1 Introduction
Graham et al. (
2005) find that chief financial officers are willing to change their firms’ operating policies to meet a financial reporting target, implying that earnings management extends beyond accruals manipulation and includes real activities. For example, managers could temporarily cut research and development (R&D) outlays to show a profit instead of a loss. Studies call such a manipulation real earnings management (REM) and measure its extent by the difference between the firm’s costs and those reported by its industry peers (Roychowdhury
2006). The literature has made considerable progress in measuring accrual manipulation, but REM estimation models remain rudimentary.
1 Researchers continue to use the original models proposed by Roychowdhury (
2006), despite the problems identified in recent studies (Siriviriyakul
2015; Cohen, et al.
2016).
2 I investigate three questions: (1) Are REM models misspecified? (2) Could researchers draw incorrect inferences because of the misspecifications? (3) Can those models be improved and, if so, how?
I show that REM models cannot distinguish between a firm’s earnings management and its competitive strategy, because both can entail different levels of costs than industry peers. Consequently, researchers could erroneously infer REM if same-industry firms pursue different strategies. I find that variations in competitive strategies within industries are large enough to cause incorrect inferences about the presence and extent of earnings management. Furthermore, competitive strategy is associated with commonly studied accounting and finance variables, such as capital structure, corporate governance, executive compensation, and disclosure policy. Researchers can therefore document spurious correlations between earnings management and strategy-driven firm characteristics. I suggest improvements in REM estimation models to address this problem.
Roychowdhury (
2006) proposes four models to measure REM, each focused on a different component of operating income. Two of these models interpret negative abnormal levels of R&D and selling, general, and administrative (SG&A) expenses as cutbacks of discretionary costs. The third model considers positive abnormal production cost [cost of goods sold [(COGS) plus changes in inventory] to be overproduction. The fourth model regards abnormal cash flow from operations as a sign of earnings management.
3 Abnormal values for each of the four variables are obtained from linear regression models at the industry-year level and rely on two assumptions. First, all firms in an industry have the same cost and cash flow patterns when they are not managing earnings. Second, sales revenue is the sole driver of costs and profitability in the normal course of business. (SG&A and R&D models make this assumption using past revenue.) Researchers then, depending on the model, deem as abnormal the portion of costs or cash flows that are unrelated to current or past sales.
I show that the two assumptions underlying the estimation models are systematically violated. The cost patterns and cash profitability of firms in a given industry could differ because firms are in different stages of their life cycles (Miller and Friesen
1984; Dickinson
2011) or they adopt dissimilar business models at the time of their formation (Stinchcombe
1965). Young firms invest more in intangibles to create product differentiation or cost advantage (Porter
1980). In addition, firms listed in the last 25 years or so have capitalized on the progress in information technology to offer innovative products and enhanced services to customers (Shapiro and Varian
1998). Hence younger cohorts are more likely to pursue customer intimacy and product leadership strategies than older cohorts at the same stage of their life cycle (Treacy and Wiersema
1993). Pursuing those strategies requires high levels of intangible inputs, which are reported in R&D and SG&A (Brown and Kapadia
2007; Govindarajan and Srivastava
2016). I find that younger cohorts incur larger R&D and SG&A expenses in the normal course of operations than older cohorts in the same industry. Furthermore, older cohorts show higher levels of cash profitability than do younger cohorts, which typically incur losses. So the old and new cohorts within industries differ in their cost and profitability patterns.
The second assumption, that current or past sales is the sole driver of nonmanipulated costs and profitability, is systematically violated in three of the four REM models. Firms make cost decisions according to their competitive strategy. For example, firms invest in innovation, strategy, market research, customer and social relationships, computerized data and software, brands, and human capital to reap long-term rewards (Wernerfelt
1984; Peteraf
1993; Eisfeldt and Papanikolaou
2013).
4 A firm’s plan for its future market share, revenues, or profits thus must be a significant determinant of its investment policy. A discretionary cost model that is based solely on past revenues should therefore be misspecified, because it excludes major determinants of planned investments. In contrast, a production cost model would be well specified because COGS, the main component of production cost, is matched to current revenues by accounting convention.
5
I find that the residuals from discretionary cost models (the portion of costs that is unrelated to current revenues) are large and strongly associated with future revenue growth. Discretionary cost models thus measure REM with errors that reflect long-term investments. Furthermore, residuals display the same cohort patterns as the reported discretionary costs—they increase from the oldest to the youngest cohorts. Because regression residuals must add up to zero, the oldest cohorts show large negative residuals, while the youngest cohorts display large positive values. A researcher would conclude that the oldest cohorts opportunistically cut discretionary costs and the youngest cohorts overinvest in intangibles.
The production cost model is better specified than the discretionary cost model and yields smaller residuals, because COGS is highly matched to current revenues.
6 Also, the difference between the residuals of the youngest and the oldest cohorts is much smaller for the production cost model than for the discretionary cost models. Thus the youngest and the oldest cohorts show no significant difference in earnings management by overproduction but appear to significantly differ in earnings management by discretionary cost curtailment. This pattern is noteworthy because studies typically find significant earnings management using discretionary cost models but not with production cost models. Stated differently, the literature shows widespread earnings management using measures that are obtained from under-specified models. But the same studies do not report significant results with measures of better-specified models. Furthermore, earlier studies typically find higher REM for large, low-growth, and highly profitable firms, which are the characteristics of older cohorts.
