SEARCH

SEARCH BY CITATION

Abstract

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Prior research
  5. 3. Research design and empirical results
  6. 4. Threats to internal validity
  7. 5. Comparing sample firms to the broad population
  8. 6. Conclusion
  9. References

A heated debate exists as to whether discontinuities in earnings distributions are indicative of earnings management. While many studies attribute discontinuities in earnings distributions to earnings management, other studies argue that earnings discontinuities are artifacts of sample selection and research design. Overall, there is limited direct evidence of a connection between earnings discontinuities and earnings management. In this study, we provide direct evidence linking earnings management to earnings discontinuities for a sample of firms that settle securities class action lawsuits and restate earnings from the alleged GAAP violation period. We compare the distribution of restated (“unmanaged”) earnings to originally reported (“managed”) earnings. We find that discontinuities are not present in the distribution of analyst forecast errors and earnings changes using unmanaged earnings but are present using managed earnings. The discontinuity in the earnings level distribution is attenuated, but not eliminated, on an unmanaged basis. These shifts among our sample of firms are caused by earnings management and cannot be explained by sample selection or research design issues. Our findings are important because many studies use earnings discontinuities as a proxy for intentional earnings manipulations and we provide the first direct evidence of a link between these two phenomena.


1. Introduction

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Prior research
  5. 3. Research design and empirical results
  6. 4. Threats to internal validity
  7. 5. Comparing sample firms to the broad population
  8. 6. Conclusion
  9. References

This paper explores the impact of earnings management on the discontinuity in earnings distributions near important benchmarks for a sample of firms where the amount of earnings management is measurable. Numerous studies (e.g., Burgstahler and Dichev 1997; Degeorge, Patel, and Zeckhauser 1999) attribute earnings discontinuities to earnings management. Consistent with this assertion, many executives indicate that they would manage earnings within generally accepted accounting principles (GAAP) to achieve earnings benchmarks (Graham, Harvey, and Rajgopal 2005). However, recent studies suggest that the discontinuities may be driven by other factors including scaling, sample selection, and the effects of special items and taxes (e.g., Durtschi and Easton 2005, 2009; Beaver, McNichols, and Nelson 2007).

In order to explore the impact of earnings management on the discontinuity in earnings distributions, one must be able to determine the amount each firm manages earnings and then compare the distribution of unmanaged earnings to the distribution of managed earnings. Unfortunately, earnings management is largely unobservable, making it impossible to directly measure its effect on earnings discontinuities in the overall population of firms. Thus, the only way to provide direct evidence of a relation between earnings management and earnings discontinuities is to investigate the matter among a sample of firms where earnings management is detected ex post and thus measurable. Our study takes precisely this approach.

We utilize a sample of settled accounting-related securities class action lawsuits where firms restated earnings from the alleged GAAP violation period. This sample provides a strong basis for inferring earnings management for two reasons. First, all firms in the sample restated earnings from the class period (i.e., the time period over which stock price was inflated due to the alleged misstatement), indicating earnings were misstated and allowing measurement of the misstatement. Second, during our sample period, U.S. securities laws were strongly biased in favor of dismissing nonmeritorious securities class action suits. To survive a motion to dismiss by the defendant firm in a Rule 10b-5 case, plaintiffs must allege with particularity both the alleged GAAP violation and accompanying facts giving rise to a strong inference of fraud. Thus, the settlement of litigation implies that the accounting misstatement was intentional.1

Our empirical analysis is straightforward. We plot the distributions of restated (i.e., unmanaged) and originally reported (i.e., managed) earnings and test whether there is a discontinuity in the respective distributions near earnings benchmarks (i.e., analysts’ consensus forecast, prior year earnings, and the zero-profit benchmark). Two features of our research design are worth emphasizing. First, because our sample of firms restated earnings related to settled securities litigation, we observe both the presence and extent of earnings management. This eliminates the need to rely on discretionary accrual models or other approaches to identify and measure earnings management. Second, our tests allow us to isolate the effect of earnings management on discontinuities because we hold all factors other than the restatement constant. The only variable that changes the shapes of the distributions we examine is earnings management.

We find no evidence of a discontinuity at the analyst forecast and prior-year earnings benchmarks when earnings are plotted on an unmanaged basis. However, we find a significant discontinuity at these benchmarks when earnings are plotted on a post-managed basis. For the zero-profit benchmark, we find similar results when earnings are scaled by contemporaneous market value of equity. However, when earnings are scaled by total assets or the post-class period market value of equity, we find evidence of discontinuities at the zero-profit benchmark even on an unmanaged basis. These discontinuities, however, are smaller than those of the managed distributions, consistent with earnings management contributing to these discontinuities. Overall, our results suggest that the discontinuities in the earnings distributions are largely driven by earnings management among our sample of firms.

Our sample selection criteria only capture instances where earnings management is detected ex post. This allows us to identify and measure earnings management, which is required to explore the effect of earnings management on the distribution of earnings. However, we do not capture all instances of earnings management and our sample likely contains the more extreme cases of earnings management. Certainly, our sample does not likely capture within GAAP earnings management or real earnings management. Therefore, we cannot quantify what percentage of the discontinuities observed in broad samples (e.g., the COMPUSTAT population) is attributable to earnings management versus competing explanations. Nevertheless, in an effort to provide evidence on how our results relate to the overall population of firms, we do compare the earnings distributions for our sample firms to those of the overall population.

If earnings management plays a role in the shape of overall earnings distributions, one would expect a subsample of known earnings managers to exhibit greater discontinuities and benchmark beating tendencies than the broad population, where the incidence and extent of earnings management is likely lower on average. On the other hand, if earnings discontinuities are driven by factors other than earnings management, one would expect earnings distributions for known earnings managers to look the same as the overall population. We find that the discontinuities in our sample are larger than those observed among the broad population of firms. In addition, the proportion of firms that just meet or beat an earnings target is higher in our sample of known earnings managers relative to the broad population. While this evidence is indirect and thus only suggestive, it is consistent with the notion that earnings management contributes at least partially to earnings discontinuities observed in broad samples.

Our paper adds important evidence to the literature. To our knowledge, this is the first study to directly measure the magnitude of earnings management and examine the impact of ex post revealed earnings management on the shapes of earnings distributions. Our approach allows us to construct a distribution of unmanaged earnings, the absence of which has hindered prior work in this area (see Kerstein and Rai 2007). Our results are also important because prior studies using earnings management proxies, either specific accruals or general abnormal accrual measures, provide indirect and mixed evidence of a link between earnings management and discontinuities in earnings distributions (see, e.g., Beatty, Ke, and Petroni 2002; Dechow, Richardson, and Tuna 2003; Beaver, McNichols, and Nelson 2003; Ayers, Jiang, and Yeung 2006). In fact, Durtschi and Easton (2009: 1) note that prior evidence supporting a link between earnings management and earnings discontinuities is “sparse, perhaps even non-existent”. Our findings among a sample of firms that were caught managing earnings indicate that actual instances of earnings management do create sizable discontinuities in earnings distributions.

We review the prior literature in section 2 and outline our research design and empirical results in section 3. Section 4 provides additional tests that examine threats to the internal validity of our study, while section 5 compares our sample firms to the broader population. Section 6 concludes.

2. Prior research

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Prior research
  5. 3. Research design and empirical results
  6. 4. Threats to internal validity
  7. 5. Comparing sample firms to the broad population
  8. 6. Conclusion
  9. References

Discontinuities in earnings distributions, earnings management, and other explanations

Many studies find discontinuities around earning benchmarks, including the profit/loss threshold (e.g., Hayn 1995; Burgstahler and Dichev 1997); prior year earnings (e.g., Burgstahler and Dichev 1997); and analysts’ consensus forecast (e.g., Degeorge et al. 1999). The most common explanation for these discontinuities is earnings management, and the discontinuities are often cited as evidence of earnings management.

Recent studies, however, point out several plausible alternative explanations for the discontinuities near earnings benchmarks, including scaling (Degeorge et al. 1999; Dechow et al. 2003; Durtschi and Easton 2005, 2009), sample selection (Durtschi and Easton 2005, 2009), and the asymmetric effect of taxes and special items on profit and loss firms (Beaver et al. 2007). Scaling by market value of equity could induce a discontinuity because firms on either side of the relevant benchmark (e.g., profit versus loss firms) have different market valuations. The sample selection criteria in certain studies (e.g., requiring lagged market value of equity or analyst coverage) could also lead to spurious discontinuities if they result in the overinclusion or overexclusion of firms on either side of the relevant benchmark (Durtschi and Easton 2005, 2009). Finally, Beaver et al. (2007) contend that the discontinuity at the profit/loss threshold arises at least in part from special items and the asymmetric accounting treatments of taxes.2

Several studies attempt to isolate the earnings management story from these alternative explanations. The three primary methods of attempting to determine whether earnings management contributes to the earnings discontinuities utilize discretionary specific accruals, discretionary total accrual models, or properties of fiscal year reporting patterns. Consistent with earnings management, Beatty et al. (2002) compare private and public banks and find that public banks are more likely to avoid small earnings declines and to use discretionary bank reserves (e.g., the loan loss reserve) to do so. Similarly, Beaver et al. (2003) utilize ex post revisions in property-casualty insurance company loss reserves and find that these reserves appear to be understated for firms that report small profits relative to small losses. On the other hand, prior research utilizing discretionary accrual models provides relatively little consistent evidence indicating that firms that meet or barely beat benchmarks actually managed earnings.3 One criticism of both discretionary-specific accruals and discretionary total accrual models is that the intentionally manipulated portion of accruals is usually unobservable, even ex post. For example, Petroni, Ryan, and Wahlen (2000) document that revisions to property and casualty reserve accruals have both discretionary and nondiscretionary components. Parsing out the intentionally managed component from the nonmanaged component is critical in determining whether the proxy appropriately captures earnings management.4 This is of particular concern as performance may be correlated with both the earnings management measures and meeting or beating the earnings benchmarks (Kothari, Leone, and Wasley 2005). A key advantage of our study is that we can identify and measure earnings management in our sample without relying on model estimates or strict assumptions about accrual misstatements (e.g., all accrual revisions represent earnings management).

