between Accruals and Cash Flows ρ( N Iit, CF Oit) Persistence Xi,t = φ0,i + φ1,ixi,t 1 + εi,t NI = Net income (172) A = Total Assets (6) NI = Net Income (172)/ Lagged Assets (6) CFO = (Cash Flows from Operating Activites (308)-Ex. Items (124))/Lagged Assets (6) NI = Net Income (172)/ Lagged Assets (6) NI = Net Income (172)/ Lagged Assets (6) CFO = (Cash Flows from Operating Activites (308)-Ex. Items (124))/Lagged Assets (6) X = Net Income (172)/Common Shares Outstanding (25) High frequency of large losses indicate high earnings quality High ratios indicate earnings smoothing Low variability indicates smoothing Large negative correlations indicate smoothing Low values of φ0,i indicate impersistent earnings 51

54 Predictability Xi,t = φ0,i + φ1,ixi,t 1 + εi,t X = Net Income (172)/Common Shares Outstanding (25) High variance of the error term indicates poor earnings quality 52

55 6.3 Accounting Quality Metrics Below is a description of how I measure earnings quality using all 16 metrics. An overview is depicted in Table 4. Accruals Models The original Jones Model measured abnormal accruals in only one year and hypothesised income decreasing earnings management, thereby having an a priori expectation of the direction of the accruals. My research design is dierent on these two parameters. First, the interest lies in estimating abnormal accruals in a longer time period, and second, I have no hypothesis concerning the direction of the earnings management, and therefore abnormal accruals can take on both positive and negative values. I therefore alter the Jones Model in two ways to deal with these issues. To solve the rst issue, namely nding the abnormal accruals in the two time periods (before and after the restatement event, respectively), I estimate The Jones Model as xed-eects panel regressions in each of the 48 Fama-French (FF) industries 14. Under this approach, I assume that the unobserved heterogeneity inuencing the level of accruals is xed over time in each FF industry. Thus, only time-varying unobserved factors remain in the error term which allows us to examine how discretionary accruals for each rm in each of the 48 industries change over time. 15 To solve the second issue, I use the solutions proposed in the methodological paper by Hribar and Nichols (2007). They describe that discretionary accruals models are increasingly used with unsigned values of abnormal accruals to test for overall dierences in earnings quality. Using the absolute rather than the signed value of discretionary accruals obviously changes the distribution of residuals, since they are truncated at 0. In addition, Hribar and Nichols (2007) show that unsigned discretionary accruals have a dierent probability function than signed discretionary accruals. More specically, they show that the expected value of absolute discretionary accruals is an increasing function of the residual variance. Therefore, they suggest that the driver of the residual variance is controlled for in research designs. In line with this, I add the volatility of operating cash ows of each rm as an additional regressor, to control for operating volatility. In accordance with with prior research (e.g. Louis and White (2007)), I include lagged accruals in the regression to control for the mean reversion of 14 Available from Siccodes48.txt. 15 It is not possible to use rm-xed eects in the Jones Model. To see this, remember that xed eects place the unobserved eects, that are constant over time, in a constant variable capturing the unobserved eect. Since the unobserved eect is removed from the residual, and the measure of discretionary accruals is the residual, the approach will underestimate the level of abnormal accruals. 53

56 accruals. Consistent with for instance Francis et al. (2005), I estimate discretionary accruals directly as the residuals, rather than using a two-stage approach. I predict that restaters have higher levels of absolute discretionary accruals in the years before the restatement. Accruals Quality Models Working capital accruals are estimated using the approach of Dechow and Dichev (2002). I follow Francis et al. (2004) in estimating Equation 4 and 5 over rolling, rm-specic windows, ten years in the before sub-sample, and a minimum of three years in the after sub-sample. The panel regressions for both the original and the modied Dechow-Dichev Model are run with rm-xed eects and the standard deviation of residuals for each rm is the Dechow-Dichev measure of accruals quality. I predict that restaters have higher standard deviation of residuals than non-restaters in the years before the restatement. Accruals Models I estimate accruals directly as the dierence between cash ows and earnings. Since the reporting of cash ows was made mandatory with the FASB Statement No. 95 (eective from 1988), this method has by far been the most widely used. More specically, my measure of accruals stems from Hribar and Collins (2002), who nd accruals as earnings before extraordinary items, less operating cash ows from continuing operations. They suggest that this method contains less measurement error than the balance sheet approach used before. I test if the level and change in accruals, respectively, are signicantly dierent for the restaters and non-restaters, and before and after the last restatement event. In accordance with Dechow and Dichev (2002), the absolute magnitude of accruals is the variable of interest. I expect to nd higher levels of and yearly changes in accruals for restaters before the restatement event. Avoiding Earnings Decreases and Small Losses I follow Burgstahler and Dichev (1997) in comparing distributions of earnings but scale with lagged assets rather than market value of equity. I tabulate results for distributions of change in earnings. It is tested if the distributions of earnings for restaters and non-restaters, respectively, are signicantly dierent from each other. Table 7 and 8 show the median of change in earnings and the dierence between the two is tested with the Wilcoxon Two-Sample Test for dierences in distributions. When comparing the frequency of small losses and small prots I use the approach of Burgstahler et al. (2006). Restaters are predicted to have a higher median of change in earnings 54

57 and a smaller proportion of small losses to small prots, than non-restaters do. Timeliness The piecewise regression suggested by Basu (1997) is estimated as a panel regression with rm-xed eects in both groups and time periods. I follow the approach of Ball and Shivakumar (2005). Table 7 and 8 show the coecient on NEGDUM NI, which is expected to be more negative for non-restaters than restaters. I tabulate the number of large losses as the percentage of the total number of observations. Smoothness The variability of change in earnings and the variability of earnings to variability of cash ows are calculated on rm level and compared for the two groups. Restaters are expected to smooth more, i.e. have lower variance than non-restaters. Spearman correlations are calculated between accruals and cash ows as the last measure of earnings smoothing. The correlations are anticipated to be more negative for restaters. Persistence In accordance with Francis et al. (2004), a one-order autoregressive regression is estimated for each rm using maximum likelihood estimation, in the before and after period, respectively. The parameter estimates on φ are outputted and compared for restaters and non-restaters. I expect that non-restaters have more persistent earnings, i.e., φ closer to 1. Predictability The autoregressive regression is estimated as described above, and I compare the standard deviation of residuals for each rm. Higher variance of the error term is expected for restaters. 6.4 Descriptive Statistics Table 5 shows all the restatements from Audit Analytics (matched with unique GVKEY) before each restating rm is matched with a non-restating counterpart. This corresponds to 5,336 restatement events shared among 3,817 unique rms. Thus, each rm on average restated 1.4 times. It is clear that the number of restatements increased signicantly in the mid-2000s. This could be due to change in SEC's identication procedure, the downturn in the American economy in the beginning of the new millennium, and the Sarbanes-Oxley Act (Scholz, 2008). Table 6 presents a comparison of restating rms in my sample to all Compustat rms. It seems as if restating rms are overrepresented in the service 55

58 Table 5: Restatements divided in Years Restatement Year Number of Restatements Percent % % % % % % % % % % % Sum 5, % 56