7 In effect, those studies conclude that older cohorts manage earnings by cutting discretionary costs when the routine business practice of those firms may be to invest less in research and development and intangibles.
The above tests do not rule out the possibility that older cohorts manage earnings to a greater extent than do younger cohorts. Three tests negate this proposition. First, older cohorts are characterized by low growth and positive cash flows. Therefore they have the least incentive to mislead external capital providers, which is arguably the strongest motive for earnings management (Dechow and Skinner
2000). Second, the serial correlation of REM proxies is as high as 0.5–0.7, indicating persistence in firms’ distinctive characteristics (Siriviriyakul
2015). This high persistence shows the stability of firms’ competitive strategies (Stinchcombe
1965; Porter
1980) and is inconsistent with the idea that REM is a temporary deviation from a firm’s optimal business practice and that it should reverse in the next period to catch up with necessary expenses.
8 Third, the oldest cohorts continue to show the largest profits year after year. This pattern contradicts the proposition that the oldest cohorts continually manipulate their operations, because a prolonged deviation from the optimal business practice must be followed by reduction in profits.
9
In addition, the oldest cohorts’ normal discretionary costs (those explained by the estimation models) are also lower than those of the youngest cohorts. These findings lead me to conclude that the oldest cohorts’ distinct ways of doing business are misinterpreted by the current models as real earnings management. Members of the oldest cohorts have survived for more than 40 years. They must have competed successfully against each other and the nonsurviving firms by following superior strategies, creating better products, or establishing more stable markets and customer bases, which now enables them to earn economic rents without having to invest as much in intangibles as younger cohorts (Amit and Schoemaker
1993; Agarwal and Gort
2002). New players, in contrast, must spend higher amounts on innovation, strategy, advertising, customer relationships, and brands to build competitive advantages or to gain from recent technological advances (Porter
1980; Shapiro and Varian
1998). These differences in competitive strategies of the oldest and youngest cohorts, in conjunction with the under-specification of REM estimation models, lead to the appearance that the oldest cohorts underinvest in intangible assets.
The main takeaway from the paper is that competitive strategy is an omitted variable in REM estimation models that should be included in the first-stage models.
10 The empirical proxies for competitive strategy are not available in financial reports, which is a major limitation of accounting (Lev and Gu
2016). I propose a sequence of corrective steps based on the available financial statement variables. First, I assume that a firm’s competitive strategy relates to its opportunity set. Thus, in the first-stage estimation, I include the proxies for opportunity set, namely, size, past profitability, and growth (Gunny
2010). Second, I assume that firms spend on intangibles to generate current revenues as well as to secure future benefits. Hence I include future revenues in the estimation models. Third, I control for the firm’s own past expenses to identify deviations from its routine behavior (Gunny
2010). Even if successfully applied, these three steps cannot correct for a firm’s optimal business response to a new economic shock in the measurement year. That response would appear as a deviation from the firm’s past expenses, which could be misinterpreted as real earnings management. Researchers can avoid this error by using a cohort adjustment, based on the assumption that firms in similar life-cycle stage and with similar technological vintage experience similar economic shocks.
11 I subtract the costs of a similar-size firm belonging to the same industry cohort from the costs of a given firm to estimate its abnormal behavior.
I demonstrate that each sequential step mitigates the measurement errors in REM proxies and reduces the portion of costs considered manipulative. Mitigation with each step, however, differs across proxies. The inclusion of one-year-forward revenues more effectively mitigates the errors for SG&A than for the production cost model, because abnormal SG&A is strongly correlated with future revenue growth and abnormal COGS is not. Cohort adjustment has the largest effect on R&D models, suggesting that firms with similar technology vintage and in similar stage of life cycle spend similar amounts on R&D. The steps I propose could change the inferences of studies, such as that of Kim and Park (
2014).
My paper makes three contributions to the literature. First, it adds to understanding of the earnings management phenomenon as measured by the current REM models. I show that the models ignore the relation between a firm’s competitive strategy and its costs, leading researchers to misinterpret strategy-related cost difference as earnings management. My findings are consistent with the ideas of Dechow et al. (
2010) that earnings properties are determined by both fundamental performance and accounting practice and of Ball (
2013) that researchers often find earnings management when no other party with greater information and incentive detects that pattern.
Second, I propose enhancements in the estimation models to lower the competitive strategy-related measurement errors in earnings management proxies. Any hypothesis test of an incentive for earnings management is a joint test of the validity of the researcher’s first-stage model and the relation between the incentive and earnings management. The enhancements I propose should improve the reliability of future tests about earnings management. As such, my contribution is analogous to that of Dechow et al. (
1995), Kothari et al. (
2005), and Owens et al. (
2017) in improving the measurement of discretionary accruals. Nevertheless, I caution researchers against mechanically applying the corrective steps I propose. These steps would overcorrect for errors when the incentive for earnings management varies with the firm’s opportunity set, when the firm routinely manages earnings, or when the members of the firm’s industry cohort manage earnings to an equal extent.
Third, results of this paper can be generalized to any model that estimates a firm’s abnormal, manipulative, or suboptimal behavior by the uniqueness of its characteristics vis-à-vis industry peers. I show that the oldest and youngest cohorts in an industry often differ in their competitive strategies, leading to systematic differences in their strategy-related financial characteristics. Thus cohort adjustment must be applied to any industry-based measurement of suboptimal or manipulative behavior.
