First Version: June SUMMARY This study investigates the relationship between cost stickiness and income smoothing. Both

Transcription

1 1 Evidence on the Trade-Off between Managerial Motives for Cost Stickiness and Income Smoothing First Version: June 2016 SUMMARY This study investigates the relationship between cost stickiness and income smoothing. Both arise as a result of managerial discretion. However, an asymmetrical reaction of costs to changes in activity levels counteracts the ambition to report smooth earnings. Hence, we expect these two managerial phenomena to be negatively associated. More precisely, we assume that managers trade off incentives for engaging in income smoothing against personal motives inducing cost stickiness. Applying both a cross-sectional model and a firm-specific measure of cost stickiness, we find evidence for a negative relationship considering both directions of causality. We further separate the discretionary component from our aggregate income smoothing measure and show that the negative relation is primarily driven by opportunistic, garbling managerial behavior. Our results corroborate the important role that managerial discretion plays in financial accounting as well as cost behavior. JEL Classifications: D22; M41; M52 Keywords: Asymmetric Cost Behavior; Income Smoothing; Managerial Discretion; Earnings Management

2 INTRODUCTION 2 Managerial discretion and its underlying motives are a fruitful area of research in management as well as financial accounting. Agency theory assumes that managers do not simply act in the best interest of the firm but also aim to maximize their personal utility, for instance, by taking accounting or resource adjustment decisions to meet specific targets (e.g., Hope and Thomas, 2008; Chen et al., 2012). Our study combines two research streams that are fundamentally based on managerial behavior but have thus far been examined entirely independent of each other: Cost stickiness and income smoothing. Cost stickiness is a well-documented property of corporate costs. On average, costs seem to decrease less during reduced firm activity than they grow for an equivalent increase (Anderson et al., 2003). Initially, the cost stickiness literature suggests that asymmetric cost behavior primarily arises because managers take rational decisions as they trade off adjustment costs of committed resources against corresponding holding costs. However, there is also the possibility that opportunistic managers discretionarily refrain from cutting costs in periods of downturn, which also induces cost stickiness (Anderson et al., 2003). Banker and Byzalov (2014) note that costs are a fundamental component of accounting earnings. As such, financial accounting decisions affect cost behavior and vice versa. Hence, prior research provides initial evidence on the interaction between motives for cost stickiness and earnings management (Kama and Weiss, 2013; Dierynck et al., 2012). They show that managerial incentives to increase reported earnings suppress motives for asymmetric cost behavior (i.e., managers are more inclined to cut costs when facing a decrease in activity). However, these studies focus on a general association of earnings management and cost stickiness. Earnings smoothing, arguably the most comprehensive and long-term form of accounting earnings management, has been disregarded by the cost stickiness-related literature. As a result, we answer prior recommendations for additional research which fur-

3 3 thers our understanding of the interplay of management and financial accounting (Weiss, 2010; Banker and Byzalov, 2014). Income smoothing refers to the deliberate moderation of fluctuations in earnings (e.g., Beidleman, 1973). Prior literature separates motives for reporting smoothed earnings into an "informational" and "garbling" part (Tucker and Zarowin, 2006). The informational role entails the provision of private information about the firm s economic stability and future performance (e.g., Sankar and Subramanyam, 2001; Kirschenheiter and Melumad, 2002). In contrast, the garbling role refers to managerial opportunistic behavior to gain private benefits (e.g., DeFond and Park, 1997). Recent literature emphasizes the importance of distinguishing between these two roles in the examination of income smoothing behavior (Cahan et al., 2008; Dou et al., 2013). We hypothesize that cost stickiness and income smoothing are negatively associated. Both phenomena are based on managerial motives to target specific objectives. However, an asymmetrical reaction of costs in response to changes in activity induces a bumpy earnings path. Thus, it counteracts managerial ambitions to report a smooth income stream. Therefore, we assume a trade-off between the managerial motives for income smoothing and cost stickiness as the underlying cause for the negative relation. First, we test our prediction by examining the impact of income smoothing on cost asymmetry replicating the cross-sectional model of cost stickiness (Anderson et al., 2003). To proxy for earnings smoothing behavior, we adopt an aggregate income smoothing score (Dou et al., 2013). Using a sample of listed North American firms over the period , we show that increased engagement in income smoothing significantly mitigates the degree of cost stickiness. More importantly, we provide evidence that firms classified as smoothers do not exhibit asymmetric cost behavior at all. Second, since managers may also prioritize taking cost decisions to minimize future adjustment costs, we reverse the direction of causality and examine the impact of cost sticki-

