A Reexamination of Real Earnings Management from a Firm-Specific Time-Series Perspective. E. SCOTT JOHNSON Virginia Tech University

Transcription

1 A Reexamination of Real Earnings Management from a Firm-Specific Time-Series Perspective E. SCOTT JOHNSON Virginia Tech University T. TAYLOR JOO New Mexico State University MICHAEL D. STUART Vanderbilt University May 2015 Early draft. Please do not cite. Keywords: Real earnings management, Earnings management, Earnings manipulation, Discretionary expenses, Firm-specific forecasts, Managerial ability. Data Availability: Data are publicly available from sources identified in the text. We thank Brooke Beyer, Bowe Hansen, Jing Huang, James Myers, Mitch Oler, Velina Popova, Sarah Stein, Michael Wolfe, and participants at the 2014 Arkansas Research Conference for helpful comments and suggestions. We also thank Peter Demerjian for generously providing access to MA-Score data.

2 A Reexamination of Real Earnings Management from a Firm-Specific Time-Series Perspective ABSTRACT: Managers use real earnings management (REM) to influence reported earnings through the manipulation of real business activities. Extant literature generally employs crosssectional analysis to identify REM, but newer studies suggest that these traditional REM measures are severely mis-specified. In this study, we employ a new methodology that utilizes firm-specific time-series characteristics of discretionary expenses to identify REM. Using this new REM measure, we find that firms engage in REM to meet or just beat zero earnings and to meet or just beat consensus analyst forecasts, a finding that is not strongly supported by traditional REM measures. Prior literature also suggests that some managers and researchers believe that REM is myopic and imposes real costs on firms, but there is mixed empirical evidence regarding the impact of REM on future operating performance. We find that firms that opportunistically engage in REM perform, on average, neither better nor worse in the future, a finding consistent with survey responses suggesting that managers are careful not to sacrifice future operating performance in order to meet or just beat earnings benchmarks. Additional analysis shows that firms with high ability managers that engage in REM have better future operating performance.

3 1. Introduction Prior literature suggests that managers engage in earnings management in order to report earnings that meet or just beat certain benchmarks. Earnings management is generally classified as either accruals management (AM) or real earnings management (REM). 1 AM refers to the use of discretionary accruals, either within or outside of generally accepted accounting principles (GAAP), to opportunistically influence earnings. Common examples of AM include underestimating bad debt and accelerating the recognition of sales. REM refers to managerial decisions that influence earnings through real business activities. Roychowdhury (2006, 337) defines REM as departures from normal operational practices, motivated by managers desire to mislead at least some stakeholders into believing certain financial reporting goals have been met in the normal course of operations. Examples of REM include opportunistically cutting discretionary expenses such as advertising, research and development (R&D), and selling, general and administrative (SG&A). Extant literature generally employs cross-sectional analysis to identify firms engaging in REM (e.g., Roychowdhury 2006; Cohen, Dey and Lys 2008; Cohen and Zarowin 2010; McInnis and Collins 2011; Zang 2012; Zhao, Chen, Zhang and Davis 2012), but newer studies suggest that traditional REM measures are severely mis-specified (Cohen, Pandit, Wasley and Zach 2014; Siriviriyakul 2014). Additionally, empirical evidence on the impact of REM on future operating performance is mixed. In this study, we employ a new methodology to identify REM that utilizes firm-specific time-series characteristics of discretionary expenses. We then examine the effect of REM on future operating performance. Emerging research questions the validity of the REM proxies used in the majority of extant studies. Cohen et al. (2014) find that traditional REM measures used in the literature are 1 Throughout this study, and consistent with prior literature, we use the terms real earnings management, real activities management, and real activities manipulation interchangeably. 1

4 severely mis-specified. Similarly, analysis of current REM models by Siriviriyakul (2014) suggests the presence of omitted variables. This mis-specification raises the possibility that the findings of prior research are not robust to alternate specifications. Prior research utilizes crosssectional analysis to estimate abnormal discretionary expenses and identify firms engaging in real earnings management. In practice, however, financial statement analysis relies on firms individual characteristics (Stickney, Brown and Wahlen 2003; White, Sondhi and Fried 2003; Lundholm and Sloan 2009; Penman 2013). Additionally, Francis and Smith (2005) argue that earnings analysis should be performed using firm-specific time-series estimations because the persistence of a firm s earnings components are unlikely to depend on the persistence of other firms similar components. Anecdotal evidence suggests that managers engage in REM to meet certain benchmarks. Based on survey results, Graham, Harvey and Rajgopal (2005) find that 80% of chief financial officers (CFOs) would decrease discretionary spending in order to meet an earnings benchmark and 55.3% would delay starting a new project to meet an earnings benchmark, even if the delay entailed a small sacrifice in value. However, respondents indicate that they would take real actions to boost earnings only if the real sacrifices are not too large. This suggests that managers understand the potential costs associated with REM, but it also suggests that they attempt to mitigate those costs even while engaging in REM to meet earnings benchmarks. Theoretical research suggests that REM imposes value-destroying costs on firms. Ewert and Wagenhofer (2005, 1102) state that real earnings management imposes real costs on the firm. Additionally, Jensen (2005, 8) states that when numbers are manipulated to tell the markets what they want to hear (or what managers want them to hear) rather than the true state of 2