The rest of the paper is organized as follows. Section
2 describes the literature on real earnings management and explains the estimation models and measurement of variables. Section
3 examines the violations of the two assumptions underlying the Roychowdhury (
2006) models. Sections
4 and
5 investigate whether model misspecifications can lead to incorrect inferences about the presence and the extent of real earnings management. Section
6 proposes a sequence of improvements in the models. Section
7 concludes.
2 Prior research, description of models, and measurement of variables
Healy and Wahlen (
1999, p. 368) state that “[e]arnings management occurs when managers use judgment in financial reporting and in structuring transactions to alter financial reports to either mislead some stakeholders about the underlying economic performance of the company or to influence contractual outcomes that depend on reported accounting numbers.” Managers can manipulate not only financial reports but also operating and financing activities to meet reporting targets.
12 This idea is confirmed in the Graham et al. (
2005) survey of financial executives, which finds that chief financial officers are willing to cut discretionary costs, such as R&D and advertising, to show higher earnings in the short term. Roychowdhury (
2006) proposes an innovative method to detect such opportunism.
Roychowdhury (
2006) suggests that deviations in production costs and discretionary costs from otherwise optimal operating decisions represent managers’ attempt to manipulate earnings. He reasons that lower discretionary costs, compared with industry peers [identified by two-digit Standard Industrial Classification (SIC) code], could indicate the reduction of soft discretionary costs. He also posits that higher production costs, relative to peers, represent overproduction of goods. He further argues that manipulation of real activities affects operating cash flow, though the direction of the effect is ambiguous. Many subsequent studies associate abnormal operating cash flow with REM, consistent with the idea that curtailment of discretionary costs increases operating cash flow.
Roychowdhury (
2006) models require two assumptions. First, in the normal course of business, all firms in a given industry need the same level of discretionary costs and production costs, and they generate the same levels of cash operating profits. (All variables are scaled by total assets at the beginning of the year.) Second, either current or past revenue is the sole determinant of optimal costs. Based on these assumptions, Roychowdhury (
2006) measures a firm’s deviations from its optimal outlays by the residuals from the regressions of SG&A, R&D, production costs, and operating cash flow on current or past revenues estimated by industry and year. He finds that regression residuals are associated with the frequency of meeting earnings benchmarks. He therefore reasons that the regression residuals represent a firm’s suboptimal behavior to manipulate financial reports. Roychowdhury’s models continue to be widely used in the literature, despite enhancements proposed by subsequent studies (e.g., Gunny
2010).
2.1 Measurement of real earnings management
Consistent with Roychowdhury (
2006), I measure discretionary costs by
SG&A (Compustat XSGA) and
R&D (XRD).
ProductionCost is calculated by adding changes in inventory (INVT) to cost of goods sold (COGS). Cash flow from operations (OANCF) is referred to as
OperatingCashFlow.
13 All variables are scaled by total assets at the beginning of the year (AT).
To determine overproduction, I follow Roychowdhury (
2006) and estimate the following cross-sectional regression for each industry (two-digit SIC code) and year.
$$ {\displaystyle \begin{array}{l}{\mathrm{ProductionCost}}_{\mathrm{i},\mathrm{t}}={\upbeta}_1+{\upbeta}_2\times \frac{1}{{\mathrm{Total}\ \mathrm{Assets}}_{\mathrm{i},\mathrm{t}-1}}+{\upbeta}_3\times \frac{{\mathrm{Sales}}_{\mathrm{i},\mathrm{t}}}{{\mathrm{Total}\ \mathrm{Assets}}_{\mathrm{i},\mathrm{t}-1}}+{\upbeta}_4\times \frac{\Delta {\mathrm{Sales}}_{\mathrm{i},\mathrm{t}}}{{\mathrm{Total}\ \mathrm{Assets}}_{\mathrm{i},\mathrm{t}-1}}\\ {}+{\upbeta}_5\times \frac{\Delta {\mathrm{Sales}}_{\mathrm{i},\mathrm{t}-1}}{{\mathrm{Total}\ \mathrm{Assets}}_{\mathrm{i},\mathrm{t}-1}}+{\upepsilon}_{\mathrm{i},\mathrm{t}}\end{array}} $$
(1)
where
∆Sales represents changes in revenues. The residual estimated on a firm-year basis represents a manipulation of the production schedule. The more positive the residual, the higher the manipulation, assuming that firms increase their production levels to spread fixed costs over a larger number of units to show higher profit margins.
