1. What is the issue being addressed in the paper? 2. What are the findings of the paper? 3. Why is this paper important to auditors, and what are the implications of this paper for the auditing profession? 4. Describe the research methodology used as a basis for the conclusions. 5. describe any limitations of the research.

Auditing: A Journal of Practice & Theory Vol. 27, No. 2 November 2008 pp. 1-29 American Accounting Association DOI: 10.2308/aud.2008.27.2.1 Revenue Manipulation and Restatements by Loss Firms Jeffrey L. Callen, Sean W. G. Robb, and Dan Segal SUMMARY: This paper investigates the relation between the extent of a rm's past and expected future losses or negative cash ows and the ex ante probability that it will manipulate revenues. When a rm has a string of losses or negative cash ows, traditional valuation models do not yield reliable estimates of rm value, and traditional price-earnings ratios are not meaningful. Evidence suggests that market participants tend to value loss rms on the basis of the level and growth in revenues, rather than cash ows and earnings, thereby motivating these rms to overstate revenue. In fact, empirical results indicate that there is a positive relation between the number of years that rms exhibit and/or anticipate losses or negative cash ows and investment in receivables after controlling for credit policy. We further show that the ex ante likelihood that rms manipulate revenue in violation of GAAP is positively associated with the history of past and expected future losses or negative cash ows, as well as with the investment in accounts receivable adjusted for credit policy. Our results suggest another indicator of manipulation that may be used by auditors and regulators in identifying rms that are more likely to overstate revenues. Keywords: revenue manipulation; earnings management; auditing; restatements. Data Availability: Data used in this study are available from public sources. INTRODUCTION The purpose of this study is to show that the greater a rm's string of past and expected future losses or past and expected future negative operating cash ows, the more likely it is to violate GAAP by overstating revenues and accounts receivable in order to induce a higher market valu- Jeffrey L. Callen is a Professor at the University of Toronto, Sean W. G. Robb is an Associate Professor at the University of Central Florida, and Dan Segal is an Associate Professor at the University of Toronto. We thank the two anonymous reviewers, Rachel Schwartz Associate Editor, Theresa Libby, Robert Mathieu, Gordon Richardson, Ping Zhang, and workshop participants at Simon Fraser University, University of Central Florida, University of Manitoba, The University of Texas at San Antonio, University of Toronto, Virginia Commonwealth University, and York University for helpful comments and suggestions. We thank Amy Hageman for research assistance. The authors gratefully acknowledge nancial support from a Social Sciences and Humanities Research Council SSHRC Standard Research Grant. Submitted: October 2005 Accepted: May 2008 Published Online: March 2009 1 2 Callen, Robb, and Segal ation. This in turn suggests a new indicator of manipulation that may be used by auditors and regulators in identifying rms that are more likely to overstate revenues.1 The linkage between loss rms and revenue manipulation has its logical genesis in the popular press and the accounting literature on Internet rms. Both of these sources maintain that, absent a sufcient time series of positive earnings and cash ows, traditional valuation models such as the discounted cash ow and discounted residual earnings models do not yield reliable estimates of rm value. Given negative earnings or negative cash ows, analysts tend to follow the price-tosales ratio instead e.g., Demers and Lev 2001. In addition, although revenues are not the only source of value-relevant information, a number of academic studies have shown that the market views revenues and revenue growth as highly important in valuing Internet rms Hand 2000; Trueman et al. 2000, 2001; Bagnoli et al. 2001; Campbell and Sefcik 2002; Davis 2002; Bowen et al. 2002. This study extends the argument to loss rms in general. If the market substitutes revenues and revenue growth for earnings and earnings growth in valuing Internet rms because losses or negative cash ows do not provide much if any value-relevant information, then the same argument applies almost as forcefully for other rms with strings of past and expected future losses or negative cash ows. The relative importance of revenues in determining the market capitalization of rms that report a string of losses or negative cash ows henceforth loss rms provides an incentive for these rms to manipulate revenues in order to achieve greater market capitalization.2 For the same reason, loss rms are less interested in manipulating expenses because earnings are not particularly value relevant. Firms for which earnings are value relevant may also attempt to manipulate revenues but, in contrast to loss rms, are just as likely to manipulate expenses e.g., Enron. Therefore, loss rms should yield a less \"noisy\" sample of revenue manipulators by comparison to other rms.3 In contrast to most studies in the earnings management literature, we investigate earnings manipulation by loss rms through the prism of revenue restatements. In most studies of earnings management, the researcher uses a proxy for earnings management and, therefore, cannot be certain that earnings have in fact been manipulated for a discussion of this issue see Marquardt and Wiedman 2004. Exceptions are precisely those studies that are based on restatement data e.g., Richardson et al. 2003. We use restatement data to infer the ex ante likelihood that a rm will manipulate revenues. Restatements arising out of accounting errors involving revenue overstatements are fairly strong indicators of revenue manipulation, and are not necessarily a result of enforcement actions.4 In what follows, we rst show that revenues are value relevant in explaining the market value of loss rms whereas, in contrast, earnings and operating cash ows are not signicantly associ1 2 3 4 On a more conceptual level, this paper also attempts and is perhaps the rst to show that the history of earnings and cash ows matters to restatements by loss rms. See Lev et al. 2008 on this issue in the context of restatements generally. Obviously, some rms may manipulate revenues in order to avoid losses. To the extent that they succeed in showing prots, they are not part of our sample of loss rms if loss rms are dened based on earnings. Nonetheless, these rms may be captured in the sample of loss rms if loss rms are dened on the basis of negative operating cash ows. Arguably, a sample of Internet rms might yield a cleaner sample. However, the universe of Internet rms that restated revenues is too small to yield a meaningful sample. The SEC perceives young growth rms to have a higher likelihood of nancial statement fraud and nancial distress Feroz et al. 1991; Beneish 1997. If the SEC also targets loss rms for ling review a possibility, although without empirical support, we would nd a higher proportion of loss rms among restaters and a potential spurious correlation between losses and earnings management as gauged by restatements. Although the SEC was involved in approximately 55 percent of the revenue restatements during the time period covered by this study, only a portion of these were initiated by the SEC. In most restatement cases, the SEC launched an investigation subsequent to a voluntary restatement by the company. The latter nding is also supported by Anderson and Yohn 2002 and Palmrose et al. 2004. Auditing: A Journal of Practice & Theory American Accounting Association November 2008 Revenue Manipulation and Restatements by Loss Firms 3 ated with the market value of these rms. We then document a positive relation between the number of years that a rm exhibits and/or anticipates losses or negative cash ows and its investment in receivables, after controlling for credit policy. This result is