5 the firm it is lying, and when real operating decisions that would maximize value are compromised to meet market expectations, real long-term value is being destroyed. However, prior empirical research provides mixed evidence on the impact of REM on future performance. On the one hand, some studies find that REM imposes value-destroying costs. Cohen and Zarowin (2010) find that seasoned equity offering (SEO) firms often engage in both forms of earnings management (i.e., REM and AM), and of the firms that suffer post-seo declines in performance, the declines are more severe for firms engaging in greater levels of REM than for firms engaging in greater levels of AM. Additionally, Francis, Hasan and Li (2011b) find that firms engaging in REM are more likely to suffer subsequent stock price crashes than firms not engaging in REM, and Zhao et al. (2012) find that abnormally low discretionary expenses and production costs lead to lower future performance in the absence of just meeting earnings benchmarks. On the other hand, some studies do not find a negative impact from REM on future performance. For instance, Taylor and Xu (2010) find that REM does not lead to a decline in subsequent operating performance, and Gunny (2010) provides evidence that firms that engage in REM to meet certain earnings benchmarks have better future operating performance than firms that either miss the benchmarks or that meet the benchmarks without engaging in REM. This study is dually motivated by (1) the need to develop an alternate method for identifying REM, given the growing evidence that traditional REM models are mis-specified, and (2) the mixed empirical evidence regarding the impact of REM on future operating performance. To conduct our analyses, we construct a preliminary sample of 31,782 firm-year observations (representing 4,383 unique firms) between 1995 and We first employ a firm- 3

6 specific time-series methodology to estimate expected discretionary expenses (i.e., the sum of advertising, R&D, and SG&A costs). We then calculate the abnormal level of discretionary expenses as the difference between actual and expected discretionary expenses. Because prior research suggests that traditional REM measures are severely mis-specified, we examine the properties of our firm-specific time-series measure of abnormal discretionary expenses (Firm- Specific REM) and compare these properties to the traditional REM measure (Traditional REM). Siriviriyakul (2014) states that the Traditional REM measure is highly persistent, which we confirm through an examination of the sign (positive or negative) of the abnormal discretionary expenses over time. The Firm-Specific REM measure proves to be much less persistent and exhibits the variation we would expect to see if managers are making myopic decisions about discretionary expenses in order to meet certain earnings benchmarks. Next, we examine whether Firm-Specific REM is associated with the likelihood of meeting certain benchmarks. Because we compare our measure to the Traditional REM measure first developed by Roychowdhury (2006), we use the same benchmarks utilized in his study: zero earnings and the consensus analyst forecast. 2 We first use the Firm-Specific REM measure and find that it is negatively associated with the likelihood of meeting or just beating both zero earnings and the consensus analyst forecast, at the 5% and 1% level, respectively. 3 This result suggests that the Firm-Specific REM measure is identifying manipulations in real activities to meet these important benchmarks. Next, we use the Traditional REM measure and find that it is negatively associated with the likelihood of meeting or just beating zero earnings, but only at the 2 It is important to note that although Roychowdhury (2006) does test for REM to meet or just beat the consensus analyst forecast, he does not use the same measure (Traditional REM) that he utilizes to test for REM to avoid losses. Instead, he uses an alternate specification that is based on a performance-matching technique advocated by Kothari, Leone, and Wasley (2005). 3 Note that lower levels of abnormal discretionary expenses result in higher net income. 4

7 10% level. We do not, however, find an association between Traditional REM and meeting or just beating the consensus analyst forecasts. Overall, these results suggest that the Firm-Specific REM measure is more accurately identifying REM than the Traditional REM measure. In supplemental analyses, we classify firms with negative abnormal discretionary expenses in year t but positive abnormal discretionary expenses in year t-1 and t+1 as blip suspect firms and examine whether these firms engage in REM to meet or just beat zero earnings or prior year s earnings. 4 A common criticism of REM research is that negative abnormal discretionary expenses are simply permanent cost cutting decisions made by management to improve future profitability or are part of the normal operations of the firm. In other words, it may not be real earnings management at all, but rather good management, a change in management philosophy, or inherent firm differences. By identifying firms that have positive abnormal discretionary expenses in year t-1, followed by negative abnormal discretionary expenses in year t, and then returning to positive abnormal discretionary expenses in year t+1, we attempt to decrease the likelihood that our suspect firms are simply cutting costs in an effort to improve future profitability. We utilize this blip in discretionary expenses to identify firms that strategically manipulate real activities in year t to meet certain benchmarks. As predicted, we find a positive and statistically significant association, at the 1% level, between blip firms, determined by utilizing the Firm-Specific REM measure, and the likelihood of meeting or just beating zero earnings or last year s earnings. However, we do not find a statistically significant association between blip firms, determined using the Traditional REM measure, and the same benchmarks. This test provides additional evidence of the validity of the Firm-Specific REM 4 Change in earnings is an important benchmark in earnings management literature that is addressed, but not tested, in Roychowdhury (2006). We add this benchmark to our tests as a robustness check for the validity of our measure of abnormal discretionary expenses, as we believe that managers may engage in REM to meet or just beat last year s earnings. 5