To determine curtailment of discretionary costs, the following cross-sectional models are estimated for each industry and year (Roychowdhury
2006).
$$ \mathrm{SG}\&{\mathrm{A}}_{\mathrm{i},\mathrm{t}}={\upbeta}_1+{\upbeta}_2\times \frac{1}{{\mathrm{Total}\ \mathrm{Assets}}_{\mathrm{i},\mathrm{t}-1}}+{\upbeta}_3\times \frac{{\mathrm{Sales}}_{\mathrm{i},\mathrm{t}-1}}{{\mathrm{Total}\ \mathrm{Assets}}_{\mathrm{i},\mathrm{t}-1}}+{\upepsilon}_{\mathrm{i},\mathrm{t}} $$
(2)
and
$$ \mathrm{R}\&{\mathrm{D}}_{\mathrm{i},\mathrm{t}}={\upbeta}_1+{\upbeta}_2\times \frac{1}{{\mathrm{Total}\ \mathrm{Assets}}_{\mathrm{i},\mathrm{t}-1}}+{\upbeta}_3\times \frac{{\mathrm{Sales}}_{\mathrm{i},\mathrm{t}-1}}{{\mathrm{Total}\ \mathrm{Assets}}_{\mathrm{i},\mathrm{t}-1}}+{\upepsilon}_{\mathrm{i},\mathrm{t}}. $$
(3)
The residuals are the inverse measures of manipulation of discretionary costs. (The model for R&D is the same as that of Roychowdhury (
2006, p. 351, footnote 24).) The more negative the SG&A and R&D residuals, the higher the curtailment of discretionary costs.
Abnormal
OperatingCashFlow is measured by estimating the following cross-sectional model by industry-year (Roychowdhury
2006).
$$ {\mathrm{OperatingCashFlow}}_{\mathrm{i},\mathrm{t}}={\upbeta}_1+{\upbeta}_2\times \frac{1}{{\mathrm{Total}\ \mathrm{Assets}}_{\mathrm{i},\mathrm{t}-1}}+{\upbeta}_3\times \frac{{\mathrm{Sales}}_{\mathrm{i},\mathrm{t}}}{{\mathrm{Total}\ \mathrm{Assets}}_{\mathrm{i},\mathrm{t}-1}}+{\upbeta}_4\times \frac{\Delta {\mathrm{Sales}}_{\mathrm{i},\mathrm{t}}}{{\mathrm{Total}\ \mathrm{Assets}}_{\mathrm{i},\mathrm{t}-1}}+{\upepsilon}_{\mathrm{i},\mathrm{t}}. $$
(4)
The residuals represent the curtailment of sales expenses.
2.2 Financial characteristics
In addition to costs, I examine variations in financial characteristics of firms in the same industry. I consider the market value of equity, lagged return on assets (ROA), and the market-to-book ratio as proxies for firm size, nonmanipulated profitability, and growth, respectively. I measure profitability by the earnings-to-price ratio and the return on assets. A firm is assumed to have just missed its earnings target if it reports a loss and the ratio of net income to beginning-of-year total assets lies between zero and − 1%. A firm is assumed to have just met the earnings target if its ratio of net income to beginning-of-year total assets is between zero and 1%. I also calculate these variables based on changes in earnings. If the change is negative but greater than −1% of beginning-of-year total assets, the firm is presumed to have just missed showing improvement in profitability. If the firm reports an increase in net income greater than zero but less than 1% of beginning-of-year total assets, it is presumed to have just shown improvement in profitability.
4 Differences in proxies of real earnings management by listing cohorts
Measurement errors in discretionary cost models would appear in regression residuals. Younger cohorts, with their larger discretionary costs and in conjunction with under-specified models, would show higher residuals than the older cohorts. But regression residuals must add up to zero. So the oldest cohorts would show negative residuals, and the youngest cohorts would show positive residuals. I test this proposition by classifying all firms into quartiles by their listing vintage within each industry. For the four largest industries (metal and mining, chemical and allied products, electronic and other electric, and business services), I present the quartile averages of abnormal components of
R&D,
SG&A,
OperatingCashFlow, and
ProductionCost in Panels A2–D2 of Fig.
1, respectively. The total value for each variable for the same cohort is presented alongside in Panels A1–D1 for easy comparison. In three of the four industries, the abnormal values of
R&D and
SG&A are negative for the oldest cohorts and positive for the youngest ones. The abnormal values of
OperatingCashFlow are positive for the oldest cohorts and negative for the youngest ones. These patterns give the appearance of significant real earnings management by the oldest cohorts. No consistent evidence on real earnings management for the oldest cohorts is obtained with abnormal production costs.
I then calculate the average values of abnormal components for the youngest and oldest cohorts for the pooled sample. Panel A of Table
3 shows that the abnormal values for
R&D,
SG&A,
OperatingCashFlow, and
ProductionCost are, respectively, 0.027, 0.204, −0.119, and 0.007 for the youngest cohorts and − 0.026, −0.144, 0.075, and − 0.021 for the oldest cohorts. All of these values are significant (except abnormal production cost for the youngest cohorts). All of the differences between the youngest and the oldest cohorts are also significant.