With respect to fiscal year reporting patterns, Jacob and Jorgensen (2007) provide evidence that the discontinuity in the annual earnings distribution relative to the zero-profit benchmark is present when earnings are measured over the annual fiscal period, but is not present when quarterly data are aggregated in four quarter periods that do not correspond to firm fiscal years. Similarly, Kerstein and Rai (2007) find that firms that are very close to the zero-profit benchmark through the third quarter year-to-date attain small annual profits at an unusually high rate. Both studies conclude that earnings management is responsible for the disproportionate number of firms that just meet the annual zero-profit benchmark.

These studies highlight another potential issue with the interpretation of discontinuities due to properties of the firm fiscal year (Durtschi and Easton 2009). Firms set earning targets with respect to the fiscal year, so the achievement of earnings benchmarks could relate to real operating actions taken by firms to achieve goals (see, e.g., Dechow et al. 2003; Roychowdhury 2006). In addition, the integral approach to interim financial reporting aligns quarters within the same fiscal year more than quarters from different fiscal years, leading to questions regarding studies relying on the fiscal year reporting period (Durtschi and Easton 2009).

Thus, the evidence linking earnings management to earnings discontinuities is mixed and there is still significant disagreement in the literature regarding whether discontinuities are evidence of earnings management.5 This limited evidence appears to buttress the claims that research design issues and real operational actions are primarily responsible for the patterns. Due to the inability of researchers to directly observe earnings management, it is impossible to gather large sample evidence on the extent of earnings management for all firms. Thus, even a small sample where earnings management is observable is of interest with respect to the debate regarding the relation between earnings management and earnings discontinuities.

Earnings management and securities class action suits

While there is no uniformly accepted definition of earnings management, a common characterization is “when managers use judgment in financial reporting and in structuring transactions to alter financial reports to either mislead some stakeholders about the underlying economic performance of the company or to influence contractual outcomes that depend on reported accounting numbers” (Healy and Wahlen 1999: 368). The key elements of this definition are that financial reports intentionally mask true performance to mislead another party. Based on this definition, a natural method to classify firms that managed earnings (i.e., intentionally misstated accounting information with the intent to mislead) is by utilizing securities class action suits.

Securities class action suits are primarily filed under Rule 10b-5 (promulgated by the Securities and Exchange Commission [SEC] pursuant to Section 10(b) of the Securities Exchange Act of 1934). Rule 10b-5 is an anti-fraud remedy allowing private recovery in connection with the purchases or sales of securities. The elements of a Rule 10b-5 violation are “(1) a misstatement or omission of (2) a material fact (3) made with intent (4) that the plaintiff justifiably relied on (5) causing injury in connection with the purchase or sale of a security” (Skinner 1994: 41). Thus, in an accounting context, a Rule 10b-5 violation includes the definition of earnings management cited above.6

Most securities class action suits evolve in a similar fashion. They almost always begin after a significant “bad news” announcement by a firm. The most common examples include restatements due to an “irregularity”, sometimes accompanied by executive dismissal and an internal or regulatory investigation, a substantial write-down of inventory or accounts receivable, or a whistleblowing report by a sell-side analyst or the financial press. Depending on the availability of material facts, plaintiffs (purchasers of securities during the time of the alleged fraud, known as the class period) usually file suit soon thereafter (within a few days to a few months). The plaintiffs identify alleged GAAP misstatements in their complaint and the period of time over which these misstatements led to inflated stock prices. The class period usually ends on the “bad news” announcement day described above.

Securities class action suits almost never proceed to trial (Black, Cheffins, and Klausner 2006). Defendant firms typically file a motion to dismiss shortly after the lawsuit is filed. Pursuant to the Private Securities Litigation Reform Act of 1995 (PSLRA), the discovery process cannot proceed until the motion to dismiss is resolved. Thus, the key procedural item for these suits is the applicable legal standard for the motion to dismiss, particularly with respect to the intent (scienter) requirement.7 As noted, plaintiffs in suits alleging accounting-related violations of Rule 10b-5 must allege a GAAP violation accompanied by scienter. Specifically, the facts alleged in the complaint must give rise to a strong inference of fraud. The Supreme Court defines a strong inference of fraud as “more than merely plausible or reasonable — it must be cogent and at least as compelling as any opposing inference of non-fraudulent intent”.8

Based on this standard, it is very unlikely that innocent, unintentional GAAP violations (i.e., those that do not constitute earnings management) will survive the motion to dismiss in a securities class action suit. In fact, the court in Reiger v. Price Waterhouse Coopers, LLP, 117 F. Supp. 2d 1003, 1010–1011 (S.D. Cal. 2000) stated that “Violations of GAAP or GAAS, standing alone, do not satisfy the particularity or strong inference requirements of the [PSLRA].”Hennes, Leone, and Miller (2008) provide empirical evidence consistent with this legal standard. They find that firms that issue restatements likely due to unintentional misstatements (errors) are rarely sued (one out of 83 cases in their sample). Alternatively, they find that firms that issue restatements due to intentional misstatements (irregularities) are sued often (84 of 105 cases in their sample).

In addition, it must be stressed that few of the firms in our sample actually admit accounting fraud (settlements are generally structured so as not to admit fraud) and we do not claim that they committed fraud.9 Rather, the settlement of the securities class action suit shows that a court found a “strong inference” of fraud. Thus, the settlement combined with a restatement strongly implies that at least earnings management occurred through “a purposeful intervention in the external financial reporting process” (Schipper 1989: 92), but it is generally not possible to determine whether fraud occurred in a particular case.10

3. Research design and empirical results

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Prior research
  5. 3. Research design and empirical results
  6. 4. Threats to internal validity
  7. 5. Comparing sample firms to the broad population
  8. 6. Conclusion
  9. References

Sample selection

This study utilizes the Securities Class Action Services (SCAS) database of historical class action data from RiskMetrics Group. The SCAS database includes all securities fraud class action lawsuits filed since the implementation of the PSLRA in December 1995. We limit our sample period to the timeframe in which the PSLRA was effective due to both data availability and because the PSLRA provides a powerful screen against frivolous lawsuits that may not be indicative of earnings management (see Choi, Nelson, and Pritchard 2009). To allow time for suits to be resolved, we end the sample with suits filed in 2005 that are dismissed or settled by March 2009.

To ensure that the sample is composed of firms that managed earnings, we exclude dismissed lawsuits and suits alleging only nonaccounting violations of securities laws. Firms that legally engage in “real” earnings management (e.g., Roychowdhury 2006) or take operational actions to meet earnings benchmarks would not be included because such suits, if filed, would contain no inference of fraud. However, firms that inflate earnings or misclassify items to deceive investors (e.g., McVay 2006) would be included.

We examine firms that restated accounting information from fiscal quarters contained in the class period, the time frame over which investors were allegedly deceived by the firms’ financial statements. Thus, we examine lawsuits that (1) settle and allege fraudulent GAAP violations (none of the suits proceeded to trial) and (2) involve restatements by firms for periods during which the alleged accounting fraud occurred.

The restatement criterion is particularly important. First, it allows us to construct unmanaged earnings distributions based on the restated financial data. Second, it represents an explicit admission by the firms that their financial statements were incorrect during the alleged fraud period. Thus, these restrictions ensure that firms included in our sample are highly likely to have committed earnings management and allow us to identify the amount of earnings management (as discussed in the “Earnings Management and Securities Class Action Suits” section). Nonsettled lawsuits do not provide an inference of earnings management (even when combined with a restatement) because the misstatement could be an error. Similarly, a lawsuit without a restatement could be settled on grounds outside of the accounting allegations because suits commonly allege both accounting and other nonaccounting violations (e.g., voluntary disclosure that was misleading). Thus, we need settled accounting-related suits with restatements to both infer and measure earnings management.