consistent with loss rms being more likely to manipulate revenues than protable rms and suggests another indicator for auditors to identify rms with potential revenue misstatements. Finally, we provide evidence that there is a positive relation between the ex ante probability of revenue manipulation and the number of years that a rm exhibits and/or anticipates losses or negative cash ows. We nd that the relation is far more signicant for negative cash ows than for earnings losses, which is to be expected since cash ow-based valuation is far more common than earnings-based valuation, especially for large rms Graham and Harvey 2001. It is precisely when cash ow valuation fails that managers have an incentive to manipulate revenues in order to maintain or increase market capitalization. We also show that the ex ante probability of revenue manipulation is positively related to the level of accounts receivable after controlling for the credit policy of the rm, leverage, the ratio of inventory to total assets, and the volatility of equity returns. While intuition suggests that small young growth rms are candidates for revenue manipulation, our results indicate that a history of losses or negative cash ows is associated with revenue manipulation even after controlling for rm age, size, and growth. The remainder of the paper is divided into ve sections. The next section reviews the literature on revenue manipulation and develops the hypotheses. This is followed by a section describing the research design, a section which details the data selection criteria, and a section which presents the results. The nal section is devoted to a discussion of the ndings and their implications for auditing practice. LITERATURE REVIEW AND HYPOTHESES DEVELOPMENT Pronouncements by the Securities and Exchange Commission SEC and the Financial Accounting Standards Board FASB indicate substantial concern about the tendency of Internet and technology rms to report misleading levels of revenue see FASB 1999, SAB No. 101, EITF 99-17. Furthermore, responding to widespread concerns that investors did not receive reliable nancial information in relatively recent periods of frenetic revenue growth, regional ofces of the SEC, the Federal Bureau of Investigation, and the United States Attorney General's ofce cooperated in a legal crackdown of accounting violations related to revenue recognition see Schoenberger 2001. Indeed, the total number of restatements due to revenue-related errors has increased substantially over time. The number of revenue-related restatement cases from 1997 to 1999 is almost twice as many as the number of cases in the period from 1988 to 1996 Callen et al. 2006. The literature on earnings management through revenue manipulation is fairly recent. Dechow et al. 1996 show that SEC enforcement actions are likely to involve revenue recognition issues. Nelson et al. 2002, 2003 provide survey data conrming that income-increasing earnings management involving revenue recognition are common occurrences. Plummer and Mest 2001 provide empirical evidence concerning the incentives rms have to \"meet or beat\" analysts' expectations through revenue manipulation, while Bagnoli et al. 2001 nd that capital markets respond to revenue surprises. Although these papers provide evidence of revenue manipulation, they do not analyze the relation between loss rms and revenue manipulation. A second stream of relevant research concentrates on revenue manipulation by young rms. Rangan 1998, Teoh et al. 1999, and Shivakumar 2000 document that rms manage earnings upwards in periods prior to issuing equity in an attempt to increase share value. Marquardt and Wiedman 2004 argue that rms issuing equity prefer to manage earnings through a mechanism that suggests to the market that the reported earnings level will persist into the future in order to Auditing: A Journal of Practice & Theory November 2008 American Accounting Association 4 Callen, Robb, and Segal maximize the proceeds from the share issuance. Consequently, relatively new rms refrain from managing income through nonrecurring items, but instead use their discretion over sales revenue or operating expenses to achieve their earnings objectives. In addition, life cycle theory suggests that a growth-maximization strategy is most cost-benecial when rms are relatively young.5 Therefore, signaling growth through aggressive revenue recognition methods may result in a positive stock price reaction. Indeed, Anthony and Ramesh 1992 show that the stock price response coefcient on unexpected sales growth is signicantly larger for young rms. A third stream of literature investigates the importance of revenues for Internet rms. Since most Internet rms report losses and negative cash ows, traditional valuation models cannot be applied and price-to-earnings ratios cannot be meaningfully calculated and compared. This in turn leads to an increase in the importance of revenues, as Internet rms are likely to be valued based upon their revenues. Indeed, Demers and Lev 2001 show that analysts follow the price-to-sales ratios of Internet companies. Hand 2000, Campbell and Sefcik 2002, Davis 2002, and Bowen et al. 2002 provide empirical evidence that the market impounds reported revenues in the stock prices of Internet rms, and Bagnoli et al. 2001 and Davis 2002 demonstrate that the market responds to revenue surprises. Furthermore, Bowen et al. 2002 show that revenue levels are strongly associated with the market's valuation of Internet rms. Taken as a whole, there is convincing evidence that Internet rms have economic incentives to manipulate reported revenues in the presence of multiple years of negative reported cash ows and earnings. Overall, the literature indicates that young rms with negative earnings have economic incentives to report articially high levels of revenue. The key reasons for revenue manipulation seem to be to create positive expectations of future growth through sales and the incentive to induce higher market capitalization.6 This paper provides the actual link between losses, revenue manipulation, and market capitalization. Specically, we show that revenues are value relevant for loss rms in general regardless of industry classication, and that the probability of revenue manipulation is increasing with the rm's string of negative cash ows and losses.7 Generally, rms manipulate revenues either through accounts receivable or unearned revenues, depending on the reason for the manipulation and the timing of cash collection. Some rms may manipulate revenues in order to smooth growth, whereas other rms may understate revenues to avoid regulatory sanctions or to minimize taxes see Healy and Wahlen 1999. In these latter cases, revenue manipulation may be achieved through manipulation of the \"unearned revenue\" account. In contrast, manipulation to overstate revenues is usually achieved by recording fraudulent sales and/or by the premature recognition of legitimate sales. These forms of manipulation generally ow through accounts receivable. Given this conjecture, we expect that the investment in accounts receivable by loss rms would be higher than by nonloss rms after controlling for credit policy.8 5 6 7 8 See Porter 1980 for a more complete discussion of life cycle theory. Additional incentives to inuence stock prices through revenue manipulation include managerial stock option compensation plans and rm access to equity capital see Bowen et al. 1995. These additional incentives are consistent with the arguments in this paper. Clearly the sample of loss rms includes rms that are headed toward bankruptcy. Although these rms may also have incentives to manipulate revenues, it is reasonable to assume that the incentives for doing so are not related to market capitalization since these rms are likely to be valued at their liquidation value. Hence, the inclusion of nancially distressed companies probably biases the analysis against nding evidence that relates revenues to market capitalization. We report the results of the analysis excluding nancially distressed companies in the sensitivity analysis section. The