Table 3Differences in appearance of real earnings management by the youngest and oldest cohorts
Youngest cohorts | 0.027*** | 0.204*** | −0.119*** | 0.007 |
Oldest cohorts | −0.026*** | −0.144*** | 0.075*** | −0.021** |
Difference | 0.053*** | 0.348*** | −0.194*** | 0.027** |
Panel B: Percentage of instances of detecting significantly positive or negative values for randomly drawn samples from youngest and oldest cohorts |
| AbnormalR&D | AbnormalSG&A | AbnormalOperatingCashFlow | AbnormalProductionCost |
Quartile of firms | < −0.01 | < −0.05 | > 0.05 | > 0.05 |
Youngest cohorts | 0% | 1% | 0% | 6% |
Oldest cohorts | 92% | 94% | 81% | 6% |
Difference | −92% | −93% | −81% | 0% |
Two patterns are noteworthy. First, the magnitude of ratios of abnormal to total value for
R&D,
SG&A, and
OperatingCashFlow is much larger than that for
ProductionCost. (Total values are presented in Panel B of Table
1.) The magnitude of ratios for the oldest cohorts is 65% (−0.026 / 0.040) for
R&D, 40% (−0.144 / 0.356) for
SG&A, and 416% (0.075 / 0.018) for
OperatingCashFlow but only 3% (0.021 / 0.729) for
ProductionCost. Thus discretionary cost and cash flow models produce relatively large residuals. Second, those residuals carry statistically significant values for the oldest and youngest cohorts but in opposite directions. (The residuals for the middle two cohorts are relatively small.) A researcher would interpret the abnormal values for the oldest cohorts as undercutting of discretionary costs. The youngest cohorts do not display real earnings management using any measure. They would appear to overspend on intangibles, as if exacerbating their already low profits.
I consider 0.05 as a threshold for interpreting significant earnings management, consistent with the anecdotal evidence that 5% of total assets is considered a material amount. I use a threshold of 0.01 for R&D, because it is a much smaller number than the other three variables. I then estimate the percentage of firms showing earnings management in randomly drawn samples from the oldest and the newest cohorts.
I draw 100 random samples of 100 observations each from the top and bottom quartiles and estimate the likelihood of obtaining materially significant values for the four abnormal components in the direction consistent with real earnings management. I calculate percentage of observations with
AbnormalR&D less than −0.01,
AbnormalSG&A less than −0.05, and
AbnormalOperatingCashFlow and
AbnormalProductionCost greater than 0.05. Panel B of Table
3 presents these likelihoods. For the oldest cohorts, the likelihood of detecting large real earnings management based on R&D, SG&A, operating cash flow, and production cost is 92%, 94%, 81%, and 6%, respectively. For the youngest cohorts, the same likelihoods are 0%, 1%, 0%, and 6%, respectively. Stated differently, randomly drawn samples from the oldest cohorts almost always appear to cut discretionary costs but do not appear to manipulate production costs. The same is not true for the youngest cohorts. This phenomenon explains a typical finding of real earnings management studies relying on the Roychowdhury (
2006) models that firms with low growth, large size, and high profitability cut discretionary costs, because these are more likely to be the characteristics of older cohorts than younger cohorts. Furthermore, studies find significant results with
SG&A,
R&D, and
OperatingCashFlow but rarely with
ProductionCost.
20
The findings of this section are consistent with the idea that the inferences of real earnings management in the literature could represent violations of two assumptions underlying the REM estimation models: (1) all firms in an industry have the same cost and cash flow patterns when not managing earnings; (2) sales revenue is the sole driver of costs and profitability in the normal course of business.
6 Improvements in measurement models
Measurement errors in empirical proxies, if randomly distributed, should merely reduce the power but not bias the results of the tests of the hypotheses. However, measurement errors in three of the four real earnings management proxies are not randomly distributed. They display cohort patterns and are manifestations of competitive strategy. This systematic measurement error could cause spurious correlations in any hypothesis test involving a firm characteristic that is driven by firm’s competitive strategy. Researchers can therefore document spurious correlations between earnings management and that strategy-driven characteristic.
I propose a sequence of corrective steps to mitigate these possible errors. First, I include the widely accepted proxies for a firm’s opportunity set of size, past profitability, and growth in the first-stage model (Gunny
2010). Second, I include future revenues in the model, because firms spend on intangibles not only to produce current revenues but also to secure future benefits. Third, I control for the firm’s own past expenses to identify deviations from the firm’s behavior in prior years (Gunny
2010). Hence I estimate eqs. (1)–(4) with the inclusion of five new variables. For example, eq. (1) is recalculated by
$$ {\displaystyle \begin{array}{l}{\mathrm{ProductionCost}}_{\mathrm{i},\mathrm{t}}={\upbeta}_1+{\upbeta}_2\times \frac{1}{{\mathrm{Total}\ \mathrm{Assets}}_{\mathrm{i},\mathrm{t}-1}}+{\upbeta}_3\times \frac{{\mathrm{Sales}}_{\mathrm{i},\mathrm{t}}}{{\mathrm{Total}\ \mathrm{Assets}}_{\mathrm{i},\mathrm{t}-1}}+{\upbeta}_4\times \frac{\Delta {\mathrm{Sales}}_{\mathrm{i},\mathrm{t}}}{{\mathrm{Total}\ \mathrm{Assets}}_{\mathrm{i},\mathrm{t}-1}}+{\upbeta}_5\times \frac{\Delta {\mathrm{Sales}}_{\mathrm{i},\mathrm{t}-1}}{{\mathrm{Total}\ \mathrm{Assets}}_{\mathrm{i},\mathrm{t}-1}}\\ {}+{\upbeta}_6\times \mathrm{LogMarketValuei},\mathrm{t}+{\upbeta}_7\times \mathrm{LagROAi},\mathrm{t}+{\upbeta}_8\times \mathrm{M}/\mathrm{Bi},\mathrm{t}+{\upbeta}_9\times \frac{{\mathrm{Sales}}_{\mathrm{i},\mathrm{t}+1}}{{\mathrm{Total}\ \mathrm{Assets}}_{\mathrm{i},\mathrm{t}-1}}+{\upbeta}_{10}\times {\mathrm{ProductionCost}}_{\mathrm{i},\mathrm{t}-1}+{\upepsilon}_{\mathrm{i},\mathrm{t}}\end{array}} $$
(7)
The new variables are in the second line of eq. (7). Even if successfully applied, these three sets of controls cannot correct for a firm’s optimal business response to an external shock or opportunity in the given year, which would appear as a deviation from the firm’s own past behavior. This factor can be controlled by a cohort adjustment motivated by the assumption that firms in similar life-cycle stage and with similar technology vintage experience similar economic shocks and have similar optimal response. A firm’s abnormal behavior is thus estimated by subtracting the activity of a same-cohort firm having similar size from the activity of the given firm to obtain a cohort-adjusted measure.