Panel A of Table 1 presents our sample selection in more detail. The primary sample begins with a comprehensive data set of all settled accounting-related securities fraud class action lawsuits filed during the period 1996–2005. We exclude suits that do not involve stockholders as plaintiffs and restrict our sample to firms with nonmissing Center for Research in Security Prices (CRSP) and COMPUSTAT data (market value of equity and income before extraordinary items). Finally, we exclude lawsuits that do not involve a restatement for fiscal quarters contained in the class period.11 Our final sample contains 329 lawsuits and 1,284 firm-quarters with originally-reported and restated earnings data. Panel B of Table 1 provides a breakdown of lawsuits by industry.12

Table 1. Sample
Panel A: Sample screens
Class action lawsuits filed in Federal Court during the period 1996–2005 involving alleged GAAP violations settled by defendants737
Less: Cases where plaintiffs do not include common stockholders(27)
 Cases where data is not available on CRSP or COMPUSTAT(151)
 Firms that did not restate quarterly earnings(230)
Final sample329
Total firm-quarters with restated and originally reported data1,284
Panel B: Industry composition
IndustryLawsuitsIndustryLawsuits
  1. Notes:

  2. We obtain lawsuit information from RiskMetrics and data on quarterly restatements from the COMPUSTAT Point-in-Time file. Industry definitions are based upon SIC codes using the definitions on Kenneth French’s website.

Aircraft1Machinery5
Autos3Measuring Equipment4
Banks16Medical Equipment3
Building Materials2Metals1
Business Services78Mining1
Clothing6Oil & Gas2
Computers36Other3
Construction3Personal Services10
Consumer Goods2Recreation3
Drugs11Restaurants4
Electrical Equipment4Retail17
Electronic Equipment24Rubber3
Entertainment5Steel2
Fabricated Products3Telecommunications12
Finance6Textiles2
Food5Transportation4
Healthcare11Utilities11
Insurance15Wholesale11
  Total329

Identifying and measuring earnings management

We use the COMPUSTAT Point-in-Time database to identify and measure managed (originally reported) earnings. COMPUSTAT has a policy of overwriting historical quarterly data with restated quarterly data in the traditional industrial files used by academic researchers. Thus, the COMPUSTAT quarterly files do not maintain the original quarterly data publicly reported by firms in their initial SEC filings (e.g., Jegadeesh and Livnat 2006). If companies restate quarterly data to correct errors or irregularities or to reflect changes in accounting principles, merger activity, or discontinued operations, COMPUSTAT overwrites the historical data with the restated data. This convention of overwriting quarterly data differs from the COMPUSTAT practice with respect to annual data, where the original data is generally retained and restated data is provided in a different data field, explicitly labeled as “restated”.

Prior to the availability of the Point-in-Time database, it was difficult for researchers to track the evolution of originally reported quarterly earnings to restated earnings in any given vintage of the quarterly industrial file. The Point-in-Time database allows researchers to track these changes. For each firm-quarter of financial data, the Point-in-Time database tracks changes in selected accounting data (such as revenue and earnings) each calendar month up to twenty quarters after the data was originally disclosed (usually in an SEC filing). These changes occur due to restatements of quarterly earnings by firms in public filings.13

We measure the restatement amount as the difference between the most recent value for earnings before extraordinary items in the Point-in-Time database and the initial value reported in the Point-in-Time database, which is obtained from the initial SEC filing. We use this approach because some firms involved in accounting scandals restate the same fiscal period more than once and the most recent value of earnings captures restated earnings after “the dust has settled”. We focus on fiscal quarters contained in the class period to ensure we capture restatements related to earnings management.14

Two additional features of our approach merit discussion. First, the Point-in-Time database does not track annual restatements, so we do not identify restatements involving only annual figures. Thus, our sample may not include all restatements by our sample firms. However, most firms that publicly announced the need to restate are captured in our sample. In an untabulated analysis, we compared our sample to a sample collected and used by Hennes et al. 2008 of all firms announcing a restatement in an 8-K filing during the period 1997–2006.15 We find that 85 percent of firms announcing a restatement in the Hennes et al. 2008 sample are classified as restating quarterly earnings using our methodology. We investigated a handful of announcing firms not captured as restating using our methodology. In many cases, despite their restatement announcement, these firms stopped filing reports for several fiscal periods, never filed an amended 10-K or 10-Q upon the resumption of filing, and did not appear to restate prior quarterly earnings in subsequent filings. Thus, while our approach may exclude some restatements, we capture the overwhelming majority of cases where firms restate financial data in SEC filings.

Second, though we focus only on restatements during the class period, it is possible that some restatements in our sample are not attributable to earnings management. For example, when a merger occurs, firms will sometimes restate prior quarterly earnings on a pro forma basis (i.e., as if the firms had been merged all along) to facilitate comparability across quarterly periods. Discontinued operations can trigger a similar treatment. This type of change in the Point-in-Time database is not necessarily indicative of earnings management. While COMPUSTAT does maintain footnote codes in the Point-in-Time database to track the incidence of mergers and discontinued operations each fiscal period, it is not clear from these footnotes whether the restatement is attributable solely to the mechanical effects described above or to other effects such as the correction of earnings management. For this reason, and because events like mergers are sometimes cited by plaintiffs in their complaints as motives to manage earnings, we include all restatements during the class period in our primary analysis to follow.16

To validate our measure of earnings management, we hand-collected a subsample of 50 firms that had no indication of M&A activity per COMPUSTAT and examine the reason for their restatement, and find that 46 (92 percent) of these cases involved restatements in Point-in-Time associated with the allegations of fraud in the lawsuit. The other four (8 percent) cases involved either M&A restatements not captured by the COMPUSTAT footnote or we could not find the restatement in SEC filings. Importantly, for these four cases, the restatements were so minor they did not even affect the bin assignment in the earnings levels distribution. We conclude that (1) our methodology using COMPUSTAT Point-in-Time is a reasonable and reliable way to capture restatements related to securities fraud with machine readable data, and (2) that any data errors that arise do not have a material effect on our inferences.

Our approach yields several advantages which alleviate concerns from the prior literature (e.g., Durtschi and Easton 2005, 2009; Beaver et al. 2007). First, the only factor that differs between the earnings distributions we examine is the level of earnings (restated versus originally reported). Our comparison of restated and originally reported earnings distribution contains exactly the same set of firms, so sample selection bias (Durtschi and Easton 2005, 2009) or the asymmetric effects of accounting items (Beaver et al. 2007) cannot explain differences between these earnings distributions. However, we do acknowledge that we capture a unique set of firms that are accused of extreme earnings management. Therefore, without a large, random sample, it is not feasible to extrapolate our results and precisely quantify how much of the discontinuities observed in broad samples (e.g., the COMPUSTAT population) are attributable to earnings management versus competing explanations.

Descriptive statistics

Table 2 provides descriptive statistics for our sample of sued firms. Announcement Return is the cumulative, market-adjusted return for the seven days centered on the end of the class period. Typically, this is the date that the earnings restatement is announced. As presented in Table 2, the mean cumulative market-adjusted return surrounding the class period end date is −22 percent, which is significantly different from zero, both statistically (untabulated) and economically. Further, the mean cumulative market-adjusted return surrounding the class period end date is similar to that observed in the prior studies (e.g., Hennes et al. 2008; Johnson et al. 2007; Karpoff, Lee, and Martin 2008). There is substantial variation in size across sued firms during the class period. The 10th percentile of market value of equity across firm-quarters is roughly $70 million, while the 90th percentile is roughly $12 billion. In addition, the mean (median) market value of equity for firms in our sample is $6.37 (0.66) billion, suggesting that we have several large firms causing skewness in our size variable. Restatement is restated earnings before extraordinary items less originally reported earnings before extraordinary items. The mean (median) Restatement per firm-quarter is −$4.10 (−$0.71) million, or roughly −$16.5 (−$2.8) million on an annual basis. In addition, the mean (median) restatements amount as a percentage of the market value of equity is 0.40 percent (0.10 percent), or 1.6 percent (0.4 percent) on an annual basis. Finally, on a per share basis, the mean (median) restatement per share is a negative five (two) cents per share, or negative 20 (8) cents per share on an annual basis.

Table 2. Descriptive statistics
VariableMean5th10th25th50th75th90th95th
  1. Notes:

  2. Announcement return is the cumulative, market-adjusted return for the seven days centered on the end of the class period specified in the plaintiff complaint. MVE is the market value of equity at the end of each quarter from the COMPUSTAT quarterly fundamental file. Restatements are measured as the latest restated value of quarterly earnings before extraordinary items less the corresponding amount originally reported by the firm. Shares outstanding are obtained from the COMPUSTAT quarterly fundamental file.

Announcement return−0.221−0.713−0.541−0.338−0.194−0.0590.0480.131
MVE (in millions of dollars)6,374.88041.00469.854191.805655.0382,650.61012,059.10025,253.680
Restatement (in millions of dollars)−4.147−43.276−18.320−4.858−0.7070.3905.22628.600
Restatement scaled by MVE−0.004−0.036−0.020−0.006−0.0010.0000.0060.022
Restatement scaled by shares outstanding−0.049−0.448−0.263−0.100−0.0200.0100.0870.251

Table 2 also indicates that not all restatements for quarters in the class period decrease earnings, with roughly 30 percent of restatements actually increasing earnings. These restatements arise from two main sources. First, when firms shift income between periods (through premature revenue recognition or “cookie jar” accounting), the restatement will by definition increase income in certain quarters and decrease income in others.17 Second, as discussed above, some restatements may be attributable to events such as mergers or discontinued operations. Nevertheless, the distribution of restatements is strongly nonsymmetric and skewed to the left with the 25th percentile of the distribution of restatement per share being roughly ten times the magnitude of the 75th percentile.