assumption that all revenue manipulation cases are accompanied by overstatements of accounts receivable biases the analysis against nding a positive association between revenue manipulation and the level of accounts receivable. Auditing: A Journal of Practice & Theory American Accounting Association November 2008 Revenue Manipulation and Restatements by Loss Firms 5 This discussion leads us to the following three hypotheses expressed in the alternative form: H1: There is a positive relation between the extent of a rm's past and anticipated future losses negative cash ows and its ratio of accounts receivable to sales. Hypothesis 1 follows from our conjecture that accounts receivable is the primary accounting mechanism by which loss rms manipulate revenues. H2: There is a positive relation between the extent of a rm's past and anticipated future losses negative cash ows and the ex ante likelihood of revenue manipulation in contravention of GAAP. Hypothesis 2 is our primary hypothesis. It follows from the conjectured economic incentives of loss rms to manipulate revenues in order to increase rm value. H3: There is a positive relation between the accounts receivable to sales ratio adjusted for credit policy and the ex ante likelihood of revenue manipulation in contravention of GAAP. Hypothesis 3 follows from H1 and H2. To the extent that accounts receivable is the primary accounting mechanism by which loss rms manipulate revenues, we expect to nd a positive relation between the level of receivables and the probability of manipulation. RESEARCH DESIGN Research Design to Test H1 We conjecture that the decision to manipulate revenues in order to maintain or increase market capitalization depends on the expectation of future losses or negative cash ows. If managers and investors do not anticipate future losses or negative cash ows, then managers will expect investors to value the rm using traditional methods that focus primarily on future earnings and cash ows rather than on sales revenue. If investors value rms using capitalized earnings, then managers may have an incentive to manipulate bottom-line earnings through expense manipulation rather than through the overstatement of revenues. Revenue manipulation is often more costly than expense manipulation since the average decrease in market value once the manipulation is discovered is much higher if the manipulation involves revenues rather than expenses see Anderson and Yohn 2002; Palmrose et al. 2004; Callen et al. 2006. In other words, the incentive to rely solely on revenue manipulation is attenuated once managers expect their rm to become protable. We dene \"loss\" rms with respect to a \"loss ratio,\" where the loss ratio for year t is computed as the proportion of years in which the rm reported negative net income operating cash ows from year t5 to year t+3 inclusive. For example, if the rm has earnings data for the period 1995-2005, the loss ratio for 2001 is computed as the proportion of loss years in the period 1996 to 2004 inclusive. The implicit assumption here is that managers have perfect foresight regarding the sign of net income in the three-year period following year t.9 This assumption is not as restrictive as it might appear since we know that analysts, who have less information about the rm than managers, routinely provide estimates of expected income. Although one may argue that 9 In the sensitivity analysis discussed below we relax the assumption of perfect foresight. Auditing: A Journal of Practice & Theory November 2008 American Accounting Association 6 Callen, Robb, and Segal analysts' estimates are far from perfect, their errors are mostly related to the level of earnings rather than the sign of the earnings. Furthermore, it is reasonable to assume that managers on average know roughly if their rm will be protable or not in the foreseeable future.10 To examine whether loss rms have higher accounts receivable balances than protable rms, we need to control for the credit policy of the rm, since extending trade credit is one of the tools used to maintain and increase competitiveness and market share.11 Petersen and Rajan 1997 provide a comprehensive overview and empirical evidence of the current theories of trade credit. Consistent with their analysis of the determinants of accounts receivable, we surmise that a rm's investment in receivables is a function of the nancial strength of the rm in general, its operational performance relative to its industry competitors, and its stage in the business cycle. The discussion that follows briey addresses each of these factors. Financially strong rms are able to extend generous credit terms in order to attract and retain customers, but, because of their wealth, are not constrained to do so. Conversely, nancially weak rms may be forced to invest in accounts receivable in order to survive, but simultaneously may be constrained by their need for cash inow. Following Petersen and Rajan 1997, we assume that large rms tend to be nancially strong and proxy nancial strength by rm size as measured by the natural log of total assets LSIZE. We proxy for the operational performance of the rm relative to its industry competitors using the four-digit SIC median adjusted growth rate in sales MDGRS_P if positive, MDGRS_N if negative, and the four-digit SIC median adjusted gross prot scaled by total sales MDGRM. MDGRS_P MDGRS_N is computed as the difference between the rm's growth rate in sales and the median growth rate in the rm's four-digit SIC industry if positive negative and 0 otherwise. MDGRM is computed as the difference between the rm's gross prot margin and the median gross prot margin in the rm's four-digit SIC industry. Following Petersen and Rajan 1997, we also include the square of MDGRM MDGRM_SQ in order to control for the potential nonlinear relation between accounts receivable and the gross margin.12 The rm's stage in the business cycle is also related to its credit policy. Young rms are more likely to extend better credit terms to their customers in order to capture greater market share and to generate superior growth rates in sales. The rm's stage in the business cycle can be proxied by the log of age LAGE and log of total assets LSIZE. To be consistent with Petersen and Rajan 1997, we also include the square of age LAGE_SQ in our analysis. Following Petersen and Rajan 1997, we predict that accounts receivable is positively associated with LSIZE, LAGE, MDGRS_P, and MDGRM, and negatively associated with MDGRS_N, LAGE_SQ, and MDGRM_SQ.13 We test H1 by regressing the rm's industry-adjusted accounts receivable to sales ratio MDARS on its loss ratio and on the proxies for the rm's credit policy discussed above, and determine whether the coefcient of the loss ratio is positive and signicant. We use both earnings-based and cash ow-based loss ratios LOSS. The earnings-based loss ratio for year t is 10 11 12 13 For many rms in our sample, analysts' forecast data are not available or there are very few analysts that cover the rm. In addition, I/B/E/S provides only one-year-ahead forecasts for the majority of their sample rms. Finally, I/B/E/S does not provide analysts forecasts of future cash ows for most rms. Nonetheless, credit sales are generally costly for two main reasons: rst, there is the risk of noncollection, and second, credit sales typically entail an implicit discount. Petersen and Rajan 1997 estimate the credit policy model with and without age and gross margin squared. The rationale for the inclusion of the squared variables is that credit policy is probably a concave function of age and gross margin, meaning that accounts receivable is positively related to the gross margin, but the slope is decreasing. Because of the detailed nature of their data from the National Survey of Small Business Finances, Petersen and Rajan 1997 are able to include in their analysis two additional variables that potentially affect credit policy; namely, whether the rm operates in an urban or a rural environment and the maximum amount that can be drawn on a line of credit. Auditing: A Journal of Practice & Theory American