8 measure and introduces a new way of identifying REM firms through outwardly-visible indicators. We also replicate Gunny (2010) to examine the impact of REM on future performance. In contrast to the findings in Gunny (2010), we do not find evidence that firms that engage in REM to report small positive earnings perform better in the future using the new Firm-Specific REM measure. However, we also do not find evidence that firms that engage in REM to report small positive earnings perform worse in the future. These results suggest that, on average, engaging in REM to meet earnings benchmarks has no discernible impact on future operating performance. This may be due to the fact that, while managers engage in REM to meet earnings benchmarks, they are careful not to sacrifice future operating performance. This finding is consistent with the Graham et al. (2005) survey findings that managers weigh the costs and benefits of engaging in REM. In additional analyses, we find some evidence that firms with high ability managers that engage in REM perform better in the future, suggesting that either high ability managers are better at mitigating the negative impact of REM on future performance or that behavior that appears to be REM, when engaged in by firms with high ability managers, is actually just good management. Although there are other studies examining the consequences of REM, there is no consensus on its effect on future performance, and given recent evidence suggesting that traditional REM models are severely mis-specified, we believe a re-examination of the topic is warranted. This study contributes to the REM literature by (1) developing a new time-series firm-specific empirical method to identify income-increasing REM, and (2) reexamining the impact of REM on future performance using this new methodology. 6

9 Our results should be of interest to investors, because they reveal that, contrary to prior research (i.e., Gunny 2010), engaging in REM to meet earnings benchmarks does not appear to lead to better future operating performance. Additionally, we find that REM firms do perform better in the future when led by high ability managers. Finally, our results should be of interest to accounting researchers as they consider the measurement of real earnings management and evaluate the effect of this behavior on future performance. Our paper proceeds as follows. Section 2 reviews prior evidence from the REM literature and develops our hypotheses. Section 3 describes our sample and research design. Section 4 presents our empirical results. Section 5 concludes. 2. Background and Hypothesis Development Anecdotal evidence suggests that managers engage in real activities manipulation to influence earnings. Graham et al. (2005) provide evidence, from surveys, that managers are willing to manipulate real activities, even if the manipulations have the potential to reduce firm value, as long as the negative impact is only a small sacrifice in value. In fact, Graham et al. (2005, 29) indicate that several CFOs argue, you have to start with the premise that every company manages earnings. Of particular importance to our study, Graham et al. (2005) report that a larger number of managers admit to reducing discretionary expenses than to engaging in other methods of real activities manipulation. Extant REM studies utilize the methodology developed by Roychowdhury (2006) to calculate abnormal discretionary expenses and identify firms engaging in REM. However, Cohen et al. (2014) find that this Traditional REM measure is severely mis-specified. Specifically, they find that the Traditional REM measure indicates the presence of REM far too often in randomly 7

10 constructed samples. Similarly, analysis of the Traditional REM model by Siriviriyakul (2014) suggests the presence of omitted variables. She states that [b]ecause real earnings management is a departure from normal activity, if the REM proxies truly capture REM activity, they should exhibit subsequent reversal (Siriviriyakul 2014, 3). Contrary to her expectation, she finds that the REM proxies are highly persistent, which she interprets as evidence of omitted correlated variables. Whereas prior research utilizes cross-sectional analysis to identify firms engaged in real earnings management, in practice, financial statement analysis is conducted using individual firm characteristics and fundamentals (e.g., Stickney et al. 2003; White et al. 2003; Lundholm and Sloan 2009; Penman 2013). In addition, Francis and Smith (2005) argue that earnings analysis should be performed using firm-specific time-series estimations because the persistence of a firm s earnings components are unlikely to depend on the persistence of other firms similar components. Therefore, we develop a firm-specific time-series model that does not rely on information about other firms within an industry and does not make the assumption that the persistence of earnings components (e.g., discretionary expenses) are the same for all firms. Thus, our model provides a method for more accurately forecasting a firm s expected level of discretionary expenses, which allows us to identify firms that deviate from those expected levels. Early empirical evidence of REM focuses on the opportunistic reduction of R&D expenditures to reduce expenses. Baber, Fairfield and Haggard (1991) find evidence suggesting that R&D investment decisions are influenced by managers concerns about reported earnings, and Bushee (1998) finds that managers reduce R&D expenditures to meet short-term earnings benchmarks. Dechow and Sloan (1991) provide evidence that CEOs reduce R&D expenditures to increase current earnings near the end of their tenure. Finally, Bens, Nagar and Wong (2002) 8