Primary results

To analyze the impact of earnings management on the distribution of earnings, we plot the same sample of firms using both restated and originally reported earnings. Specifically, we plot the empirical distribution of restated and originally reported analyst forecast errors, seasonal earnings changes, and earnings levels. We assign observations to appropriate “bins” in the earnings distributions using conventions from prior literature (e.g. Burgstahler and Dichev 1997; Degeorge et al. 1999).

For analyst forecast errors, we assign firms to one cent bins where bin zero corresponds to firms that exactly met the last consensus analyst forecast in I/B/E/S measured at least three days before the earnings announcement date. We calculate restated earnings by subtracting the income statement correction in the Point-in-Time database from the actual I/B/E/S earnings number. Specifically, the correction is equal to the change in income before extraordinary items, net of any changes in special items, scaled by shares used to calculate the appropriate EPS metric (diluted or basic). The share figures are taken from COMPUSTAT, while the diluted or basic indicator is taken from I/B/E/S. Results are inferentially similar if we make no adjustment for special items or simply use shares outstanding from COMPUSTAT as the scaling variable. We use the actual I/B/E/S earnings number as the originally reported number because it is not adjusted for earnings restatements. In addition, consistent with a vast prior literature on analyst forecasts, using I/B/E/S actual earnings helps maintains consistency with the earnings definition implicit in the I/B/E/S consensus forecast.

For earnings levels, we scale restated and originally reported income before extraordinary items by the market value of equity at the end of the quarter, and assign observations to bins in 0.005 increments. For restated and originally reported earnings changes, we scale seasonally differenced changes in income before extraordinary items by market value of equity at the end of the quarter, and assign observations to bins in increments of 0.0025. To calculate restated earnings changes, we use both restated current quarterly earnings and restated earnings from the same quarter of the prior year, if prior year earnings were restated.

Figure 1 plots the distribution of analyst forecast errors (in cents per share) during the class period for the restated and originally reported earnings distributions. The top left graph plots the restated (unmanaged) distribution, while the top right graph plots the originally reported (managed) distribution. The bottom graph combines the distributions so that changes in each bin are more apparent.

image

Figure 1.  Analyst forecast errors

Download figure to PowerPoint

The asymmetry noted in prior literature is drastically attenuated for the restated sample. Using originally reported forecast errors, there is an asymmetry about zero. In addition, consistent with Degeorge et al. 1999, there is not a noticeable “dip” at bin negative one.18 Roughly 25 percent of the sample falls in bin zero and 15 percent of the sample falls in bin one. In contrast, only 7 percent of the sample reports earnings that miss analyst forecasts by a penny (bin negative one).19 Thus, firms appear much more likely to meet or just beat analyst forecasts as opposed to miss forecasts by a penny. In addition, as we move from the restated to the originally reported distribution, we observe significant increases in the proportion of observations in bins zero through three, and a decrease in observations in bins negative three through ten.20 In untabulated results, the proportion of observations that meet or beat analyst forecasts increase from roughly 40 percent in the restated distribution to roughly 70 percent in the originally reported distribution (< 0.01).

We test the statistical significance of the observed shifts using two discontinuity tests. First, we measure the discontinuity in the distribution of forecast errors using the test devised by Burgstahler and Dichev 1997 (BD hereafter). The BD test measures the extent to which the number of observations in bin negative one (just below the benchmark) appears too low given the number of observations in neighboring bins. The BD test assumes the expected number of observations in any given bin is equal to the average number of observations in the two adjoining bins. No parametric assumptions about the underlying distributions are imposed. Following BD, we calculate a test statistic equal to the difference between the actual and expected number of observations, divided by the estimated standard deviation of this difference. Under the null of a smooth and continuous distribution, this test statistic is distributed asymptotically standard normal due to the Central Limit Theorem. Negative values for the test statistic imply the actual number of observations is too low, indicative of a discontinuity.

The second test comes from Degeorge et al. 1999 (DPZ hereafter), who suggest a specific test when the suspected discontinuity coincides with the mode of the distribution (here, bin zero). DPZ reason that if firms manage earnings to meet the zero forecast error threshold, then there should be an asymmetry in the magnitude of the distribution slope around bin zero. That is, the change in the proportion of observations from bin negative one to zero should be larger in magnitude than the change in proportion from bin zero to bin one. Likewise, this magnitude difference should be unusual relative to other symmetric comparisons across the distribution (e.g., the slope from bin negative two to negative one compared to the slope from bin one to two). Thus, the DPZ test statistic is simply the difference in absolute slope magnitudes about bin zero, divided by the sample standard deviation of this difference across other symmetric bin ranges (we use −11 to 11, similar to DPZ). Again, negative values indicate a discontinuity. The precise distribution of this test statistic is unclear because it is based on only a handful of slope observations and thus the Central Limit Theorem does not apply. Using bootstrap simulations, DPZ find that a test statistic exceeding two in magnitude is quite infrequent under the null of perfect symmetry and use this cutoff to assess statistical significance.

Confirming the visual evidence in Figure 1, the discontinuities are virtually nonexistent in the restated distribution, as both the magnitude and the statistical significance of the discontinuity at zero are quite small. However, both the BD test and the DPZ test indicate strong discontinuities in the distribution of originally reported forecast errors in panel A of Table 3. In addition, panel A of Table 3 tabulates changes in the proportion of observations in bins at and just above zero from the restated to the originally reported distribution. The dramatic changes in the proportion of observations observed in Figure 1 are significant at the 1 percent level using a chi-square test of proportions.

Table 3. The effect of restatements on discontinuities and benchmark beating
Panel A: Analyst forecast errors
 Discontinuity tests Proportion tests
BD (1997) test DPZ (1999) test  Bin 0Bins 0 to 3
Size of discontinuityTest statisticSize of discontinuityTest statistic
Restated0.0010.06−0.0190.06Restated0.0740.210
Originally reported−0.066−6.26−0.078−6.68Originally reported0.2500.562
N 1,094   Difference0.176***0.352***
Panel B: Earnings changes
 Discontinuity tests Proportion tests
BD (1997) test DPZ (1999) test  Bin 0Bins 0 to 3
Size of discontinuityTest statisticSize of discontinuityTest statistic
Restated0.0050.48−0.0040.58Restated0.1130.301
Originally reported−0.016−1.78−0.068−5.07Originally reported0.1420.402
N 1,281   Difference0.029**0.101***
Panel C: Earnings levels
 Discontinuity tests Proportion tests
BD (1997) test  Bin 0Bins 0 to 3
Size of discontinuityTest statistic
  1. Notes:

  2. Panel A contains discontinuity tests and differences in proportion tests for the analyst forecast error distribution using originally reported and restated data. Forecast errors are calculated as actual earnings (originally reported or adjusted for restatement) from I/B/E/S less the last consensus (median) forecast at least three days prior to the announcement of earnings. Observations are placed into bins with intervals of one cent per share. See Figure 1. The BD (1997) test comes from Burgstahler and Dichev 1997. The size of discontinuity at bin negative one and zero is equal to the actual proportion of observations in bin negative one less the expected proportion. The expected proportion is equal to the average proportions of the two adjacent bins (bin zero and negative two). This deviation from the expected value is then divided by its standard error to yield a test statistic, which is distributed asymptotically standard normal. The DPZ (1999) test comes from Degeorge et al. 1999 and is designed for situations where the suspected discontinuity involves the mode of the distribution (bin zero in this case). The size of discontinuity at bin negative one and zero is equal to the difference in the magnitude of the change in proportions of observations from bin zero to negative one versus bin zero to positive one, less the sample average of this difference across symmetric bin ranges from −11 to 11. See the text for more details. This deviation from the sample mean is then divided by its sample standard deviation to yield a test statistic, which follows a t distribution if the underlying distribution of earnings is Gaussian. Based upon simulation analysis, DPZ use a cutoff of 2.00 to assess significance. Differences in the proportions of observations in various bins across the distribution of originally reported versus restated earnings are tested using a chi-square test of proportions. Panels B and C present similar analyses for the distributions of earnings changes and earnings levels, respectively. Earnings changes are originally reported or restated earnings before extraordinary items less originally reported or restated earnings before extraordinary items from the same quarter of the prior year, deflated by MVE from COMPUSTAT. Observations are placed into bins with intervals of 0.0025. See Figure 2. Earnings levels are originally reported or restated earnings before extraordinary deflated by MVE from COMPUSTAT. Observations are placed into bins with intervals of 0.005. See Figure 3.