Accounting Association November 2008 Revenue Manipulation and Restatements by Loss Firms 7 computed as the proportion of years in which the rm reported negative net income from year t 5 to year t+3 inclusive. The cash ow-based loss ratio for year t is computed as the proportion of years in which the rm reported negative cash ows from year t5 to year t+3 inclusive. We perform separate regressions on the earnings and cash ow loss ratios because of the high correlation 0.75 between these two variables. Formally, we estimate the following model: MDARSit = 0 + 1LOSSt + 2LSIZEit + 3LAGEit + 4LAGE _ SQit + 5MDGRS _ Pit + 6MDGRS _ Nit + 7MDGRM it + 8MDGRM _ SQit + year dummies + it 1 where i denotes the rm and t is a time index. We reject H1 if the estimated coefcient of LOSS 1 is less than or equal to 0. Research Design to Test H2 and H3 To determine if loss rms manipulate revenues through the overstatement of accounts receivable, we examine whether the ex ante probability of revenue manipulation is positively associated with the loss ratio and the level of accounts receivable, after controlling for factors that affect the credit policy of the rm. Since the probability of manipulation is unobservable, we utilize error restatement data to estimate the probability of manipulation using the two-stage sequential \"partial observability\" probit model of Poirier 1980 and Abowd and Farber 1982.14 Krishnan and Krishnan 1996 apply this model to the auditor's qualication decision. We use the two-stage probit model as opposed to a standard probit model for two main reasons. First, not every rm that is able to manipulate revenues does so. There is a decision calculus involved. Failure to account for this decision results in a sample selectivity problem and potentially biased coefcients.15 Second, the manipulation of revenues and the discovery of the manipulation take place at two sequential points in time the manipulation at time t and the discovery of the manipulation and the need to restate at time t+x, x 0. For example, consider a rm that was discovered in year 2002 to have overstated year 2000 revenues and, as a consequence, the rm is required to restate year 2000 revenues. From the perspective of an external observer, the restatement of year 2000 is a two-stage process. In the rst stage, management decides to manipulate revenues in year 2000. In the second stage, the manipulation remains dormant until it is discovered in year 2002. Recognizing that the manipulation of revenues and the discovery of the manipulation take place at two sequential points in time implies that of the total set of variables that explain the probability of restatement, some will be specic to the probability of manipulation, such as LOSS, and other variables will be specic to the probability of not discovering the manipulation when the reports are actually manipulated, such as auditor experience EXP. Focusing the analysis on each set of factors separately, one set that determines the probability of manipulation and one set that determines the probability of detecting the manipulation, is more likely to lead to the correct parameter values for each separate probability and hence the product of these probabilities than using all variables together to try to explain the overall product of these probabilities; that is, the overall probability of a restatement. In short, properly recognizing the structure of the restatement process should yield better parameter estimates. 14 15 Poirier 1980 rst developed the two-stage \"partial observability\" probit model in a simultaneous events context. Abowd and Farber 1982 further extended the model to a sequential events context. See the discussion of the \"partial observability\" model by Maddala 1983, 362-364. Auditing: A Journal of Practice & Theory November 2008 American Accounting Association 8 Callen, Robb, and Segal In probabilistic terms, the probability of a restatement can be expressed as the joint probability of these two stages. Let P. denote the probability of an event and let P./. denote the conditional probability. Furthermore, let: REV_RESit the event of a restatement of year t nancial statements by rm i in year t+x, x 0; M it the event of revenue manipulation in year t by rm i; and UM i,t+x the event that revenue manipulation by rm i remains undiscovered until year t+x, x 0. Using basic probability theory, we can write the probability of a restatement of year t nancial statements due to revenue manipulation as the product of the probability that the rm manipulates revenues in year t and the probability that the manipulation remains undiscovered until year t+x, x 0, conditional on revenue manipulation in year t: PREV_RESit = PM it and UM i,t+x = PM it*PUM i,t+x/M it. 2 Assuming that the probability that the rm manipulates its revenues in year t is a positive linear function of a vector of the rm's observed characteristics Xi and a white noise innovation term i, allows us to express PM it as: PM it = PXit + it 0 3 where is the vector of parameters to be estimated. Similarly, assume that the probability that the revenue manipulation remains undiscovered until year t+x given that the rm manipulated revenues in year t is a positive linear function of a vector of the rm's observed characteristics Zit and a white noise innovation term it. Then, we can express PUM i,t+x / M it as: PUM i,t+x/M it = PZit + it 0 4 where is the vector of parameters to be estimated. Substituting Equations 3 and 4 into Equation 2 yields the unconditional probability of a restatement expressed as the product: PREV_RESit = PXit + it 0*PZit + it 0. 5 Estimating Equation 5 by maximum likelihood yields consistent estimates of the parameter vectors and . We conjecture that the ex ante probability of revenue manipulation PM it = PXit + i 0 is a function of the LOSS ratio and other benets and costs of manipulation. Lacking direct measures of the benets of manipulation, we proxy for these benets by rm characteristics that are shown in the literature to be related to the benets of manipulation. Specically, we measure the benets of revenue manipulation by the log of rm age LAGE, growth in sales GRS, leverage LEV, nancial distress as measured by Altman's Z-score ALT_Z, business risk as measured by the volatility of the rm's equity returns SDR, computed as the standard deviation of the residuals of the regression of daily stock returns on the value weighted market return in the scal year, and the level of inventories as measured by inventory to total assets INV. Excess sales growth has been shown to be associated with fraud Beasley 1996; Bell and Tabor 1991; Loebbecke et al. 1989. Firms with greater leverage are more likely to violate their debt covenants, providing an incentive to manipulate. Firms in nancial distress and rms with high business risk are likely to manipulate revenues to avoid bankruptcy, while the likelihood of fraud is positively associated with nancial distress and the risk of bankruptcy Bell and Tabor 1991; Loebbecke et al. 1989. The age of the rm, volatility of stock returns, and the ratio of Auditing: A Journal of Practice & Theory American Accounting Association November 2008 Revenue Manipulation and Restatements by Loss Firms 9 inventory to total assets are also associated with the likelihood of fraud Stice 1991; Pratt and Stice 1994. We also control for the impact of the credit policy of the rm on accounts receivable by using the residual MDARS_RES from the credit policy model of Equation 1 as a regressor instead of the raw accounts receivable to sales ratio. The residual reects the excess accounts receivable to sales ratio after controlling for growth in sales, size, age, and protability.16,17 We measure the cost of manipulation by a proxy for managerial turnover that is a consequence of manipulation being discovered TURNOVER. Hennes et al. 2007 show that managerial turnover is highly positively correlated with the size of the Cumulative Abnormal Return CAR around the restatement announcement; the more negative the CAR the greater the likelihood of managerial turnover. To obtain an ex ante measure of turnover, we