11 provide empirical evidence that firms engage in REM for purposes other than meeting a specific earnings benchmark. They find that firms experiencing significant employee stock option exercises shift resources away from real investments, including R&D, toward the repurchase of their own stock. Consistent with Bushee (1998), they also suggest that this imposes real costs on the firm. More recent research finds evidence of REM employing multiple types of real activities manipulation. Roychowdhury (2006) develops empirical methods to detect these manipulations in large samples and provides evidence that managers reduce discretionary expenses to avoid reporting annual losses. Using a modified version of his abnormal discretionary expense measure, Roychowdhury (2006) also finds less robust evidence that firms use REM to meet consensus analyst forecasts. Cohen et al. (2008) confirm the findings of Roychowdhury (2006) and document that, while AM declined after the passage of the Sarbanes-Oxley Act (SOX) in 2002, REM increased significantly. This suggests that firms shifted away from using AM to REM after SOX. Badertscher (2011) finds that managers engage in earnings management to support overvalued equity and that they generally exhaust AM strategies before moving to REM. Finally, Zang (2012) finds that managers use real activities management and accrual-based earnings management as substitutes and engage in real activities management when its relative cost is low. Prior research finds mixed results on the consequences of REM. Cohen and Zarowin (2010) provide evidence that managers engage in earnings management around seasoned equity offerings (SEOs) and that declines in post-seo performance due to REM are more severe than declines due to AM. Their evidence is important, because it shows that post-seo underperformance is driven not just by accruals reversal, but also reflects real consequences of 9

12 operational decisions made to manage earnings. Bhojraj, Hribar, Picconi and McInnis (2009) show that firms that beat analyst forecasts using both REM and AM have worse future operating performance and returns, in the subsequent three years, than firms that just miss analyst forecasts but do not engage in any form of earnings management. Their study, however, does not attempt to disentangle the effect of AM and REM individually, but rather uses a measure that incorporates both forms of earnings management. Francis et al. (2011b) find that firms engaging in REM are more likely to suffer subsequent stock price crashes than firms not engaging in REM. Additionally, Kim and Park (2013) investigate the impact of REM on auditor s clientretention decisions and find that auditor resignations are more likely when firms manage earnings through opportunistic operating decisions. These results illustrate the negative consequences and risk associated with REM. Alternatively, Taylor and Xu (2010) find that REM does not lead to a significant decline in subsequent operating performance. Similarly, Gunny (2010) provides evidence that firms that engage in REM to meet or just beat earnings benchmarks have better operating performance, in the subsequent three years, than firms that either miss the earnings benchmarks or meet the earnings benchmarks but do not engage in REM. She suggests that engaging in REM is not opportunistic, but rather a way for managers to produce current-period benefits and to signal superior future earnings. Zhao et al. (2012) find that although abnormal real activities, in general, are associated with lower future performance, real earnings management intended to just meet earnings benchmarks is associated with higher future performance, consistent with real earnings management conveying a signal of superior future performance. One possible reason for the inconsistent results in prior research is that different managers could be engaging in real activities manipulation to achieve different goals, or perhaps 10

13 that the association between REM and future performance varies with the ability levels of managers. It is possible that behavior that appears to be REM is actually good management or similarly that REM may be used to signal strong future performance. This may be particularly true for firms with high managerial ability. Demerjian, Lev, and McVay (2012) develop a managerial ability score that is calculated in two-stages: the first stage provides an estimate of firm-level operational efficiency and the second stage controls for various firm characteristics to isolate the effects of the manager. 5 Utilizing this managerial ability measure allows us to determine whether managers with higher ability are better able to mitigate the costs of REM on future performance and signal stronger future performance, or whether they are perhaps not engaging in REM at all, but rather engaging in good management that leads to better future performance. Demerjian, Lev, Lewis, and McVay (2013) find that better managers are less likely to engage in earnings management when it reduces financial reporting quality, and Demerjian, Lewis-Western, and McVay (2015) find that better managers are able to intentionally smooth earnings in a way that is low-cost to their firms. These findings suggest that managers may also be able to engage in REM without negatively impacting future operating performance. The lack of consensus in prior literature regarding the effects of REM on future operating performance, and the need for a new methodology of identifying firms suspected of engaging in REM, motivate this study. 3. Research Design Managers use earnings management to meet certain earnings benchmarks, achieve internal bonus targets, appear more attractive prior to initial public offerings (IPOs) and SEOs, 5 The MA-Score data developed by Demerjian et al. (2012) is publicly available for download at 11

14 avoid debt covenant violations, meet regulatory requirements, avoid taxes, and for numerous other reasons (Graham et al. 2005). Examples of REM include cutting discretionary expenses such as advertising, R&D, and SG&A, in order to achieve desired outcomes. Theoretically, REM should be an infrequent event that reverses as firms return to normal levels of activity in subsequent years. REM, however, is difficult to measure because we cannot observe the intentions of management. Prior research (e.g., Roychowdhury 2006; Gunny 2010) uses crosssectional models to estimate the expected level of discretionary expenses and then compares the estimate to actual discretionary expenses to determine abnormal discretionary expenses. These cross-sectional models are estimated by year and industry. Cohen et al. (2014) and Siriviriyakul (2014) express concern that the traditional models used to measure real earnings management are severely mis-specified and do not accurately identify firms that depart from normal levels of expenses. Furthermore, because of the importance of firm-specific characteristics in financial statement analysis, the analysis is performed at the firm level in practice (e.g., Stickney et al. 2003; White et al. 2003; Lundholm and Sloan 2009; Penman 2013). Consistent with this argument, Francis and Smith (2005) suggest that earnings analysis should be performed using firm-specific time-series estimations because the persistence of a firm s earnings components are unlikely to depend on the persistence of other firms earnings components. To address the concerns raised in prior literature, we develop a new methodology of measuring REM. Specifically, we use firm-specific time-series estimations. 3.1 Firm-specific persistence of discretionary expenses Firms that deviate from their normal levels and report abnormally low levels of discretionary expenses to meet earnings targets are likely engaging in REM. In our analysis, we 12