4 4 ness on income smoothing. Applying a firm-specific cost stickiness measure (Weiss, 2010), we find a significantly negative association with income smoothing. Cost stickiness seemingly constrains the ability to smooth earnings. Furthermore, we decompose aggregate income smoothing into the informational and garbling component (Dou et al., 2013). Our findings show that cost asymmetry negatively influences both components, but the effect is stronger and only significant for the garbling role. Our empirical findings are robust to a wide range of modifications. For example, the main inferences hold when employing alternative individual income smoothing measures, altering sample selection or variable generation processes or controlling for additional potentially influential factors. Nevertheless, we acknowledge that the documented statistical associations do not necessarily imply causality. The remainder of this paper is organized as follows. Section 2 provides the theoretical background as well as hypotheses development. Section 3 explains our research design choices and describes the samples. Section 4 discusses our empirical findings. Section 5 concludes. BACKGROUND AND HYPOTHESES DEVELOPMENT Research at the interface between financial and management accounting with respect to cost stickiness is rather scarce. A growing body of evidence shows an asymmetrical reaction of costs to changes in firm activity (i.e., sales). Particularly, Anderson et al. (2003) show that selling, general, and administrative (SG&A) 1 costs increase, on average, by 0.55% per 1% growth in sales, but decrease only by 0.35% per 1% decline in sales revenue. Hence, costs are sticky. They increase more than they decrease for an equal change in activity. The first reason why managers might refrain from reducing idle capacity are adjustment costs (i.e., costs of reducing capacity and replacing resources when activity is restored). 1 Asymmetric cost behavior has also been shown for other cost categories, such as costs of goods sold (Subramanian and Weidenmier, 2003), operating costs (e.g., Calleja et al., 2006; Kama and Weiss, 2013) or labor costs (Dierynck et al., 2013).

5 5 They are determined by monetary factors, such as selling costs of assets, severance payments for dismissed employees or integration costs for new employees (e.g., search and training costs). However, especially in the long-term, also non-pecuniary, indirect costs, like lower morale among remaining staff, loss of reputation or specialist employees, can be considered as adjustment costs. In comparison, holding costs of retaining capacity are rather monetary and include, for instance, wages, rents or energy costs. Consequently, asymmetrical cost behavior occurs when managers perceive adjustment costs to outweigh holding costs. Besides, Anderson et al. (2003) identify personal considerations by self-interested managers as a second reason for the existence of cost stickiness. Some managerial motives might include fear of job loss when downsizing a division or revising past overambitious investment decisions or the anguish of dismissing familiar employees. Recent work in this area focuses on the relationship between asymmetric cost behavior and (upward) earnings management. 2 Since costs are a fundamental determinant of earnings, manipulating accounting data may also influence cost behavior (Banker and Byzalov, 2014). As such, Dierynck et al. (2012) find that managerial incentives to avoid reporting losses and to meet the zero earnings benchmark are negatively associated with cost asymmetry. Executives seem to be more willing to cut labor costs when facing a sales decrease. Similarly, Kama and Weiss (2013) find that motives to meet short-term earnings targets induce deliberate resource adjustments. Cumulatively, both papers show that upward earnings management fosters more decisive cost cutting and, hence, reduces cost stickiness. To enhance our understanding at the intersection of cost and financial accounting, we address the link to income smoothing, a specific form of earnings management that has been entirely disregarded in cost stickiness-related literature thus far. Income smoothing refers to the "intentional dampening of fluctuations about some level of earnings that is currently 2 Other examples on the association of cost stickiness with properties of financial accounting include Banker et al. (2016) who investigate the confounding effect of sticky costs on conditional conservatism.