rank all restating rms by size deciles and compute the average CAR around the restatement for each size decile. Expected managerial turnover is then measured for each rm whether it had to restate or not by the absolute value of the average CAR in the three days around the restatement announcement in the size decile to which the rm belongs. Finally, we also control for rm size SIZE measured by the log of market value of equity. Formally, we assume that PM it = P0 + 1LOSSit + 2MDARS _ RESit + 3LEVit + 4GRSit + 5TURNOVERit + 6LAGEit + 7ALT _ Zit + 8SDRit + 9INVit + 10SIZEit + it. 6 We predict that the coefcients of LOSS, MDARS_RES, LEV, GRS, ALT_Z, SDR, and INV are positively related to the probability of revenue manipulation while the coefcients of TURNOVER, LAGE, and SIZE are negatively related. The probability that revenue manipulation remains undiscovered until year t+x, given revenue manipulation in year t, is assumed to be a function of whether the auditor is one of the Big 8 AUD, the auditor's expertise in the industry EXP measured as the log of the number of contemporaneous audit clients in the same four-digit SIC code that employ the same auditor, and rm size SIZE. This probability is hypothesized to be negatively related to AUD and EXP since Big 8 auditors and auditors with more industry experience are more likely to discover the manipulation, and so it is less likely to remain undiscovered.18 We make no prediction about SIZE. Although large rms are scrutinized more closely by auditors, the ability to hide fraud is likely easier in large rms. More formally, we assume that: PUM i,t+x/M it = P0 + 1AUDit + 2EXPit + 3SIZEit+it. 7 DATA Financial statement and price data are collected from the annual Compustat and the monthly CRSP databases, respectively. We begin by identifying all rms included in Compustat from 1992-2005 with nonmissing net income before extraordinary items DATA18 and cash ow from operations DATA308. Using this sample, we compute the loss ratio. We then eliminate observations with missing values of sales DATA12, growth in sales, accounts receivable DATA2, inventory DATA3, or Altman's Z-score. In addition, we eliminate industries which have no revenue-related restatement cases. 16 17 18 Note that the credit policy model was estimated without including the loss ratio among the independent variables. In a sensitivity analysis we estimate the model by including all the variables of the credit policy model along with the ratio of accounts receivable to sales among the independent variables. In a separate analysis, we also estimate the model replacing MDARS_RES with the actual accounts receivable to sales ratio. The results obtained for both analyses are similar to those reported. The negative relation between EXP (AUD) and the second-stage probability can be demonstrated analytically. Auditing: A Journal of Practice & Theory November 2008 American Accounting Association 10 Callen, Robb, and Segal We also impose restrictions related to stock returns and market values. We compute annual stock returns from monthly CRSP data returns including dividends. Returns are computed over a period starting nine months before and ending three months after the scal year-end. If the rm was de-listed we include the de-listed return. We also require valid market values of equity three months after the scal year end. Finally, we remove four-digit SIC codes with fewer than four rms and throughout the analyses we further eliminate the top and bottom 1 percent of each of the variables in the different regressions. The nal sample consists of 22,821 3,997 rm-years rms. Visual inspection of our sample rms shows them to be distributed widely across fourdigit SIC groups with no unusually large concentrations in any specic industry sector. We collect nancial statement restatement data from the General Accounting Ofce's GAO website and by searching Lexis for restated nancial statements for the years starting in 1993 and ending in 2002. On January 17, 2003, the GAO issued a revised restatement report GAO 2003, GAO-03-395R that identied 919 corporate nancial restatements due to accounting irregularities between January 1, 1997, and June 30, 2002. Since the GAO database only contains the date when restatements are announced, we hand-collect data for the actual year of restatements. Eliminating restatements due to in-process R&D and mergers and acquisitions reduces the GAO sample to 705 corporate nancial restatements representing 1,057 restated scal years. Searching Lexis we identify an additional 275 corporate nancial restatements representing 470 scal years. Together, the combined restatement sample consists of 980 unique restatements representing 1,527 scal years. Merging this total restatement sample with the nancial data reduces the number of restated scal years to 521, of which 262 years involve restatements of revenues. Table 1 provides descriptive statistics for the sample rms. Panel A of Table 1 shows that the sample rms are generally of medium size; the mean median market value of equity is about $1,717 $110 million; the mean median total assets is $1,704 $117 million. The mean loss ratio computed based on net income is 0.34, whereas the mean loss ratio computed based on cash ows from operations is 0.29. These loss ratios indicate that the average rm in the sample reports losses or negative cash ows from operations in about one third of the years during the sample period. The sample rms are also relatively young median age is 9 and grow faster than their peers the mean industry adjusted growth rate is 12 percent. Panel B of Table 1 shows summary statistics for quintiles formed by the loss ratio. We divide the sample rms into quintiles based on the magnitude of the loss ratiorms with a loss ratio less than 20 percent, rms with a loss ratio between 20 percent and 40 percent, and so on. Of the total number of observations, 14,691 65 percent had a loss ratio less than 40 percent while 2869 13 percent have a loss ratio greater than 80 percent. The table indicates that the median market value of equity and the median total assets decrease with the loss ratio; the median market value total assets with a loss ratio less than 20 percent is $295 $309 million, whereas the median market value total assets with a loss ratio greater than 80 percent is $45 $27 million. Although there is no discernable pattern in the book-to-market BM and price-to-sales PS ratios, the median BM is the lowest and the median PS is the highest for rms with the highest loss ratios. In addition, median age is consistently decreasing with the loss ratio. The median age for the lowest highest loss ratio quintile is 11 6, indicating that young rms are more likely to report losses, consistent with the business cycle theory of Porter 1980. Finally, the table also shows that rms in the lowest highest quintile have the lowest highest ratio of accounts receivable to sales, 0.158 and 0.184, respectively. This nding is in line with our conjecture that loss rms are more likely to manipulate revenues and that the manipulation ows through receivables. Panel C shows the Spearman and Pearson correlations among key variables. Consistent with our predictions, the correlation between the loss ratios is positively associated with the ratio of accounts receivable to sales, indicating that rms with higher loss ratios have relatively higher receivables. In addition, the table shows that the Pearson correlations among the loss ratios and Auditing: A Journal of Practice & Theory American Accounting Association November 2008 Revenue Manipulation and Restatements by Loss Firms 11 TABLE 1 Descriptive Statistics Panel A: Summary Statisticsa Mean MVE $M ASSETS $M LOSS_NI LOSS_CF MDARS MDGRS AGE MDGRM INV SDR EXP 1,717 1,704 0.34 0.29 0.02 0.12 13 0.05 0.17 0.003 12 Std. 10,003 7,071 0.33 0.33 0.13 0.99 10 0.52 0.14 0.007 15 Panel B: Median Statistics by Loss Ratios Quintilesb Loss n MVE ASSETS 0%-20% 21%-40% 41%-60% 61%-80% 81%-100% Sample 10,319 12,217 4,372 3,796 2,744 2,236 2,517 2,115 2,869 2,457 22,821 Panel C: Correlation Tablec LOSS_NI LOSS_NI LOSS_CF ARS MDARS MDGRS AGE a 295 291 