15 focus on advertising, R&D, and SG&A expenses, which prior research suggests are commonly manipulated to meet certain earnings benchmarks. We include advertising and R&D because they are expensed as incurred and because the future benefit of these expenditures may be uncertain. As a result, firms may cut advertising and R&D if they feel pressure to hit important earnings benchmarks in a given year, especially if the benefit from the expenses will not be realized until future periods. We include SG&A because prior research suggests that SG&A costs are sticky (Anderson et al. 2003), 6 and a significant decrease in SG&A costs in a particular year is potentially suspicious. Furthermore, SG&A frequently includes costs such as employee training, maintenance and travel, etc., which are highly discretionary in nature. To identify abnormal changes in a firm s discretionary expenses, we sum advertising, R&D, and SG&A expenses and estimate the following firm-specific regression: DisExpt = α + β1disexpt-1 + εt, (1) where: DisExpt = the sum of advertising, R&D, and SG&A expenses, scaled by total assets at the beginning of year t. 7 We estimate model (1) for each firm-year using a 10-year rolling window. Using the estimated values of the intercept and β1 from model (1), we derive the expected level of discretionary expenses in year t. The abnormal level of discretionary expenses in year t (AbnDisExpt) is calcualted as the difference between the actual level of discretionary expenses in year t and the predicted level of discretionary expenses in year t. Because our calcualtion of abnormal discretionary expenditures is performed at the firm level, our measure addresses the 6 Anderson et al suggest that SG&A costs are sticky because they decrease less when sales decrease than they increase when sales increase by an equal amount. This is attributed to both the fixity of certain costs and managerial decisions to avoid cutting slack resources in anticipation of a rebound in sales. 7 In order to maximize the number of sample observations, and following prior research, we set advertisting expense and R&D expense to zero if missing. 13

16 concern that some firms are inherently different from other firms in their respective industry. It is important to note that abnormally low levels of discretionary expenditures are income increasing. 3.2 Identification of suspect firms Prior research suggests that firms face intense pressure to meet certain earnings benchmarks, and we expect that firms engage in REM to meet these benchmarks. Following Roychowdhury (2006), we utilize the following benchmarks to identify firms that are more likely to be engaging in REM: reporting earnings that meet or just beat zero and meeting or just beating the consensus anlayst forecast. With respect to the positive earnings benchmark, prior literature (e.g., Burgstahler and Dichev 1997; Roychowdhury 2006) finds a significant spike in the number of firms reporting earnings in the interval just to the right of zero (i.e., earnings scaled by lagged total assets between and 0.005) when compared with the interval immediately to the left of zero (i.e., firms that just missed reporting positive earnings). This suggests that the spike in earnings just to the right of zero is the result of firms managing earnings to report profits. Therefore, consistent with prior research, we classify observations that fall into the interval immediately to the right of zero as suspect firm-years (Bench_NIt). 8 Prior research finds that firms that meet or just beat consensus analyst forecasts suffer severe stock price delcines (Skinner and Sloan 2002), and these penalties may provide ample incentive for firms to engage 8 Focusing on the interval just to the right of zero presents some potential problems. First, firms that meet or just beat zero earnings are not likely to be the only firms that engage in real earnings management. Specifically, firms may use real earnings management to report earnings growth, achieve internal bonus targets, appear more attractive prior to initial public offerings (IPOs) and secondary equity offerings (SEOs), avoid debt covenant violations, meet regulatory requirements, avoid taxes, and for numerous other reasons. Second, firms with pre-managed earnings safely to the right of zero have an incentive to use income-decreasing real earnings management to create reserves, thus making it easier to achieve earnings targets in the future. This income-decreasing REM could place the firms just to the right of zero. These problems lower the power of the subsequent empirical tests; however, they are shortcomings of all earnings management studies that utilize these benchmarks to identify suspect firms. 14

17 in REM. Therefore, consistent with prior research, we classify observations that meet or just beat consensus analyst forecasts by one cent or less as suspect firm-years (Bench_AFt). 3.3 Model testing the association between REM and meeting earnings benchmarks To examine whether firms use REM to meet or just beat earnings benchmarks, we estimate the following model from Gunny (2010): AbnDisExpt = α + β1sizet + β2mtbt + β3roat + β4bencht + εt (2) where: AbnDisExpt = Firm-Specific REM or Traditional REM; Firm-Specific_REM = the difference between the actual and expected level of discretionary expenses in year t derived from model (1); Traditional_REM = the difference between the actual and expected level of discretionary expenses derived following Roychowdhury (2006); Sizet = the natural log of total assets at the beginning of year t; MTBt ROAt Bencht Bench_NIt Bench_AFt = the ratio of the market value of equity to the book value of equity at the beginning of year t; = the difference between the firm-specific income before extraordinary items in year t, scaled by lagged total assets, and the median income before extraordinary items in year t, scaled by lagged total assets, for the same industry (two-digit SIC); = Bench_NIt or Bench_AFt. = an indicator variable equal to 1 if income before extraordinary items in year t, scaled by lagged total assets, is between 0 and 0.005, 0 otherwise; and = an indicator variable equal to 1 if the firm meets or just beats the final consensus analyst forecast before the fiscal year end by one cent or less, 0 otherwise. 15