Restated−0.013−1.37Restated0.1440.497
Originally reported−0.038−4.87Originally reported0.1260.559
N 1,284 Difference−0.0180.062***

Figure 2 displays similar patterns for the earnings changes distribution. The restated distribution appears much more symmetrically distributed in the nearby bins around zero, and the number of observations in bin negative one in the restated distribution appears more in line with what one would expect given the number of observations in bins negative two and zero. The combined graph indicates increases in observations in bins zero through three and decreases in bins negative one, two and four from the restated to the originally reported distribution. The statistical tests in panel B of Table 3 confirm these visual findings. Both the BD and DPZ tests indicate significant changes in the discontinuities at zero. In addition, the chi-square proportion tests confirm that the change in the proportion of observations at and just above zero is statistically significant. The decreases in observations from the restated to the originally reported distribution just below bin zero are statistically significant (< 0.01, untabulated).

image

Figure 2.  Earnings changes

Download figure to PowerPoint

In Figure 3, we plot the distribution of restated and originally reported earnings levels. The discontinuity at bin zero appears attenuated in the restated distribution, though not virtually eliminated as observed in Figures 1 and 2 with analyst forecast errors and earnings changes. The discontinuity is more noticeable in the originally reported distribution. The combined graph indicates the proportion of observations in bin zero is slightly higher in the restated distribution, though this difference is not statistically significant (see Table 3) and is not as large in magnitude as the decrease from the restated to the originally-reported earnings in bin negative one. There is also a marked increase in the number of observations in bins one through three from the restated to the originally reported distribution. In short, the distribution of earnings shifts to the right when we move from restated (unmanaged) earnings to originally reported (managed) earnings. This shift appears to be largely from bin negative one to bins one through three.

image

Figure 3.  Earnings levels

Download figure to PowerPoint

In panel C of Table 3, the difference in the size of the discontinuity in the earnings level distribution is both economically and statistically significant. The discontinuity in the restated distribution is roughly one-third the size of the discontinuity in the originally reported distribution, and the test statistic is insignificant.21 In addition, the proportion tests indicate that the decrease in observations in bin zero in the originally reported distribution is insignificant, though the total increase in bins zero through three observed in Figure 3 is significant (< 0.01). In untabulated tests, we also find a significant (< 0.01) decrease in the number of observations in bin negative one when moving from the restated to the originally-reported distribution.

In sum, Figures 1 through 3 and Table 3 provide strong evidence that restated (unmanaged) earnings among sued firms for quarters during the class period are associated with dramatically smaller discontinuities in the distribution of analyst forecast errors, earnings changes, and earnings levels compared to originally-reported (managed) earnings. This evidence indicates that earnings management plays a significant role in the discontinuities observed in the earnings distributions of our sample firms.

In the next two sections, we conduct a series of additional tests to complement our main findings. Section 4 considers threats to internal validity (i.e., concerns that our sample findings are not really due to earnings management). Section 5 considers the more general implications of our findings and compares the shapes of the earnings distributions for our sample firms to the broad COMPUSTAT population.

4. Threats to internal validity

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Prior research
  5. 3. Research design and empirical results
  6. 4. Threats to internal validity
  7. 5. Comparing sample firms to the broad population
  8. 6. Conclusion
  9. References

A potential concern with our main findings is that they may be induced by factors other than earnings management. These factors include the effects of scaling variables and the possibility that restatements in our sample represent some form of “noise”. The overarching concern is that both factors yield restated distributions in our sample that cannot necessarily be characterized as ex-ante “unmanaged” earnings distributions. We elaborate on these factors in more detail below.

Scaling variables

One concern is that our results may be driven by the choice to scale by market value of equity, which is inflated by earnings management during the class period. If, for example, the restatement changes a profit to a loss, this loss will be scaled by an inflated market value, which could pull observations on the left of bin zero closer to bin zero. In a related fashion, the restated distribution is not a true “unmanaged” distribution because market value of equity would have been lower on average if no earnings management had taken place. Due to this concern, and concerns in the prior literature related to the influence of the scaling variable on results (e.g., Durtschi and Easton 2005), we also scale the earnings level and earnings changes distributions by both total assets and the post-class period market value of equity (calculated in the quarter after the alleged fraud is revealed) and perform our main analysis above.22 The distributions of earnings changes and earnings levels using these alternate scalars are presented in Figure 4.23

image

Figure 4.  Alternate scaling variables — Earnings changes and earnings levels

Download figure to PowerPoint

For the earnings changes scaled by total assets distribution, the shifts are similar, but more dramatic, than those reported earlier. We observe statistically significant decreases (increases) in the proportion of observations in bins negative one and two (zero, one, and two) (moving from the restated to the originally reported earnings). Using the BD test (untabulated), the size of the post-managed earnings discontinuity is 0.004 (test statistic = 0.45) and the size of the unmanaged earnings discontinuity is −0.014 (test statistic = −1.52). Using the DPZ test, the discontinuity shifts from 0.033 (test statistic = 6.35) for unmanaged earnings to −0.016 (test statistic = −2.30) for managed earnings, indicating firms are disproportionately more likely to fall in bin negative one on a restated (unmanaged) basis. Results are similar utilizing the post-class period market value of equity. Thus, the tenor of our results with respect to earnings changes does not appear sensitive to the scaling variable.

For the earnings level scaled by total assets distribution, we observe a statistically significant decrease in the proportion of observations in bin negative one, and significant increases in bins two through four (in both cases, moving from the restated to the originally reported earnings). In untabulated tests, the size of the discontinuity is roughly one-third smaller in the unmanaged earnings distribution, with a BD test statistic of 3.97 (compared to 6.82 in the managed distribution). However, a noticeable and statistically significant discontinuity is still present. Results are similar for the earnings level distribution scaled by post-class period market value of equity.

In summary, the discontinuity in the earnings changes distribution is eliminated on a restated basis, regardless of the scaling variable used. Results are more mixed for the earnings level distribution. While restating earnings for the effects of earnings management does diminish the discontinuity in the earnings level distribution when alternate scalars are utilized, a sizable discontinuity exists even on an unmanaged basis. The existence of a discontinuity in the earnings level distribution on an unmanaged basis could indicate that factors other than earnings management contribute to this discontinuity in our sample, consistent with recent research. Overall, however, these tests indicate the inference that earnings management contributes significantly to discontinuities in our sample is robust to alternate scaling variables.

Are restatements a form of noise?

An additional concern with our design is that, in the context of adjusting earnings distributions, restatements may represent some form of “noise”. At least three arguments can be made along these lines. First, adding random amounts to observations in a kinked distribution will tend to produce a smoother and more symmetric distribution (Dechow et al. 2003), and one might contend restatements represent such random noise. If this were true, then our results could be mechanical because adding noise to the managed distribution would likely produce a smoother distribution and lead to similar results (i.e., the discontinuity would be eliminated or attenuated — depending upon the magnitude of the random amounts added). However, this effect does not apply to our setting. Our restatements are not simply random noise, because our strict sample selection criteria help ensure the restatements in our sample represent intentional, nonrandom manipulations of earnings.24

We validate that our restatements are not noise in two ways. First, we rerun our analysis with further restrictions to eliminate instances where the inference of earnings management might be questioned. We initially repeat our tests on the intersection of our sample and the Hennes et al. 2008 irregularities sample, which includes 823 firm-quarters from our main sample. These results, reported in Table 4, are very similar to the primary results. Across all three panels, discontinuities are present (with t-statistics of 2.00 or greater) using originally reported earnings, but are insignificant at conventional levels using restated data. We also repeat our tests using only firms where the settlement exceeded 0.5 percent of the defendant’s market value of equity, which is a proxy for meritorious settlements (see Johnson et al. 2007). Untabulated results using these observations, which account for 1,079 out of 1,284 observations from our main sample, are again very similar to the main results. Across all three distributions, discontinuities are significant (with t-statistics of 2.00 or greater) using originally reported earnings, but are not significant at conventional levels using restated data.

Table 4. Intersection of litigation sample and Hennes et al. 2008 irregularity sample
Panel A: Analyst forecast errors
 Discontinuity tests
BD (1997) test
Size of discontinuityTest statistic
Restated0.0030.28
Originally reported−0.069−5.18
N 697 
Panel B: Earnings changes
 Discontinuity tests
BD (1997) test DPZ (1999) test
Size of discontinuityTest statisticSize of discontinuityTest statistic
Restated0.0040.35−0.002−0.15
Originally reported−0.026−2.24−0.069−3.90
N 820   
Panel C: Earnings levels
 Discontinuity tests
BD (1997) test
Size of discontinuityTest statistic
  1. Notes:

  2. This table contains discontinuity tests identical to those in Table 3, for the intersection of the primary litigation sample and the Hennes et al. 2008 irregularities sample. See Table 3 for descriptions of test statistics. Note that the DPZ tests was not conducted for the analyst forecast error distributions because the mode of the restated distribution is not bin zero, which renders this test inappropriate.

Restated−0.008−0.67
Originally reported−0.044−4.84
N 823 

Second, as discussed above, we hand-collect a sample of 50 firms in our sample with a restatement in the Point-in-Time data and find that 46 of the firms’ restatements were directly related to the allegations of fraud. In the other four cases, it was not clear whether the restatements were related to the fraud allegations, but the restatements were so minor that the firms did not change bin assignments. Thus, we find no cases where firms moved bins in our random sample due to possible noise.

A second argument related to noise involves the analyst forecast error distribution. Durtschi and Easton (2005) conjecture that the asymmetry in this distribution is not attributable to earnings management but rather to properties of analyst forecasts. Specifically, if analyst pessimism is asymmetric relative to their optimism (i.e., forecast errors tend to be larger in magnitude when negative), the distribution of analyst forecast errors may be asymmetric about zero. Thus, restating reported earnings could add “noise” to this forecast error distribution because the nonfraudulent earnings were not the object or target of analyst forecasts. Thus, our restated distribution may not be an “unmanaged” distribution so much as it is a “noised-up” distribution of errors in analyst forecasting patterns. This effect could contribute to our main findings.