23
I next investigate whether my sequence of steps mitigates the three characteristics that a valid earnings management measure should not display: (1) persistence, (2) greater earnings management by older cohorts, and (3) more frequent rejection of the null hypothesis of no real earnings management measures for randomly drawn samples from older cohorts. The results are presented by sequentially applying size, past profitability, and growth; future revenues; past expenditures; and cohort adjustment.
Results for persistence are presented in Panel A of Table
5. Persistence dramatically declines from 0.67 to 0.11, 0.67 to 0.08, 0.51 to 0.06, and 0.58 to 0.10, respectively, for
AbnormalR&D,
AbnormalSG&A,
AbnormalOperatingCashFlow, and
AbnormalProductionCost, after I apply the four corrective steps. This panel also shows how different steps dissimilarly reduce persistence. Size, past profitability, and growth most significantly mitigate persistence for
AbnormalSG&A (arguably by controlling for firm’s opportunity set) and
AbnormalOperatingCashFlow (arguably by controlling for underlying profitability). Inclusion of one-year-forward revenues further reduces persistence for
AbnormalSG&A, because SG&A produces benefits in the next year. It makes no difference for the production cost model, because COGS bears little correlation with future revenues. Controlling for the past values, perhaps mechanically, reduces persistence for all variables. But cohort adjustment also significantly reduces persistence.
Table 5Improvement in the measures of real earnings management
AbnormalR&D | 0.67 | 0.64 | 0.63 | 0.25 | 0.11 |
AbnormalSG&A | 0.67 | 0.34 | 0.26 | 0.11 | 0.08 |
AbnormalOperatingCashFlow | 0.51 | 0.26 | 0.25 | 0.19 | 0.06 |
AbnormalProductionCost | 0.58 | 0.55 | 0.57 | 0.16 | 0.10 |
Panel B: Differences between the original and revised measures of real earnings management for the youngest and oldest cohorts |
| Original measure | Revised measure 1 | Revised measure 2 | Revised measure 3 | Cohort-adjusted measure |
AbnormalR&D |
Youngest cohorts | 0.027*** | 0.013*** | 0.012*** | 0.008*** | 0.00 |
Oldest cohorts | −0.026*** | −0.015*** | −0.014*** | −0.002* | 0.00 |
Difference | 0.053*** | 0.028*** | 0.026*** | 0.01*** | 0.00 |
Percentage reduction in difference | | 49% | 51% | 81% | 100% |
AbnormalSG&A |
Youngest firms | 0.204*** | 0.043* | 0.010 | 0.000 | −0.004 |
Oldest firms | −0.144*** | −0.049*** | −0.031*** | −0.002 | 0.005 |
Difference | 0.348*** | 0.091*** | 0.041*** | 0.003 | −0.009 |
Percentage reduction in difference | | 74% | 88% | 99% | Sign reverses |
AbnormalOperatingCashFlow |
Youngest cohorts | −0.119*** | −0.041*** | −0.035** | −0.029** | −0.019 |
Oldest cohorts | 0.075*** | 0.009* | 0.006 | 0.005 | −0.005 |
Difference | −0.194*** | −0.05*** | −0.041*** | −0.035** | −0.014 |
Percentage reduction in difference | | 74% | 79% | 82% | 93% |
AbnormalProductionCost |
Youngest cohorts | 0.007 | −0.018* | −0.020* | −0.004 | −0.007 |
Oldest cohorts | −0.021*** | 0.004 | 0.008 | 0.004 | −0.008 |
Difference | 0.027** | −0.022* | −0.027** | −0.008 | 0.001 |
Percentage reduction in difference | | Sign reverses | Sign reverses | Sign reverses | 99% |
Panel C: Differences in detecting significant real earnings management measures between youngest and oldest cohorts for randomly drawn samples |
| Original measure | Revised measure 1 | Revised measure 2 | Revised measure 3 | Cohort-adjusted measure |
Abnormal R&D < −0.01 |
Youngest cohorts | 0% | 0% | 2% | 0% | 21% |
Oldest cohorts | 92% | 74% | 73% | 1% | 1% |
Difference | −92% | −74% | −71% | −1% | 20% |
AbnormalSG&A < −0.05 |
Youngest cohorts | 1% | 12% | 23% | 19% | 38% |
Oldest cohorts | 94% | 45% | 31% | 4% | 13% |
Difference | −93% | −33% | −8% | 15% | 25% |
AbnormalOperatingCashFlow > 0.05 |
Youngest cohorts | 0% | 2% | 2% | 3% | 19% |
Oldest cohorts | 81% | 0% | 0% | 0% | 2% |
Difference | −81% | 2% | 2% | 3% | 17% |
AbnormalProductionCost > 0.05 | | | | | |
Youngest cohorts | 6% | 0% | 2% | 1% | 11% |
Oldest cohorts | 6% | 3% | 3% | 0% | 1% |
Difference | 0% | −3% | −1% | 1% | 10% |
Panel D: Percentage improvement in true positives and false negatives |
| Revised measure 1 | Revised measure 2 | Revised measure 3 | Cohort-adjusted measure |
AbnormalR&D |
Percentage improvement in true positives | −1.59% | −4.06% | 16.94% | 69.58% |
Percentage reduction in false negatives | 12.09% | 10.99% | 36.27% | 54.94% |
AbnormalSG&A |
Percentage improvement in true positives | 8.63% | 10.34% | 43.53% | 17.20% |
Percentage reduction in false negatives | 31.17% | 36.36% | 51.08% | 43.72% |
AbnormalOperatingCashFlow |