88 56 50 41 37 39 45 50 110 309 339 112 70 64 45 42 34 27 23 117 LOSS_CF ARS 0.71 0.75 0.14 0.08 0.08 0.28 0.21 0.13 0.11 0.32 0.13 0.21 0.91 0.02 0.10 Q1 Median 29 31 0.00 0.00 0.03 0.12 5 0.09 0.05 0.001 3 110 117 0.25 0.14 0.00 0.00 9 0.00 0.14 0.002 7 Q3 537 580 0.571 0.50 0.05 0.15 18 0.09 0.25 0.003 15 BM PS AGE ARS 0.50 0.53 0.60 0.64 0.63 0.56 0.54 0.46 0.35 0.29 0.52 0.96 0.93 0.78 0.68 0.74 0.81 0.86 1.09 2.67 3.26 0.96 11 12 10 9 7 7 8 7 6 5 9 0.16 0.16 0.18 0.18 0.17 0.19 0.17 0.19 0.18 0.20 0.17 MDGRS AGE 0.08 0.00 0.04 0.07 0.26 0.31 0.10 0.04 0.18 MDARS 0.02 0.07 0.74 0.02 0.05 0.10 Panel A shows descriptive statistics of the main variables used in the analysis. Mean is the mean value, Std. is the standard deviation, Q1 Q3 is the rst third quartile value, and Median is the median value. MVE is the market value of equity measured three months after scal year-end. LOSS_NI LOSS_CF in year t is the proportion of years in which (continued on next page) Auditing: A Journal of Practice & Theory November 2008 American Accounting Association 12 Callen, Robb, and Segal the rm reports negative net income before extraordinary items operating cash ows from year t5 to year t+3 inclusive. b Panel B shows median statistics by loss ratios quintiles. The top number in each cell shows the median statistic for the loss ratio computation based upon income before extraordinary items. For example, the top number in the top MVE cell indicates that the median MVE of companies with an earnings-based loss ratio between 0 percent and 20 percent is 295. The number in parentheses corresponds to the median statistic for the loss ratio computation based upon operating cash ows. Thus, the number in parentheses in the top MVE cell indicates that the median MVE of companies with an operating cash ow-based loss ratio between 0 percent and 20 percent is 291. The loss ratio based on income cash ow in year t is computed as the proportion of years in which the rm reports negative net income before extraordinary items negative operating cash ows from year t5 to year t+3 inclusive. MVE is the market value of equity $millions, and n is the number of rm-year observations in each loss ratio quintile. c Panel C shows Pearson Spearman correlations below above the diagonal among key variables. All the correlations are signicant at less than 1 percent except for the correlation between MDGRS and LOSS_CF. LOSS_NI LOSS_CF in year t is the proportion of years in which the rm reports negative net income before extraordinary items negative operating cash ows from year t5 to year t+3 inclusive. Variable Denitions: AGE rm's age in years; ARS ratio of accounts receivable to sales; ASSETS total assets $millions; BM book-to-market ratio, computed as the book value of equity divided by the market value of equity three months after the scal year-end; EXP number of rms that are in the same four-digit SIC and that employ the same auditor as the rm observation; INV ratio of inventory-to-total assets; MDARS four-digit SIC median-adjusted ratio of accounts receivable to sales; MDGRM four-digit SIC median-adjusted gross prot scaled by total sales, computed as the difference between the rm's gross prot margin and the median gross prot margin in the industry; MDGRS four-digit SIC median-adjusted growth in sales, computed as the difference between the rm's growth rate in sales total sales in year t divided by total sales in year t1 and the median growth rate in the industry; PS price-to-sales ratio, computed as the market value of equity three months after the scal year-end divided by total sales; and SDR variance of abnormal returns. median adjusted growth rate in sales are positive and signicant, yet the Spearman correlation is negative and signicant. Finally, the correlation between the loss ratios is greater than 0.75, suggesting that rms that report long strings of losses also report long strings of negative cash ows and vice versa. EMPIRICAL RESULTS The Value Relevance of Revenues Bowen et al. 2002 investigate whether revenues are value relevant for Internet rms by regressing the market value of equity on earnings and revenues. They report that the coefcient on revenue for loss rm quarters is positive and signicant, indicating that revenues are still value relevant after controlling for earnings in explaining market values. To examine whether market value of equity is related to revenues after controlling for earnings, we follow Bowen et al. 2002 and regress the market value of equity on earnings, revenues, and book value of equity total assets are included in the regression to control for size:19 MVEit = 0 + 1Assetsit + 2 P _ BVEit + 3L _ BVEit + 4 P _ Earningsit + 5L _ Earningsit + 6 P _ Revenueit + 7L _ Revenueit + year dummies + it 8 where i denotes the rm, t is a time index and the s are parameters to be estimated. The variables 19 Note that our hypothesis is that loss rms manipulate revenues in order to enhance market capitalization. Hence the value relevance of revenues is examined in levels space rather than return space. Auditing: A Journal of Practice & Theory American Accounting Association November 2008 Revenue Manipulation and Restatements by Loss Firms 13 in Equation 8 are dened as follows: MVEit market value of equity measured three months after the scal year-end; Assetsit total assets Compustat item DATA6; P_BVEit book value of equity Compustat item DATA60 if net income before extraordinary items is positive and 0 otherwise; L_BVEit book value of equity Compustat item DATA60 if net income before extraordinary items is negative and 0 otherwise; P_Earningsit net income before extraordinary items Compustat item DATA18 if net income before extraordinary items is positive and 0 otherwise; L_Earningsit net income before extraordinary items Compustat item DATA18 if net income before extraordinary items is negative and 0 otherwise; P_Revenueit total revenue Compustat item DATA12 if net income before extraordinary items is positive and 0 otherwise; L_Revenueit total revenue Compustat item DATA12 if net income before extraordinary items is negative and 0 otherwise; year dummies a dummy variable indicating the year of the data observation; and it white noise innovation term. The Earnings column in Table 2 shows the OLS estimation results. Standard errors are adjusted for rm-level clustering. Consistent with Bowen et al. 2002 and Collins et al. 1999, the coefcient on book value of equity is positive and signicant p-value 10 percent. For protable rms, the coefcients on both earnings P_Earningsit and revenues P_Revenueit are positive and highly signicant p-value 1 percent. For loss rms we nd that the coefcient on earnings L_Earningsit is not signicantly different from 0. This observation is also consistent with Collins et al. 1999, who show that the coefcient on earnings for loss rms is not signicantly different from 0 for most of their sample years.20 The coefcient on loss rms' revenues L_Revenueit, however, is positive and signicantly different from 0 p-value 1 percent, indicating that revenues, and not earnings, are the main driver of market value of equity for rms that report losses. Taken as a whole, our results indicate that revenues are value relevant over and above earnings for all rms. However, while earnings are value relevant for protable rms, they are not value relevant for loss rms. These results demonstrate the singular importance of revenues for loss rms.21 The Cash Flow column shows the results of the regression where the conditioning is based on the sign of operating cash ows. Specically, P_BVEit and P_Revenueit are the book values of equity and revenues, respectively, if operating cash ows are positive and 0 otherwise. Similarly, L_BVEit and L_Revenueit are the book value of equity and revenues, respectively, if operating cash ows are negative and 0 otherwise. In addition, we replace the earnings variables P_Earningsit and L_Earningsit with P_CFOit and L_CFOit, respectively. P_CFOit represents operating cash ows if cash ows from operations DATA308 are positive and 0 otherwise. Similarly, L_CFOit represents cash ows from operations if negative and 0 otherwise. The results are similar to those reported in the Earnings column. The coefcient on revenues for rms that report negative cash ows is positive and signicant indicating that revenues are value relevant for 20 21 Although the mean coefcient of earnings for loss rms is marginally signicant overall, in 12 out of 18 years considered 1975-1992 the coefcient of earnings is not signicantly different from 0. To ensure that our results are not affected by size, we repeat the analysis after scaling all variables by book value of equity and total assets, respectively. In both cases the coefcients are qualitatively similar in sign and signicance to those reported. Auditing: A Journal of Practice & Theory November 2008 American Accounting Association 14 Callen, Robb, and Segal TABLE 2 Value Relevance Analysis (standard errors in parentheses) Variable Predicted Sign Intercept / Earnings Cash Flow 5.00 65.13 180. Assets / 0.75*** 0.28 P_BVE 1.58* 0.84 293. 