The third argument is that restatements involve discretion by firm management, and therefore may be a noisy and biased measure of earnings management during the class period. Firm managers may have incentives to understate restated earnings to create “cookie jar” reserves to be used later to boost performance. Balancing this concern is the fact that sample firms just suffered substantial costs, including legal fees, reputational damage, and the settlement amount as a result of accounting improprieties. Thus, firms may be reluctant to manage earnings via the restatement when they were just penalized for such behavior. Nevertheless, when viewed in this light, the restated distributions in our sample may not truly be “unmanaged” distributions, as they may actually contain downward earnings management. This effect could contribute to our main findings.

To address the second and third arguments above, we compare originally reported earnings distributions during the class period to post-class period earnings distributions for sued firms. The advantage of this approach is that the post-class period distributions are not constructed with restatement data in any way, yet firms have strong incentives to not manage earnings upward in the post-class period. Thus, if earnings management drives discontinuities in the earnings distributions of our sample firms during the class period, we expect these discontinuities to be drastically attenuated in the post-class period. On the other hand, if our main results are driven by the effects described above, we expect discontinuities to remain because (1) restatements would not be present to “noise up” analyst forecast errors, and (2) firms could release the “cookie jar” reserves from the restatement to meet future earnings targets.

Figure 5 plots the distribution of analyst forecast errors, earnings changes, and earnings levels for our sample firms during the class period and the post-class period.25 The large discontinuities apparent in distributions during the class period (using originally reported data) disappear in the post-class period. In untabulated tests, we find that none of the discontinuities in the post-class period are statistically significant. These results are consistent with firms engaging in less earnings management in the post-class period, and give some assurance that the noise effects of restatements described above are not contributing significantly to our main findings. Instead, these results confirm inferences from our main analysis. The post-class period distributions, like our restated class period distributions, represent more “natural” earnings distributions, relatively free of earnings management. These more “natural” earnings distributions exhibit almost no signs of discontinuities for our sample firms.

image

Figure 5.  Comparison of sample firms to post-class period

Download figure to PowerPoint

5. Comparing sample firms to the broad population

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Prior research
  5. 3. Research design and empirical results
  6. 4. Threats to internal validity
  7. 5. Comparing sample firms to the broad population
  8. 6. Conclusion
  9. References

Our main analysis indicates earnings management contributes significantly to discontinuities in the earnings distributions of firms in our sample. However, because our sample of sued firms is small and is not selected from the population at random, we cannot readily generalize our findings to the broader population of firms.26 Thus, our main analysis does not definitively resolve the debate as to whether earnings management drives the discontinuities in broad samples. In addition, our sample firms comprise a relatively small portion of the broad population, approximately 1 percent, and thus do not contribute significantly to the discontinuity in the broad sample. Nevertheless, examining the earnings distributions of our sample firms (who are known earnings managers) is informative to this debate, particularly when our sample firms are compared to the broader population of firms.

To see why, note that the proportion of firms that manage earnings among the broad sample of firms is likely lower than among the subset of firms that settled securities litigation and restated their financials. Thus, if earnings management makes it easier to achieve a benchmark, then earnings management should impact the proportion of firms that meet or beat the target and the size of the discontinuity. Hence, one should observe that our known earnings management firms (1) have more pronounced discontinuities, and (2) fall in bins at or just above zero at a disproportionate rate among the broader sample of firms. If this is not the case, and the earnings distributions for known earnings managers look the same as distributions for the broad sample with less earnings management, it would be difficult to argue that earnings management per se plays a meaningful role in the shape of overall earnings distributions.

We therefore compare our sample (“litigation sample”) to the broader population of firms in COMPUSTAT from 1993 to 2006 for the earnings levels and earnings changes distributions. For the forecast error distribution, we use the COMPUSTAT–I/B/E/S population over the same time frame as our broad comparison sample. Figure 6 and Table 5 provide the comparisons.

image

Figure 6.  Comparison of sample firms to broad population of firms

Download figure to PowerPoint

Table 5. Comparison of litigation sample versus broad population — Discontinuities and benchmark beating
Panel A: Analyst forecast errors
  N Discontinuity tests Proportion tests
BD (1997) test DPZ (1999) test  Bin 0Bins 0 to 3
Size of discontinuityTest statisticSize of discontinuityTest statistic
Population189,347−0.028−38.36−0.069−11.54Population0.1510.439
Litigation sample1,094−0.066−6.26−0.078−6.68Litigation sample0.2500.562
Difference −0.038−3.56−0.009−0.72Difference0.099***0.123***
Panel B: Earnings changes
  N Discontinuity tests Proportion tests
BD (1997) test DPZ (1999) test  Bin 0Bins 0 to 3
Size of discontinuityTest statisticSize of discontinuityTest statistic
Population189,347−0.019−39.70−0.036−6.04Population0.1280.329
Litigation sample1,281−0.016−1.78−0.068−5.07Litigation sample0.1420.402
Difference 0.0030.32−0.032−2.14Difference0.014*0.073***
Panel C: Earnings levels
  N Discontinuity tests Proportion tests
BD (1997) test  Bin 0Bins 0 to 3
Size of discontinuityTest statistic
  1. Notes:

  2. Panel A contains discontinuity tests and differences in proportion tests for the analyst forecast error distribution using the litigation sample (originally reported) and the COMPUSTAT–I/B/E/S population during the period 1993–2006. Forecast errors are calculated as actual earnings from I/B/E/S less the last consensus (median) forecast at least three days prior to the announcement of earnings. Observations are placed into bins with intervals of one cent per share. The BD (1997) test comes from Burgstahler and Dichev 1997. The size of discontinuity at bin negative one and zero is equal to the actual proportion of observations in bin negative one less the expected proportion. The expected proportion is equal to the average proportions of the two adjacent bins (bin zero and negative two). This deviation from the expected value is then divided by its standard error to yield a test statistic, which is distributed asymptotically standard normal. The DPZ (1999) test comes from Degeorge et al. 1999 and is designed for situations where the suspected discontinuity involves the mode of the distribution (bin zero in this case). The size of discontinuity at bin negative one and zero is equal to the difference in the magnitude of the change in proportions of observations from bin zero to negative one versus bin zero to positive one, less the sample average of this difference across symmetric bin ranges from −11 to 11. See the text for more details. This deviation from the sample mean is then divided by its sample standard deviation to yield a test statistic, which follows a t distribution if the underlying distribution of earnings is Gaussian. Based upon simulation analysis, DPZ use a cutoff of 2.00 to assess significance. Differences in the proportions of observations in various bins across the distribution of the litigation sample versus the COMPUSTAT–I/B/E/S population are tested using a chi-square test of proportions. Panels B and C present similar analyses for the distributions of earnings changes and earnings levels, respectively. The broad population sample in panels B and C is the COMPUSTAT universe during the period 1993–2006. Earnings changes are originally reported or restated earnings before extraordinary items less originally reported or restated earnings before extraordinary items from the same quarter of the prior year, deflated by MVE from COMPUSTAT. Observations are placed into bins with intervals of 0.0025. Earnings levels are deflated by MVE from COMPUSTAT. Observations are placed into bins with intervals of 0.005.

Population189,347−0.018−48.59Population0.0730.399
Litigation sample1,284−0.038−4.87Litigation sample0.1260.559
Difference −0.020−2.53Difference0.053***0.160***

Relative to the broad population, firms in the litigation sample exhibit a significantly larger discontinuity in the analyst forecast error distribution using the BD methodology (−0.066 versus −0.028, test statistic = −3.56). Using the DPZ discontinuity methodology, there is no statistically significant difference between the litigation sample and the broad population (−0.078 versus −0.069, test statistic = −0.72). However, firms in the litigation sample are much more likely to exactly meet analyst forecasts (25.0 percent versus 15.1 percent, p-value < 0.01) or meet or beat forecasts by three cents or less (56.2 percent versus 43.9 percent, p-value < 0.01).

For earnings changes, there is no significant difference in discontinuities using the BD methodology, but litigation firms do have a larger discontinuity than the broad population using the DPZ methodology (−0.068 versus −0.036, test statistic = −2.14). In addition, litigation sample firms are more likely to report small earnings increases in bins zero through three (40.2 percent versus 32.9 percent, p-value < 0.01) relative to the broad population. For earnings levels, litigation sample firms exhibit a larger discontinuity than the overall population of firms (−0.038 versus −0.018, test statistic = −2.53). Additionally, litigation firms are more likely to report small profits in bins zero through three (55.9 percent versus 39.9 percent, p-value < 0.01).

Overall, known earnings managers have earnings distributions that exhibit greater discontinuities than the overall population of firms, where the proportion of firms that manage earnings management is likely smaller. Moreover, observations related to known earnings management are disproportionately more likely to fall in bins at or just above zero in the distributions of forecast errors, earnings changes, and earnings levels among the broader population of firms.27 This evidence is consistent with earnings management playing a role in the shape of broader earnings distributions, but the results are not conclusive proof of a link.28

Finally, we note that while our research design holds issues such as sample selection and the scaling variable constant by using each firm as its own control, the potential bias of scaling variables and non-random sample selection in the broader population would actually bias against our finding differences between our sample firms and the broader population.29 Interestingly, our results (the effect of earnings management on the discontinuity and the difference between the discontinuity for our sample firms and the broader population) are the strongest for analyst forecast errors, the test where potential scaling and sample selection issues should be least relevant.