Percentage improvement in true positives | 16.29% | 11.83% | 10.15% | 23.10% |
Percentage reduction in false negatives | 35.99% | 36.74% | 40.15% | 47.35% |
AbnormalProductionCost |
Percentage improvement in true positives | 72.44% | 82.99% | 106.66% | 103.65% |
Percentage reduction in false negatives | 38.52% | 40.57% | 50.82% | 40.16% |
I next examine the effect of the corrective steps on cohort patterns. In addition to presenting the values for the oldest and youngest cohorts and their difference after each step, I present the percentage reduction in the oldest cohort-youngest cohort difference. As in persistence tests, the controls for size, past profitability, and growth most significantly mitigate the cohort difference for SG&A and operating cash flow-based proxies. This step also mitigates the difference for AbnormalR&D by almost 50%. By the third step, the differences are significantly reduced by 81%, 99%, and 82% for the abnormal components of R&D, SG&A, and OperatingCashFlow, respectively. Finally, cohort adjustment mechanically reduces the difference in all variables, making it statistically insignificant.
Another noteworthy result is the decline in the absolute value of the abnormal components with each step. For example, for the oldest cohorts, the absolute value of
AbnormalSG&A reduces from 0.144 to 0.049, with the application of size, past profitability, and growth, and further to 0.002, with control for future revenues and lagged values—a total reduction of 98.7%. These results indicate that real earnings management, as measured by the residuals from Roychowdhury (
2006) models, could be overstated (Ball
2013).
I next examine the frequency of rejecting the null hypothesis of no earnings management for the oldest cohorts. As in Table
4 tests, I use the absolute value of 0.01 as the threshold for
AbnormalR&D and 0.05 as the threshold for
AbnormalSG&A,
AbnormalOperatingCashFlow, and
AbnormalProductionCost. Panel C shows that the frequency of rejection of the null hypothesis declines with each sequential step. The difference between the oldest and youngest cohorts disappears and flips its sign after the fourth step is implemented. Results after the application of corrective steps indicate that the youngest cohorts manage earnings more than the oldest cohorts, a pattern consistent with their capital market incentives.
I next assume that any firm displaying earnings management in a randomly selected sample represents a false negative. From a randomly selected sample of 10% of firms, I call observations with AbnormalR&D less than −0.01, AbnormalSG&A less than −0.05, and AbnormalOperatingCashFlow and AbnormalProductionCost greater than 0.05 as false negatives. I subtract the average of false negatives for each revised measure from the average for the original measure and call the difference the percentage reduction in false negatives. For the same random 10% firm sample, I induce real earnings management. That is, I subtract 0.01 from R&D and 0.05 from SG&A and add 0.05 to OperatingCashFlow and ProductionCost. I then calculate real earnings management proxies, using the original and the revised methods. I calculate a ratio of average number of firms showing abnormal values beyond the induced levels, for induced versus uninduced cases. I reason that this ratio represents the model’s ability to identify true positives. I call it the true positive ratio. I subtract that ratio for the original measure from that for each revised measure and call the difference the percentage improvement in true positives.
I find that my proposed methods mitigate false negatives and improve true positives for all four proxies. The percentage reduction in false negatives for AbnormalR&D, AbnormalSG&A, AbnormalOperatingCashFlow, and AbnormalProductionCost, after implementing the proposed four steps, is 54.94%, 43.72%, 47.35%, and 40.16%, respectively. The percentage improvement in identification of true positives is 69.58%, 17.20%, 23.10%, and 103.65%, respectively.
I demonstrate the application of the sequential-correction method to Kim and Park (
2014), who examine real earnings management for firms that change their auditors.
24 I replicate their Table
4 using data from Audit Analytics. (They supplement those data with hand-collection.) Their Table
4 shows significant differences between real earnings management proxies for firms with auditor resignations and for firms with continuing auditors. Similar to their results, my Table
6 shows significant differences between the two groups in the abnormal components of operating cost and SG&A but not for production cost.