1.49*** 0.33 2.75*** 0.59 L_BVE 2.07*** 0.44 0.70 P_Earnings 18.37*** L_Earnings P_CFO L_CFO P_Revenue L_Revenue 4.41*** 2.38 0.08 0.12 8.11*** 2.63 6.68 8.56 0.30** 0.14 0.71*** 0.67*** 0.24 1.03** 0.27 Adj. R2 n 0.45 0.58 0.43 22135 22135 ***, **, * Indicates signicance at the 1 percent, 5 percent, and 10 percent levels, respectively. The table shows the estimation results of the regression of market value of equity MVE on earnings, revenues, book value of equity and total assets, conditional on the sign of earnings before extraordinary items Earnings column and the regression of market value of equity on cash ows, revenues, book value of equity, and total assets, conditional on the sign of operating cash ows Cash Flow column. MVE is the market value of equity measured three months after the scal year end. Assets represent total assets DATA 6. n is the number of observations. In the Earnings column, P_BVE is book value of equity DATA 60 if net income before extraordinary items is positive, and 0 otherwise. L_BVE is book value of equity DATA 60 if net income before extraordinary items is negative, and 0 otherwise. P_Earnings L_Earnings is net income before extraordinary items DATA18 if net income before extraordinary items is positive negative, and 0 otherwise. P_Revenue L_Revenue is total revenue DATA 12 if net income before extraordinary items is positive negative, and 0 otherwise. In the Cash Flow column, P_BVE L_BVE is book value of equity if operating cash ows is positive negative, and 0 otherwise. P_Revenue L_Revenue is total revenue if operating cash ows is positive negative, and 0 otherwise. P_CFO L_CFO is cash ows from operations if operating cash ows DATA 308 is positive negative, and 0 otherwise. The standard errors are adjusted for rm-level clustering. To reduce the effect of outliers, we remove the top and bottom percentile of all variables. Auditing: A Journal of Practice & Theory American Accounting Association November 2008 Revenue Manipulation and Restatements by Loss Firms 15 rms that report negative cash ows. In contrast, the coefcient on negative cash ows L_CFO is not signicantly different from 0. Credit Policy and the Loss Ratio Panel A of Table 3 presents the regression results from the OLS estimation of Equation 1. We use two proxies for the loss ratio, one based on net income and the other based on cash ows from operations. The signs of the coefcients on the control variables are mostly consistent with Petersen and Rajan 1997. Specically, in the regression where the loss ratio is computed based on earnings Earnings column, the level of accounts receivable is positively associated with rm size p-value 1 percent, which proxies for the rm's nancial strength and its ability to extend credit. The level of accounts receivable is also positively related to growth in sales. The coefcient on MDGRS_P p-value 1 percent indicates that rms with industry adjusted positive growth in sales tend to extend more generous credit terms to their customers. The coefcient on the negative industry adjusted growth rate MDGRS_N is negative but not signicant. The coefcient on current period protability, which is proxied by the gross margin ratio MDGRM, is positive and signicant p-value 1 percent, indicating that protable rms extend more generous credit terms to their customers. The coefcient on the square of the gross margin MDGRM_SQ is also positive and signicant p-value 1 percent, indicating that the level of accounts receivable is a convex function of the current period gross margin. The coefcient on age LAGE is positive contrary to our expectations but not signicant. Most importantly, consistent with H1, the coefcient on the loss ratio LOSS is positive and signicant p-value 1 percent, indicating that rms that experience a sequence of negative earnings report relatively higher ratios of accounts receivable to sales than more protable rms. The results of the regression where we compute the loss ratio based on cash ows from operations are very similar, and the coefcients on the control variables are almost identical to those reported in the Earnings column. Most importantly, the coefcient on the loss ratio LOSS is also positive and signicant p-value 1 percent, again indicating that rms that experience a sequence of negative operating cash ows report relatively higher ratios of accounts receivable to sales.22 Panel B of Table 3 replicates Panel A of Table 3 for loss ratio benchmarks of 0.5, 0.7, and 0.9. The loss ratio LOSS_D variable in Panel B is a dummy variable with a value of 1 if the rm's loss ratio is greater than the specied benchmark and 0 otherwise. We estimate the regressions for both the earnings-based loss ratio and the cash ow-based loss ratio. For example, the third sixth column from the left shows the estimation results of the model where LOSS_D takes on the value 1 if the earnings cash ows based loss ratio is greater than 0.5. In order to conduct the median adjusted analysis, we eliminate four-digit SIC codes containing fewer than three rms with loss ratios less than the benchmark ratio. Panel B of Table 3 shows that the estimated coefcients on the loss ratio dummies LOSS_D are positive and statistically signicant across all loss ratio benchmarks, again indicating a positive association between the ratio of accounts receivable to sales and the incidence of losses. The coefcients on the control variables are consistent with those reported in Panel A: the coefcients on size, positive growth 22 In order to check the robustness of our results to different specications, we conducted the following sensitivity analyses: 1 we estimated the regressions omitting the square variables, 2 we included Altman's Z in order to control for the risk of bankruptcy, 3 we replaced the net income margin with the ratio of cash ows from operations to total sales, 4 we eliminated rm-years with an Altman's Z ratio less than 1.8, 5 given the potential correlation between the loss ratio and the gross income margin ratio, we omitted the gross income margin ratio and the square of the gross income margin ratio from the independent variables, and 6 we estimated the model using raw variables instead of industry median adjusted variables. The results across all specications are qualitatively similar to those presented in the tables. Auditing: A Journal of Practice & Theory November 2008 American Accounting Association 16 Auditing: A Journal of Practice & Theory American Accounting Association TABLE 3 The Determinants of Accounts Receivable (standard errors in parentheses) Panel A: Loss Ratio as Continuous Variablea MDARSit = 0 + 1LOSSit + 2LSIZEit + 3LAGEit + 4LAGE_SQit + 5MDGRS_Pit + 6MDGRS_Nit + 7MDGRM it + 8MDGRM_SQit + year dummies + it . Variable Predicted Sign Intercept / Earnings 0.010 0.008 LOSS 0.024*** 0.005 LSIZE 0.004*** 0.000 LAGE Cash Flows 0.040*** 0.008 0.06*** 0.005 0.006*** 0.000 LAGE_SQ 0.004 0.009 0.006 0.007 0.002 0.002 MDGRS_P 0.021*** MDGRS_N 0.011 November 2008 0.008 MDGRM 0.036*** 0.009 0.001 0.014*** 0.003 0.003 0.008 0.046*** 0.009 (continued on next page) Callen, Robb, and Segal 0.003 0.003** Revenue Manipulation and Restatements by Loss Firms