18 Consistent with Roychowdhury (2006) and Gunny (2010), we include Sizet to control for size effects and MTBt to control for growth opportunities. Additionally, we include ROAt because real earnings management may be related to performance. We estimate model (2) in a pooled regression. All variables are winsorized at the top and bottom 1% of the sample distribution to mitigate the impact of outliers. Bencht is our variable or interest. We expect a negative coefficient on Bencht, indicating that firms that report small positive earnings or that meet or just beat consensus analyst forecasts report abnormally low levels of discretionary expenses, which is consistent with firms engaging in REM to meet earnings benchmarks. 3.4 Model testing the association between blip suspect firms and meeting earnings benchmarks We perform an additional test to validate that the Firm-Specific REM measure is capturing real earnings management. We create a blip variable for both the Firm-Specific REM and the Traditional REM measures. The blip variable is an indicator variable equal to 1 if abnormal discretionary expenses are negative in year t, but positive in both years t-1 and t+1. A common criticism of REM studies is that abnormally low discretionary expenses may represent a change in philosophy by managers to lower expenses moving forward, in order to produce better future operating performance, or may indicate that some firms are inherently different from other firms in their industry. This measure is intended to mitigate that criticism by identifying firms that have abnormally low discretionary expenses in a year that is surrounded by positive abnormal discretionary expenses. To examine the likelihood that firms with a blip are more likely to meet or just beat zero earnings or zero earnings growth benchmarks, we estimate the following model: SuspectMBt = α + β1blipt +β2sizet + β3mtbt + β4bignt + β5roat + β6losst + εt (3) 16

19 where: SuspectMBt Blipt Firm-SpecificBlipt TraditionalBlipt BigNt Losst = an indicator variable equal to 1 if income before extraordinary items, scaled by lagged total assets, is between 0 and or if the change in income before extraordinary items, scaled by lagged total assets, between years t-1 and t is between 0 and 0.005, 0 otherwise; = Firm-Specific_Blipt or Traditional_Blipt; = an indicator variable equal to 1 if Firm-Specific REM is negative in year t and positive in both years t-1 and t+1, 0 otherwise; = an indicator variable equal to 1 if Traditional REM is negative in year t and positive in both years t-1 and t+1, 0 otherwise; = an indicator variable equal to 1 if the firm s auditor is one of the Big N audit firms, 0 otherwise; = an indicator variable equal to 1 if the firm reported negative income available to common shareholders before extraordinary items in year t, 0 otherwise, and all other variables as previously defined. We run the model separately using the Firm-Specific_Blipt measure and then the Traditional_Blipt measure. A positive coefficient on Blipt would indicate that the likelihood of meeting earnings benchmarks in year t is greater for firms with an unexpected drop in discretionary expenses in that year. This test also provides evidence on whether the REM measures are effectively capturing the underlying real earnings management construct. 3.5 Model testing the association between REM and future performance Prior theoretical research suggests that REM is value-destroying and imposes real costs on firms (Ewert and Wagenhofer 2005). Furthermore, Jensen (2005) argues that when firms make non-value maximizing decisions, firm value is destroyed. However, empirical findings are mixed on the impact of REM on future operating performance. To examine the impact of real 17

20 earnings management on future operating performance we follow Gunny (2010) and estimate the following model: ROAt+1 or CFOt+1 = α + β1beatt +β2justmisst + β3bench_nit + β4neg_abndisexpt or Low_AbnDisExpt + β5bench_ni*(neg_abndisexpt or Low_AbnDisExpt ) + β6roat or CFOt + β7sizet + β8mtbt + β9returnt + β10zscoret-1+ εt, (4) where: ROAt+1 CFOt+1 Beatt JustMisst Bench_NIt Neg_AbnDisExpt Low_AbnDisExpt CFO t Returnt ZScoret-1 = the difference between the firm-specific income before extraordinary items in year t+1, scaled by lagged total assets, and the median income before extraordinary items in year t+1, scaled by lagged total assets, for the same industry (two-digit SIC); = the difference between the firm-specific cash flow from operations in year t+1, scaled by lagged total assets, and the median cash flow from operations in year t+1, scaled by lagged total assets, for the same industry (two-digit SIC); = an indicator variable equal to 1 if income before extraordinary items, scaled by lagged total assets, is greater than 0.005, 0 otherwise; = an indicator variable equal to 1 if income before extraordinary items, scaled by lagged total assets, is greater than or equal to and less than 0.000, 0 otherwise; = an indicator variable equal to 1 if income before extraordinary items in year t, scaled by lagged total assets, is between 0 and 0.005, 0 otherwise; = an indicator variable equal to 1 if AbnDisExpt is negative, 0 otherwise; = an indicator variable equal to 1 if AbnDisExpt is in the lowest quintile ranked by industry and year, 0 otherwise; = the difference between the firm-specific cash flow from operations in year t, scaled by lagged total assets, and the median cash flow from operations in year t, scaled by lagged total assets, for the same industry (two-digit SIC); = the buy and hold return of the firm over the 12 months of year t minus the buy and hold return of a portfolio of firms within the same CRSP decile during year t; and = a measure of bankruptcy risk calculated as 3.3*(pretax income/lagged total assets) + (sales/lagged total assets) *(retained earnings/lagged 18