6 6 considered to be normal for a firm" (Beidleman, 1973, 653). 3 As such, both income smoothing and cost stickiness result from managerial discretion based on specific incentives. While some motives induce a smooth earnings path, others result in an unbalanced reaction of costs in the wake of activity changes, implying a bumpy earnings path. Prior findings of Dierynck et al. (2012) and Kama and Weiss (2013) already indicate an incompatibility between some incentives for earnings management and cost stickiness since "some deliberate decisions induce sticky costs while others diminish sticky costs, depending on the underlying motivations" (Kama and Weiss, 2013, 203). However, the direction of causality between income smoothing and cost stickiness decisions is unclear. Therefore, we initially predict a general trade-off between income smoothing and cost stickiness. H1: Income smoothing and cost stickiness are negatively associated. There are two different theoretical approaches to specify H1 regarding the direction of causality. Dierynck et al. (2012) and Kama and Weiss (2013) state that managers facing specific earnings targets are more inclined to adjust unutilized resources. Hence, they argue that executives base their decisions on earnings management incentives and changes in the level of cost asymmetry are merely a result. Following this line of reasoning, a manager will put the emphasis on the decision whether or not to report smooth earnings. If there are sufficient personal and firm-related incentives, she will try to prevent volatile earnings. Since costs are a fundamental part of earnings (Banker and Byzalov, 2014), adjusting costs is a possible smoothing instrument. Cyert and March (1963) provide a theoretical example for the use of costs for smoothing considerations. They argue that managers will absorb excess resources as costs in positive periods and institute cost reduction arrangements in periods of downturn. In this scenario, the adopted income smoothing measures consequently align costs with the activity level and thus mitigate cost stickiness. This leads to our next, more specific hypothesis, in line with H1: 3 A vast amount of research provides evidence on the existence of income smoothing. Already Gordon (1964) develops a theoretical model predicting that managers will smooth earnings as long as they can exercise discretion over accounting choices.

7 7 H2: Increased income smoothing mitigates cost stickiness (i.e., firms engaged in earnings smoothing activities exhibit a lower degree of cost asymmetry). As shown above, cost stickiness is a phenomenon based on managerial discretion. Consequently, assuming a setting where managers prioritize cost over accounting choices, the direction of causality can also be reversed. In this context, cost behavior will have implications on income smoothing. Anderson et al. (2003) argue that managers weigh the adjustment and holding costs in periods of downturn in order to decide whether to cut or retain committed resources. However, adjustment costs are rather indirect, subjective and forwardlooking (e.g., Anderson et al., 2003; Banker et al., 2014) and managers moreover consider personal interests (e.g., Anderson et al., 2003; Chen et al., 2012). Therefore, from the accounting perspective, not to adapt unnecessary resources and cut costs leads to even more decreased earnings in periods with reduced revenues. This heightens earnings volatility between periods and thus restricts managers in targeting smoothing-related incentives. Based on these considerations, again consistent with H1, we state the third hypothesis as: H3a: Increased cost stickiness constrains engagement in income smoothing activities (i.e., cost sticky firms are less likely to report smooth earnings). If H3a holds, the relationship can be further specified. Recent studies show the importance of distinguishing between the garbling and informational role of income smoothing (Tucker and Zarowin, 2006; Cahan et al., 2008; Dou et al, 2013). The informational role of smoothing is to provide private information about future performance in order to stabilize forward-looking stakeholder relationships (Dou et al., 2013) or debt-related contracting (Gassen and Fülbier, 2015), and can thus be considered as a rational corporate action. In contrast, garbling smoothing reflects personal motives of self-interested managers (Tucker and Zarowin, 2006). These can be manifold and range from personal job protection (DeFond and Park, 1997), securing high incumbency rents (Fudenberg and Tirole, 1995) to

8 8 compensation incentives (e.g., Grant et al., 2009). Following this distinction, garbling smoothing can be considered as an even more pronounced form of managerial discretion. According to similar motivations, it is also possible to differentiate between "good" and "bad" cost stickiness (Brüggen and Zehnder, 2014). While good stickiness results from common cost-benefit considerations in the interest of the firm, bad stickiness is the consequence of discretionary decisions by opportunistic managers. Based on these considerations, we conjecture, in line with H1, that particularly the incentives causing managerial opportunism in cost behavior and earnings smoothing are incompatible and cannot be targeted simultaneously. This discussion amplifies H3a and results in our final hypothesis: H3b: Increased cost stickiness primarily constrains managerial engagement in the garbling part of income smoothing activities (i.e., the relationship in H3a is primarily driven by the garbling role of earnings smoothing). RESEARCH DESIGN AND SAMPLES Aggregate Income Smoothing Score To proxy for income smoothing, we follow Dou et al. (2013) and create an aggregate score from three individual measures developed in prior studies. The first two are 3 and 4, adopted from Burgstahler et al. (2006), and the third is, developed by Tucker and Zarowin (2006). However, while Dou et al. (2013) calculate the variables at the country-industry level, we compute them at the firm- as well as the firm-year level. The following decomposition procedure requires at least a variable at the firm- level and only a firm-year level measure can be deployed in the pooled cross-sectional cost stickiness regression model. 3 is calculated as the firm-level ratio of the cross-sectional standard deviations ( ) of operating income ( ) and operating cash flow ( ), both deflated by lagged total assets ( ). It captures the extent to which managers diminish the variability among earnings through the use of accruals. Since a considerably smaller standard deviation for income than