Auditing: A Journal of Practice & Theory TABLE 3 (continued) Variable MDGRM_SQ Predicted Sign Earnings Cash Flows 0.033*** 0.031*** 0.010 n 0.010 21986 Adj. R 2 0.015 21985 0.030 Panel B: Loss Ratio as a Dummy Variableb MDARSit = 0 + 1LOSS_Dit + 2LSIZEit + 3LAGEit + 4LAGE_SQit + 5MDGRS_Pit + 6MDGRS_Nit + 7MDGRM it + 8MDGRM_SQit + year dummies + it . Earnings-Based Loss Ratio Greater Than or Equal to: Expected Sign Intercept / 0.5 0.7 0.9 0.004 0.002 0.002 0.006 LOSS_D 0.008*** 0.002 LSIZE 0.003*** 0.000 0.006 0.010*** 0.002 0.003*** 0.000 0.008 0.011*** 0.004 0.003*** 0.000 0.5 0.7 0.018*** 0.017*** 0.006 0.028*** 0.002 0.005*** 0.000 0.9 0.007 0.007 0.030*** 0.003 0.024*** 0.004 0.004*** 0.000 0.004 0.003*** 0.000 (continued on next page) 17 November 2008 American Accounting Association Variable Cash Flows-Based Loss Ratio Greater Than or Equal to: 18 Auditing: A Journal of Practice & Theory American Accounting Association TABLE 3 (continued) Earnings-Based Loss Ratio Greater Than or Equal to: Variable Expected Sign 0.5 0.7 0.006 0.007 LAGE LAGE_SQ MDGRS_P 0.022*** MDGRS_N 0.021*** MDGRM MDGRM_SQ 0.005 0.003*** 0.001 0.003 0.006 0.039*** 0.005 0.062*** 0.006 n Adj. R 0.005 0.003*** 0.001 0.020*** 0.003 0.019*** 0.007 0.036*** 0.005 0.050*** 0.006 Cash Flows-Based Loss Ratio Greater Than or Equal to: 0.9 0.010* 0.006 0.004*** 0.001 0.021*** 0.003 0.021*** 0.008 0.024*** 0.006 0.030*** 0.005 0.5 0.006 0.005 0.003*** 0.001 0.017*** 0.003 0.014*** 0.006 0.046*** 0.005 0.057*** 0.006 0.7 0.9 0.012*** 0.005 0.006 0.004*** 0.001 0.004*** 0.001 0.018*** 0.003 0.019*** 0.003 0.015*** 0.007 0.018*** 0.008 0.041*** 0.005 0.031*** 0.006 0.044*** 0.005 0.010 0.029*** 0.005 20,561 2 18,484 14,425 19,217 16,920 13,702 0.015 0.014 0.012 0.023 0.021 0.014 (continued on next page) Callen, Robb, and Segal November 2008 ***, ** Indicates signicance at the 1 percent and 5 percent levels, respectively. To reduce the effect of outliers, we remove the top and bottom percentile of growth in sales and gross prot scaled by sales. n is the number of observations. i t is the rm time index. Revenue Manipulation and Restatements by Loss Firms Auditing: A Journal of Practice & Theory TABLE 3 (continued) Panel A shows the estimation results of the accounts receivable model. LOSS in the Earnings Cash Flows column is the proportion of years in which the rm reports negative net income before extraordinary items negative cash ows from operations from year t5 to year t+3 inclusive. The standard errors are adjusted for rm-level clustering. b Panel B shows the estimation results of the accounts receivable model where the loss variable is dened as a dummy labeled LOSS_D equal to 1 if the loss ratio exceeds the benchmark, and 0 otherwise. The benchmarks are provided at the top of the columns. For example, the third column from the left shows the estimation results of the model where LOSS_D takes the value of 1 if the earnings based loss ratio is greater than 0.5. The standard errors are adjusted for White's 1980 heteroscedasticity correction. Variable Denitions: MDARS four-digit SIC median-adjusted ratio of accounts receivable to sales; LSIZE natural log of total assets; LAGE natural log of the rm's age; LAGE_SQ square of LAGE; MDGRS_P four-digit SIC median-adjusted growth in sales, computed as the difference between the rm's growth rate in sales and the median growth rate in the industry if positive, and 0 otherwise; MDGRS_N four-digit SIC median-adjusted growth in sales, computed as the difference between the rm's growth rate in sales and the median growth rate in the industry if negative, and 0 otherwise; MDGRM four-digit SIC median-adjusted gross prot scaled by total sales, computed as the difference between the rm's gross prot margin and the median gross prot margin in the industry; and MDGRM_SQ square of MDGRM. a 19 November 2008 American Accounting Association 20 Callen, Robb, and Segal rate in sales, gross margin, and the square of gross margin are positive and signicant, whereas the coefcient on negative growth rate in sales is negative and signicant. Overall, the ndings in Table 3 indicate that the industry-adjusted ratio of accounts receivable to sales increases with the loss ratio after controlling for size, age, growth, and current period protability. This result is consistent with our hypothesis that loss rms overstate revenues in order to inate their market values.23 Restatements and Comparing Revenue-Restaters with Nonrestaters Table 4, Panel A, shows statistics on the sample of restatements by year.24 The panel shows that the overall number of restatements consistently increases over time from ve in 1993 to 324 in 2000 before beginning to decline in 2001. There were 596 restatements that involve revenues and 931 restatement cases that do not involve revenues. The RES_RATIO column shows the ratio of restatement cases to the total number of rms in Compustat. The data suggest that the relative proportion of restated years has increased from 0.06 percent in 1993 to 3.7 percent in 2000. Untabulated results show that revenue manipulators are scattered around 54 industries based on four-digit SIC codes, indicating that revenue manipulation is not an industry-specic phenomenon. Table 4, Panel B, compares key variables between companies that restated their nancial statements due to revenue manipulation and companies that did not restate their nancial statements. This panel indicates that companies that restated their revenues have signicantly higher market value of equity, a higher loss ratio, and a higher industry-adjusted ratio of accounts receivable to sales. The panel also shows that there is no signicant difference in rm age and variability of returns. Finally the difference in the median Altman's Z ratio is marginally signicant p-value = 0.097. These ndings are consistent with our hypotheses that revenue manipulation ows through accounts receivable and that there is a positive link between revenue manipulation and the loss ratio. Overall, Table 4 indicates that the incidence of revenue restatements increased during the 1990s, and that companies that restate revenues have higher loss ratios and accounts receivable to sales ratios than do companies that do not restate their nancial statements. Below, we examine whether the ex ante probability of revenue manipulation increases with the loss ratio and the accounts receivable to sales ratio after controlling for the credit policy of the rm. The Probability of Revenue Manipulation and the Loss Ratio Table 5 shows the results of the \"partial observability\" two-stage probit model. The Earnings column shows the estimation results where the loss ratio is computed based on net income, and the Cash Flows column shows the estimation results where the loss ratio is computed based on cash ow from operations. The rst two columns of results measure the loss ratio as the proportion of years in which the company reported negative net income before extraordinary items cash ows from operations from year t5 to year t+3 inclusive. We also re-estimate the model with an \"historical\" loss ratio LOSS_HIST measured from year t5 to year t in the last two columns of Table 5. All regressions are signicant at less than the 1 percent level based on the Wald and Lagrange 23 24 Loss rms are likely prone to selling or securitizing their receivables because of their nancing needs. If anything, the sale of receivables biases against our nding that loss rms over-invest in receivables. Compustat does not provide data on the sale of receivables or securitizations and therefore we are unable to account for them in our analysis. Note that the sample of restatements consists of all restated years. The number of restated years used in subsequent analyses is smaller due to the lack of Compustat or CRSP data for many of the companies in the restatement sample. Auditing: A Journal of Practice & Theory American Accounting Association November 2008 Revenue Manipulation and Restatements by Loss Firms 21 TABLE 4 Restatement Statistics Panel A: Number of Restatementa Year REV_RES EXP_RES ALL_RES RES_RATIO 1993 3 2 5 0.1% 1994 8 11 19 0