21 all other variables as previously defined. total assets) + 1.2*((current assets current liabilities)/lagged total assets), and We include ROAt to control for firm performance and Sizet to control size. We include MTBt to control for growth opportunities. We include Returnt to control for the association between stock return performance and future earnings, and we include ZScoret to control for the financial health of the firm. We estimate model (4) in a pooled regression. We include industry and year fixed effects to control for time and industry effects in our sample. In addition, we control for potential time-series correlation by clustering standard errors by firm and year. All variables are winsorized at the top and bottom 1% of the sample distribution to mitigate the impact of outliers. Our coefficient of interest is β5, the interaction between Bencht and AbnDisExpt. Prior research suggests that real earnings management is potentially valuedestroying because firms sacrifice potential value creating activities at the expense of meeting earnings benchmarks. However, an alternate view exists. Gunny (2010) finds that firms that use real earnings management to meet earnings benchmarks have better future performance. She suggests that firms engage in real earnings management in order to attain benefits that allow the firm to perform better in the future and to signal stronger future performance. Additionally, Graham et al. (2005, 35) report that more than half of the CFOs they surveyed would delay a project even if it entailed a...small sacrifice in value. This suggests that, although managers admit to manipulating real activities to meet earnings benchmarks, they are cognizant of the potential costs and attempt to keep these costs as low as possible. In light of the two competing arguments provided by prior research, coupled with the survey evidence of Graham et al. (2005), we do not make a directional prediction on β5 and expect to find that using REM to meet earnings benchmarks has neither a positive nor negative impact on future operating performance. 19

22 Prior research also suggests that high ability managers are able to manipulate income, through intentional smoothing, in a way that imposes a very low cost on their firms (Demerjian et al. 2015). Additionally, Demerjian et al. (2013) find that high ability managers are less likely to engage in earnings management when it reduces financial reporting quality. Therefore, we use the managerial ability measure developed in Demerjian et al. (2012) to examine whether high ability managers who engage in REM are able to mitigate the potential negative impacts on future operating performance. We modify Gunny (2010) and estimate the following model: ROAt+1 or CFOt+1 = α + β1firm-specific_remt +β2hamt + β3firm-specific_remt*hamt + β4roat or CFOt + β5size + β6mtbt + β7returnt + β8zscoret-1+ εt, (5) where: HAMt = an indicator variable equal to 1 when the managerial ability score is above the median for the industry (two-digit SIC) and year, 0 otherwise, and all other variables as previously defined. Our coefficient of interest in this model is β3, the interaction between Firm- Specific_REMt and HAMt. If high ability managers are able to engage in REM without negatively impacting future performance, or if they are not engaging in REM at all, but rather prudently cutting costs as good managers, then we expect the coefficient on β3 to be positive. 3.6 Sample Selection To examine the relation between REM and future operating performance, we identify all firm-year observations from the Compustat database between 1995 and Our sample consists of firms with available data from the Center for Research in Security Prices (CRSP), firms where the sum of advertising, R&D, and SG&A expenses is non-zero, and firms with 20

23 enough data to calculate abnormal discretionary expenses using our firm-specific methodology. We develop our measure of high managerial ability from publicly available managerial ability score data. We eliminate firms in regulated industries (SIC codes 4400 to 5000) and banks and financial institutions (SIC codes 6000 to 7000). These restrictions result in a preliminary sample of 31,782 firm-year observations (representing 4,383 unique firms). We also winsorize all variables at the top and bottom 1% of the distribution to eliminate the impact of extreme observations. We perform the multivariate analyses that follow using the maximum number of observations with complete data available for each test. Because of this, the number of observations varies across specifications. 4. Results 4.1 Descriptive Statistics We report sample descriptive statistics in Table 1. The sample mean (median) of Firm- Specific_REMt is ( ) with a standard deviation of , indicating a wide variance of abnormal discretionary expenses among the sample firms. The sample mean (median) of Traditional_REM is ((0.0063) with a standard deviation of (0.2008). The mean (median) of the natural log of total assets (Sizet) is (5.8020) and the mean (median) of the market to book ratio is (1.6854). The mean (median) of firm industry-adjusted return on assets (ROAt) is (0.2030), and the mean (median) of firm industry adjusted cash flow from operations (CFOt) is (0.1122). Approximately 2% of our firm-year observations report income before extraordinary items, scaled by lagged total assets, between 0 and (Bench_NIt), and approximately 12% meet or just beat consensus analyst forecasts by one cent or less (Bench_AFt). Approximately 6% of our firm-year observations report income 21