6. Conclusion

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Prior research
  5. 3. Research design and empirical results
  6. 4. Threats to internal validity
  7. 5. Comparing sample firms to the broad population
  8. 6. Conclusion
  9. References

This study provides evidence linking earnings management to discontinuities in earnings distributions among a sample of firms that settled accounting-related securities class action lawsuits and restated earnings from the alleged violation period. Our tests compare the distribution of restated (“unmanaged”) earnings to originally reported (“managed”) earnings. We find robust evidence that earnings management drives the discontinuities in the distribution of analyst forecast errors and earnings changes for our sample firms. We find more mixed evidence with the earnings level distribution. While we find clear evidence that the discontinuity in the earnings level distribution is linked to earnings management, the strength of this link is sensitive to the scaling variable. This sensitivity could indicate that factors other than earnings management contribute to this discontinuity in our sample (see, e.g., Durtschi and Easton 2005, 2009; Beaver et al. 2007).

These results add important evidence to the earnings management literature. Recent studies provide several plausible alternative explanations for the discontinuities in earnings distributions near earnings benchmarks including sample selection, scalars, analysts’ optimism, and the effects of special items and taxes. The only way to definitively determine whether earnings management contributes to the discontinuities in earnings distributions near earnings benchmarks is to capture the amount by which each firm manages earnings. In the general population of firms, this task is impossible. However, we can observe the amount and impact of earnings management on our sample firms’ earnings distributions. Such measurable earnings management is the only opportunity to objectively determine the impact of earnings management on earnings distributions.

Though we cannot readily extrapolate our findings to a broader sample of firms, our evidence indicates that actual instances of earnings management do create sizable discontinuities in our firms’ earnings distributions. While we acknowledge that factors outside earnings management could also contribute to discontinuities, we believe it is implausible that we have captured every instance of earnings management to beat earnings benchmarks. Rather, our sample is comprised of firms that managed earnings by a sufficient magnitude to be named in a securities class action suit. It seems likely that other firms commit less egregious earnings management (that goes undetected) in order to meet earnings benchmarks. This study is therefore important in considering whether earnings management plays a role in the discontinuities in various earnings distributions documented by prior studies.

Footnotes
  • 1

     Our sample selection criteria likely capture only “extreme” forms of earnings management and do not capture all types of earnings management (e.g., within-GAAP earnings management or real earnings management).

  • 2

     Only profitable firms pay taxes, which pushes positive earnings toward the zero-profit benchmark, while firms with small losses tend to report negative special items in hopes of improving future performance, pushing loss firms away from the zero-profit benchmark.

  • 3

     For example, Dechow et al. (2003) find no evidence that firms that report small profits manage earnings more than firms that report small losses. Phillips, Pincus, and Rego (2003) report mixed evidence using abnormal accrual estimates, although they do find evidence consistent with earnings management when they use deferred tax expense as proxy for earnings management. On the other hand, Ayers et al. (2006) find evidence of a positive association between abnormal accruals and the likelihood of beating analysts’ expectations.

  • 4

     This problem is particularly salient if the discretionary specific or total accrual estimates represent unintentional forecast errors or “noise”, because subtracting noise from observations in a distribution will mechanically smooth a kinked distribution. See section 4, subsection “Are Restatements a Form of Noise?” for further explanation on the impact of noise on earnings distributions.

  • 5

     Several studies examine the use of specific accounts to meet earnings benchmarks such as analyst forecasts (e.g., Dhaliwal, Gleason, and Mills 2004; Moehrle 2002). However, these studies do not specifically examine the effect on the discontinuity generated by the firms which meet or exceed the benchmarks, as they do not distinguish between the amounts by which firms exceed the benchmark (e.g., one cent or ten cents for analyst forecasts) or the number of firms falling just short of the benchmark.

  • 6

     Approximately 87 percent of our sample involves exclusively Rule 10b-5 claims. The remaining cases include Section 11 claims, which are available only when firms offer securities for sale. These claims are subject to less stringent pleading requirements in that plaintiffs need not demonstrate that a misstatement was intentional to obtain relief. This lower intent standard is balanced by the fact that audit firms perform more work to ensure the quality of financial reports in registration statements as evidenced by higher fees when clients are offering securities (see Venkataraman, Weber, and Willenborg 2008). This increased audit effort likely reduces innocent mistakes such that firms which later settle suits likely tried to conceal something. In addition, investors have no incentive to sue over innocent mistakes that do not substantially affect share prices when disclosed, as no substantive damages would exist.

  • 7

     The pleading standard provided by the PSLRA prevents discovery until the suit survives the dismissal decision. This is in sharp contrast to standard civil litigation. In most civil cases, a plaintiff’s complaint is dismissed only if there is no legal theory upon which relief might be granted, and discovery is allowed from the onset of litigation, providing an incentive for defendants to settle even nuisance suits. The provisions of the PSLRA are designed to avoid such nuisance settlements from “strike” suits (e.g., Casey 2008).

  • 8

    Tellabs, Inc. v. Makor Issues & Rights, Ltd., 551 U.S. 308 (2007).

  • 9

    For the firms that do admit fraud, there is general agreement that fraud is a form of earnings management along with other, lesser forms of earnings management. For example, in summarizing the SEC’s view of earnings management, Dechow and Skinner (2000: 238) state that “while financial-reporting choices that explicitly violate GAAP can clearly constitute both fraud and earnings management, it also seems that systematic choices made within GAAP can constitute earnings management”.

  • 10

     The absence of securities litigation does not, of course, imply the absence of earnings management. Our sample likely does not isolate even a substantial minority of all cases of earnings management. Lesser forms of earnings management are undoubtedly present which do not result in a securities class action suit.

  • 11

     Discussion with an official in the corporate finance division at the SEC revealed that firms involved in large and convoluted restatements may need several fiscal periods to ascertain precisely what occurred and what the appropriate accounting treatments should have been. In these cases, it is not uncommon for firms to consult with SEC staff on a case-by-case basis to determine what must be disclosed in SEC filings regarding the restatements. In these cases, registrants might not go back and restate every quarterly period affected by the restatement.

  • 12

     An untabulated chi-square test of homogeneity indicates the industry composition of our sample differs from the COMPUSTAT universe during the period 1996–2005. Consistent with prior work (e.g., Johnson, Nelson, and Pritchard 2007), our sample is more heavily weighted toward high-tech firms.

  • 13

     Restatements of prior quarterly data can be contained either in amended quarterly filings (10-QAs) or in ordinary quarterly or annual filings (10-Qs and 10-Ks). COMPUSTAT records restatements from both sources.

  • 14

     The class period is the time period for which the plaintiff alleges stockholders were misled. The class period typically ends on the date the intentional misstatement is revealed to the market and extends back to the beginning of the alleged fraud.

  • 15

     We thank Andrew Leone for graciously sharing the data from Hennes et al. 2008.

  • 16

     To examine whether mergers/discontinued operations could influence our results, we repeat our primary tests excluding all observations where the sales footnote code in the Point-in-Time database indicated a merger or discontinued operations activity for the quarter in question. This screen discards roughly 30 percent of the sample. All primary results are very similar and inferences are unchanged. We therefore conclude that any noise introduced into our sample from nonfraudulent restatements has no significant impact on our inference that earnings management is associated with discontinuities in earnings distributions.

  • 17

     For example, after an internal investigation in 2003, Micromuse Inc. (a firm in our sample) restated prior earnings up for some quarters and down for other quarters due to a “cookie jar” scheme involving the inappropriate accrual of expenses. In a more publicized case, Freddie Mac, another member of our sample, allegedly engaged in a massive “cookie jar” scheme to shift income to future periods. Consequently, some of the restatements for Freddie Mac are income-increasing in certain quarters. To ensure that quarters that may merely serve as an “offset” for other quarters through the building of cookie jar reserves do not drive our results, we also rerun all analyses excluding income-decreasing restatement quarters (418 of 1,284 firm-quarters). Consistent with our main results, across all three distributions we find that discontinuities are significant (with t-statistics of 2.00 or larger, except the BD tests for earnings changes where the t-statistic is only 0.90) using originally reported earnings, but are insignificant at conventional levels using restated earnings.

  • 18

     While we acknowledge there is a technical difference between a discontinuity (lack of smoothness) and an asymmetry (first derivatives of different magnitudes) in a distribution, many papers in the prior literature do not focus on this distinction. For convenience, we generally refer to both features as a “discontinuity”.

  • 19

     Figure 1 also suggests that not all firms manage earnings to just meet or beat analysts’ forecasts. Certainly, firms have other targets (e.g., the profit loss threshold, prior year earnings, debt covenants, and internal budget targets) that are important to the firm. In addition, it is possible that they manage earnings to achieve multiple targets, making it difficult to just meet or beat each target. Thus, we do not expect all firms to be in bins zero through three.

  • 20

     There is an increase in the proportion of observations in bin negative one in the originally reported earnings, consistent with the notion that some firms managed earnings but still miss analyst forecasts by a penny. Some may find this surprising because Burgstahler and Eames (2006) document an unexpectedly low number of observations in the bin to the left of the benchmark. The inference from their findings is that firms manage earnings to move from bin negative one to bin zero. However, prior research does not document what bin firms move from or to because earnings management is unobservable. Furthermore, prior studies (e.g., Dechow et al. 2003; Ayers et al. 2006) that compare the magnitude of abnormal accruals of bins negative one to that of bin zero find mixed results and suggest that in many cases there is no difference in the magnitude of abnormal accruals between bins negative one and zero. Finally, the prior literature (Durtschi and Easton 2005, 2009) provides alternative explanations for the dip in the distribution of earnings. Thus, ex ante, it is not clear whether one would expect the number of observations in bin negative one to dip when earnings are restated or not.