Table 6An application of revised measures of real earnings management
AbnormalOperatingCashFlow |
| −0.0996 | −0.0180 | 0.0083 | −0.0816** | −0.1079*** |
Original measure in my paper | −0.1010 | −0.0565 | 0.0147 | −0.0445*** | −0.1157*** |
Revised measure 1 | 0.0117 | −0.0134 | 0.0062 | 0.0252 | 0.0056 |
Revised measure 2 | 0.0153 | −0.0095 | 0.0058 | 0.0247 | 0.0094 |
Revised measure 3 | −0.0007 | −0.0112 | 0.0024 | 0.0105 | −0.0031 |
Cohort-adjusted measure | 0.0137 | −0.0164 | 0.0007 | 0.0301 | 0.0130 |
AbnormalProductionCost |
| −0.0113 | −0.0019 | −0.0267 | −0.0094 | 0.0154 |
Original measure in my paper | −0.0903 | −0.0357 | −0.0541 | −0.0546 | −0.0362 |
Revised measure 1 | −0.0945 | −0.0375 | −0.0535 | −0.0571 | −0.0410 |
Revised measure 2 | −0.1066 | −0.0488 | −0.0489 | −0.0578 | −0.0577 |
Revised measure 3 | −0.0462 | −0.0355 | −0.0262 | −0.0107 | −0.0200 |
Cohort-adjusted measure | −0.0690 | −0.0236 | −0.0273 | −0.0454 | −0.0417 |
AbnormalSG&A |
| −0.2150 | 0.2023 | 0.1035 | −0.4173*** | −0.3185*** |
Original measure in my paper | −0.2121 | 0.0885 | 0.0135 | −0.3005*** | −0.2256*** |
Revised measure 1 | 0.0014 | −0.0247 | 0.0069 | 0.0261 | −0.0055 |
Revised measure 2 | −0.0093 | −0.0236 | 0.0050 | 0.0144 | −0.0142 |
Revised measure 3 | 0.0000 | −0.0178 | 0.0047 | 0.0179 | −0.0047 |
Cohort-adjusted measure | −0.0292 | −0.0290 | 0.0054 | −0.0002 | −0.0347 |
I then apply my sequence of steps and find two significant differences from Kim and Park (
2014). First, the absolute values of real earnings management proxies decline with each sequential step, indicating that the proxies estimated after controlling for competitive strategy are not as large as previously observed. The implication is that the existence of real earnings management for both groups is overestimated. Second, the difference between the earnings management proxies for the two groups of companies becomes insignificant. In fact, just the inclusion of size, past profitability, and growth in the estimation models almost eliminates the difference in real earnings management proxies for the two groups. These results demonstrate that an inference regarding the presence and the extent of real earnings management studies could change if the proxies were calculated using the methods I propose.
7 Conclusion
This study shows that the commonly used industry-year-based models for estimating real earnings management are misspecified, because they do not control for within-industry variations in competitive strategy. As a result, wrongful inferences could be drawn about the presence of and the cause and effect relation with real earnings management, when the researcher’s study variables are associated with competitive strategy.
I propose four steps to reduce competitive strategy-related measurement errors in the proxies of real earnings management. First, I include the widely accepted proxies for a firm’s opportunity set of size, past profitability, and growth in the estimation model. Second, I include future revenues, because firms spend on intangibles not just for producing current revenues but also in expectation of future benefits. Third, I control for a firm’s own past expenses, to identify deviations from its usual behavior in other years. Fourth, the activity of a similar-sized firm belonging to the same industry cohort is deducted from the given firm’s activity to identify its abnormal activity.
I show that the implementation of the four steps significantly mitigates the measurement errors prevalent in the current proxies for real earnings management. The enhancements in models I propose should thus improve the inferences of tests involving real earnings management. Nevertheless, these enhancements would overcorrect for misspecifications if earnings management varies with the firm’s opportunity set, the firm habitually manages earnings, or the members of the firm’s cohort also equally manage earnings.
The thesis of this paper can be generalized to any model that estimates a firm’s abnormal or manipulative behavior by difference from that of other industry players. Researchers should be cautious in interpreting their results, because that difference could represent the firm’s unique competitive strategy.
Acknowledgements
I thank Brad Badertscher, Jeremy Bertomeu, Sanjay Bissessur, Dirk Black, Patricia Dechow (editor), Ian Eggleton, Luminita Enache, Réka Felleg, Joseph Gerakos, Wayne Guay, Rachel Hayes, Jonas Heese, Kalin Kolev, Pepa Kraft (discussant), Scott Lee, Alvis Lo, Pablo Machado (discussant), Raj Mashruwala, Sarah McVay, Ken Merkley, Sugata Roychowdhury, Richard Sansing, Bryce Schonberger, Mani Sethuraman, Ventsislav Stamenov, Phil Stocken, Xiaoli Tian (discussant), Senyo Tse, Tony van Zijl, David Veenman, Charles Wang, Paul Zarowin, Colin Zeng (discussant), two anonymous reviewers, and the seminar participants at the 2015 annual meeting of the American Accounting Association, 2016 meeting of the European Financial Management, 2016 Tuck Accounting Conference, 2017 Financial Accounting Research sectional meeting of the American Accounting Association, University of Amsterdam, University of Baltimore, University of Calgary, University of New Hampshire, Victoria University of Wellington, and University of Nevada for suggestions that have considerably improved the paper. I also acknowledge financial support from Daniel R. Revers T’89 Faculty Fellowship at Tuck School of Business, Dartmouth College, Haskayne School of Business, University of Calgary, and Canada Research Chair program of the Government of Canada.
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.