24 before extraordinary items, scaled by lagged total assets, or the change in income before extraordinary items, scaled by lagged total assets, between 0 and (SuspectMBt). Approximately 6% of our firm-year observations have negative Firm-Specific REM in year t and positive Firm-Specific REM in years t-1 and t+1 (FirmSpecific_Blipt), and approximately 2% have negative Traditional REM in year t and positive Traditional REM in years t-1 and t+1 (Traditional_Blipt). Approximately 56% of our firm-year observations are during years where the firm was audited by a Big N auditor (BigNt), and approximately 30% are years with negative income (Losst). The mean (median) of one-year ahead return on assets (ROAt+1) is (0.2141) and the mean (median) of one-year ahead cash flow from operations (CFOt+1) is (0.1148)., Approximately 68% of our firm-year observations have income before extraordinary items, scaled by lagged total assets, greater than (Beatt), and approximately 1% have income before extraordinary items, scaled by lagged total assets, between and 0 of lagged assets (JustMisst). Approximately 56% of our firm-year observations have negative abnormal discretionary expenses (Neg_AbnDisExpt), and approximately 18% have abnormal discretionary expenses ranked in the lowest quitile their industry (Low_AbnDisExpt). The mean (median) 12 month buy and hold return (Returnt) is ( ). The mean (median) Z Score (ZScoret) is (1.9341). Finally, approximately 42% of our firm-year observations are under the direction of high ability managers. [Insert Table 1 here] In Table 2, we present Pearson correlations for select variables of our sample observations. We find that the correlations between our Firm-Specific measure of abnormal discretionary expenses (Firm-Specific_REMt) is negatively and significantly associated with Sizet, MTBt, ROAt, Firm-Specific_Blipt, Traditional_Blipt, Beatt, Neg_AbnDisExpt, and 22

25 Low_AbnDisExpt, and is positively and significantly associated with HAMt. We find that the correlations between the Traditional measure of abnormal discretionary expenses (Traditional_REMt) is negatively and significantly associated with ROAt, Bench_NIt, Suspect_MBt, Firm-Specific_Blipt, Traditional_Blipt, Beatt, JustMisst, Neg_AbnDisExpt, and HAMt, and is positively and significantly associated with Sizet, MTBt, Bench_AFt, and Low_AbnDisExpt. [Insert Table 2 here] In Figure 1, we examine and compare the persistence of the Firm-Specific REM measure with the persistence of the Traditional REM measure. Siriviriyakul (2014) contends that the Traditional REM measure is highly persistent and mis-specified, which is an indication of correlated ommitted variables. If a firm engages in REM to achieve certain earnings benchmarks, we would expect the firm to exhibit abnormally low levels of abnormal discretionary expenses in the manipulation years and then a return to higher levels of abnormal discretionary expenses as the expenses are shifted, or return to normal levels, in prior and subsequent periods. We present the percentage of surviving firms that have abnormal discretionary expenses with the same sign (either positive or negative) in all years prior, for both the Traditional REM measure and the Firm-Specific REM measure. We also compare the measures to the likelihood that abnormal discretionary expenses would exhibit the same sign (either positive or negative) under the assumption of random chance. The figure illustrates that the Traditional REM measure is highly persistent, a criticism echoed by Siriviriyakul (2014). In fact, using the traditional methodology, 13 out of the 21 firms (or 61.9%) that survive the entire eighteen year sample period have the same abnormal discretionary expenses sign. In other words, using the traditional methodology, nearly two-thirds of the surviving firms have either eighteen consecutive years of positive 23

26 abnormal discretionary expenses or eighteen consecutive years of negative abnormal discretionary expenses. By comparison, none of the sample firms have the same sign of abnormal discretionary expenses for the full eighteen year period using our firm-specific methodology. This is compared to the % chance that a firm would have the same sign of abnormal discretionary expenses for the full eighteen years under the assumption of random chance. Thus the Firm-Specific REM measure appears to alleviate the concerns expressed by Siriviriyakul (2014). [Insert Figure 1 here] 4.2 Empirical Results In Table 3 we present results from estimating model (2), where we examine whether firms that meet or just beat zero earnings or consensus analyst forecast benchmarks report lower levels of abnormal discretionary expenses. In Columns (1) and (2), we report results where Bencht is equal to Bench_NIt, and the dependent variable in Column (1) (Column (2)) is Firm- Specific_REM (Traditional_REM). Consistent with our expectations, in Column (1) we find that the coefficient on Bench_NIt is negative and significant, at the 5% level, suggesting that firms use real earnings management to avoid reporting negative earnings. In Column (2), we also find that the coefficient on Bench_NIt is negative and significant, at the 10% level, confirming the findings from Roychowdhury (2006) that firms use real earnings management to avoid reporting negative earnings, albeit at a lower level of statistical significance. In Columns (3) and (4), we report results where Bencht is equal to Bench_AFt, and the dependent variable Column (3) (Column (4)) is Firm-Specific_REM (Traditional_REM). Consistent with our expectations, we find that the coefficient on Bench_AFt in Column (1) is negative and significant, at the 1% level, 24