  • 21

     We do not conduct the DPZ test because bin zero is not the mode of the earnings level distribution.

  • 22

     We utilize slightly wider bin widths (0.00625 for earnings changes and 0.0125 for earnings levels) for distributions scaled by post-class period market value.

  • 23

     In addition to the scalars used in Figure 4, we also scale earnings levels and changes by shares outstanding, but our limited sample size inhibits inferences with this scalar. When scaling by shares outstanding, using one cent bin ranges for our sample size is problematic because the number of observations in bins −10 to 10 ranges from roughly 10 to 30 observations per bin. Roughly 75 percent (60 percent) of the earnings levels (changes) distribution falls outside this range, which compromises statistical inferences precisely in those bins where our interest lies. However, the tenor of our results does not change when using wider bin ranges (e.g., blocks of 10 cents per share).

  • 24

     As noted in the “Sample Selection” section, our sample comprises firms that settled securities litigation. These cases survived the motion to dismiss, which requires that plaintiffs allege facts yielding a strong inference of fraudulent intent. This requirement coupled with a restatement from the alleged GAAP violation period, suggests that we capture firms that intentionally violated GAAP.

  • 25

     The majority of sample observations fall in bins −10 to 10 even in the post-class period.

  • 26

     As noted earlier in the paper, we acknowledge that our sample likely contains more extreme forms of earnings management. Therefore, our sample likely does not contain more subtle forms of earnings management, within GAAP earnings management, and real earnings management. Although our research design choice allows us to isolate the impact of earnings management on discontinuities observed among our sample of firms, it makes it difficult to generalize our results to that of the overall population of firms.

  • 27

     The fact that our sample firms meet benchmarks at a higher rate does not indicate that their meet-or-beat tendencies likely caused their litigation. This is because we include only suits that survived a motion to dismiss, indicating that a court found that the plaintiffs’ complaint presented a strong inference of fraud. This is a difficult standard to meet. For instance, courts presume that restatements are innocent errors absent evidence to the contrary (e.g., Reiger v. Price Waterhouse Coopers, LLP, 117 F. Supp. 2d 1003 [S.D. Cal. 2000]). Under such a standard, merely meeting or beating a benchmark would clearly be insufficient to withstand a motion to dismiss.

  • 28

     To establish conclusive proof, we would need to show that (1) distributions with earnings management have different shapes than distributions with less or no earnings management, and (2) earnings management observations are prevalent enough to make a difference in the broad population of firms. Our main analysis and the analysis in this section provide support for (1), and this evidence is the main contribution of our study. However, because earnings management is largely unobservable in the broad population of firms, offering comprehensive evidence on (2) is impossible. Thus, definitive proof of a link between earnings management and discontinuities in earnings distributions for the broad sample of firms will likely never be provided.

  • 29

     See Durtschi and Easton 2005, 2009 and Burgstahler and Chuk 2011 for more information about the effects of scaling and sample selection on earnings distributions.

References

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. Prior research
  5. 3. Research design and empirical results
  6. 4. Threats to internal validity
  7. 5. Comparing sample firms to the broad population
  8. 6. Conclusion
  9. References
  • Ayers, B. C., J. Jiang, and P. E. Yeung. 2006. Discretionary accruals and earnings management: An analysis of pseudo earnings targets. The Accounting Review81 (3): 61752.
  • Beatty, A. L., B. Ke, and K. R. Petroni. 2002. Earnings management to avoid earnings declines across publicly and privately held banks. The Accounting Review77 (3): 54770.
  • Beaver, W. H., M. F. McNichols, and K. K. Nelson. 2003. Management of the loss reserve accrual and the distribution of earnings in the property-casualty insurance industry. Journal of Accounting and Economics35 (3): 34776.
  • Beaver, W. H., M. F. McNichols, and K. K. Nelson. 2007. An alternative interpretation of the discontinuity in earnings distributions. Review of Accounting Studies12 (4): 52556.
  • Black, B. S., B. R. Cheffins, and M. Klausner. 2006. Outside director liability: A policy analysis. Journal of Institutional and Theoretical Economics162 (1): 520.
  • Burgstahler, D., and E. Chuk. 2011. Correcting disinformation about discontinuity: Evidence of earnings management. Working paper, University of Washington.
  • Burgstahler, D., and I. Dichev. 1997. Earnings management to avoid earnings decreases and losses. Journal of Accounting and Economics24 (1): 99126.
  • Burgstahler, D., and M. Eames. 2006. Management of earnings and analysts’ forecasts to achieve zero and small positive earnings surprises. Journal of Business Finance and Accounting33 (5–6): 63351.
  • Casey, L. L.2008. Class action criminality. The Journal of Corporation Law34 (1): 153222.
  • Choi, S. J., K. K. Nelson, and A. C. Pritchard. 2009. The screening effect of the private securities litigation reform act. Journal of Empirical Legal Studies6 (1): 3568.
  • Dechow, P. M., S. A. Richardson, and I. Tuna. 2003. Why are earnings kinky? An examination of the earnings management explanation. Review of Accounting Studies8 (2): 35584.
  • Dechow, P. M., and D. J. Skinner. 2000. Earnings management: Reconciling the views of accounting academics, practitioners, and regulators. Accounting Horizons14 (2): 23550.
  • Degeorge, F., J. Patel, and R. Zeckhauser. 1999. Earnings management to exceed thresholds. Journal of Business72 (1): 133.
  • Dhaliwal, D. S., C. A. Gleason, and L. F. Mills. 2004. Last-chance earnings management: Using the tax expense to meet analysts’ forecasts. Contemporary Accounting Research21 (2): 43159.
  • Durtschi, C., and P. Easton. 2005. Earnings management? The shapes of the frequency distributions of earnings metrics are not evidence ipso facto. Journal of Accounting Research43 (4): 55792.
  • Durtschi, C., and P. Easton. 2009. Earnings management? Erroneous inferences based on earnings frequency distributions. Journal of Accounting Research47 (5): 124981.
  • Graham, J. R., C. R. Harvey, and S. Rajgopal. 2005. The economic implications of corporate financial reporting. Journal of Accounting and Economics40 (1): 373.
  • Hayn, C. 1995. The information content of losses. Journal of Accounting and Economics20 (2): 12553.
  • Healy, P. M., and J. M. Wahlen. 1999. A review of the earnings management literature and its implications for standard setting. Accounting Horizons13 (4): 36583.
  • Hennes, K. M., A. J. Leone, and B. P. Miller. 2008. The importance of distinguishing errors from irregularities in restatement research: The case of restatements and CEO/CFO turnover. The Accounting Review83 (6): 1487519.
  • Jacob, J., and B. N. Jorgensen. 2007. Earnings management and accounting income aggregation. Journal of Accounting and Economics43 (2–3): 36990.
  • Jegadeesh, N., and J. Livnat. 2006. Revenue surprises and stock returns. Journal of Accounting and Economics41 (1–2): 14771.
  • Johnson, M. F., K. K. Nelson, and A. Pritchard. 2007. Do the merits matter more? The impact of the Private Securities Litigation Reform Act. Journal of Law, Economics, and Organization23 (3): 62752.
  • Karpoff, J. M., D. S. Lee, and G. S. Martin. 2008. The cost to firms of cooking the books. Journal of Financial & Quantitative Analysis43 (3): 581611.
  • Kerstein, J., and A. Rai. 2007. Intra-year shifts in the earnings distribution and their implications for earnings management. Journal of Accounting and Economics44 (3): 399419.
  • Kothari, S. P., A. J. Leone, and C. E. Wasley. 2005. Performance matched discretionary accrual measures. Journal of Accounting and Economics39 (1): 16397.
  • McVay, S. E.2006. Earnings management using classification shifting: An examination of core earnings and special items. The Accounting Review81 (3): 50131.
  • Moehrle, S. 2002. Do firms use restructuring charge reversals to meet earnings targets?The Accounting Review77 (2): 397413.
  • Petroni, K. R., S. G. Ryan, and J. M. Wahlen. 2000. Discretionary and non-discretionary revisions of loss reserves by property-casualty insurers: Differential implications for future profitability, risk and market value. Review of Accounting Studies5 (2): 95125.
  • Phillips, J., M. Pincus, and S. O. Rego. 2003. Earnings management: New evidence using deferred tax expense. The Accounting Review78 (2): 491521.
  • Roychowdhury, S. 2006. Earnings management through real activities manipulation. Journal of Accounting and Economics42 (3): 33570.
  • Schipper, K. 1989. Commentary on earnings management. Accounting Horizons3 (4–5): 91102.
  • Skinner, D. J.1994. Why firms voluntarily disclose bad news. Journal of Accounting Research32 (1): 3860.
  • Venkataraman, R., J. P. Weber, and M. Willenborg. 2008. Litigation risk, audit quality and audit fees: Evidence from initial public offerings. The Accounting Review83 (5): 131545.