9 9 cash flow indicates accrual smoothing activities, lower values for 3 suggest more smoothing. However, to avoid potential confusion, we multiply the term by minus one such that larger values imply more income smoothing: 4 3, /, (1), /, 4 likewise captures accrual smoothing activities. It is defined as the firm-level Spearman correlation ( ) between the change in total accruals ( ) and the change in operating cash flows. 5 The measure is expected to be negative because a certain degree of earnings smoothing occurs naturally due to accrual reversals (Dechow, 1994). However, a more negative correlation suggests that managers deliberately use accruals to counteract volatility in cash flows (e.g., Leuz et al., 2003). Since the term is again multiplied by minus one, greater values imply more earnings smoothing: 4,,, (2) is the firm-level Spearman correlation between the change in discretionary accruals ( ) and the change in pre-managed income ( ; i.e., net income minus discretionary accruals). Accordingly, Tucker and Zarowin (2006) assume an underlying pre-managed income series that managers try to smooth through discretionary accruals. Discretionary accruals are estimated employing the performance matched cross-sectional Jones model, as proposed by Kothari et al. (2005).,, 1/,,,,,,,, (3),, 1/,,,,,,, (3a) 4 5 Since extreme downward outliers are observed, 3 is winsorized only at the 1 st percentile to avoid additional data loss (Burgstahler et al. (2006) truncate both the 1 st and 99 th percentile). The correlation-based variables 4 and have natural limits (1 and 1) and need not to be winsorized. Following Dechow et al. (1995), the accrual component of earnings is calculated as:, where signifies the change over the last fiscal year. is calculated as operating income less total accruals.

10 10,,,,, (3b) Total accruals ( ), change in sales ( ) and gross property, plant and equipment ( ) are each deflated by lagged total assets ( ). Return on assets ( ) is net income scaled by lagged total assets. To avoid stringent data requirements and industryspecific effects, regression (3) is estimated on all firms in the same two-digit standard industrial classification (SIC), separately for each year. 6 Non-discretionary accruals ( ) are the fitted values and discretionary accruals the residuals of the regression. A more negative correlation between discretionary accruals and pre-managed income implies artificial earnings smoothing. The term is again multiplied by minus one to ensure that higher values of indicate more income smoothing:,,, (4) In order to compute firm-year level measures, we modify the presented equations (1), (2) and (4). Consistent with Francis et al. (2004), 3, is additionally computed by calculating the standard deviations over a rolling five-year window. 7 Furthermore, 4, and, are calculated using the current year s and past four year s observations (Cahan et al., 2008; Tucker and Zarowin, 2006). These modifications also alter the underlying assumption of stable smoothing behavior. Instead, firm-year level smoothing values take possibly changing incentives for smoothing during the sample period into account. To obtain the aggregate income smoothing score ( ), we follow Dou et al. (2013) and extract the first component of a principal component analysis (PCA). 8 Corresponding to the individual measures, a higher value of indicates greater income Consistent with Tucker and Zarowin (2006), we require at least 10 observations for each industry-year crosssection and we winsorize the regression variables at three standard deviations each year. Francis et al. (2004) deploy a rolling ten-year period. However, we employ a five-year period attributable to data restrictions and consistency with the firm-year level calculation of 4 and in Cahan et al. (2008) and Tucker and Zarowin (2006). The value of the Kaiser-Meyer-Olkin criterion exceeds 0.6 for every model indicating an acceptable implementation of the PCA. Furthermore, the extracted component explains approximately 60 per cent of the variance (Eigenvalue 1.7; Uniqueness 0.4), on average.

11 11 smoothing. Moreover, the aggregate measure has the advantage of mitigating potential measurement errors in the individual scores. Building on the continuous variable, we further create a dichotomous indicator variable ( _ ) to distinguish between smoothing and non-smoothing firms or firm-year observations. The binary variable is coded as 1 if the value for is in the upper quartile, and 0 otherwise. We take industry peculiarities into account by calculating the quartiles for each 2-digit SIC industry, separately. 9 Decomposition of Income Smoothing In H3b, we hypothesize that the negative impact of cost stickiness on income smoothing is primarily driven by a trade-off between managerial motives for cost stickiness and the garbling role of earnings smoothing. To decompose the income smoothing measure into its garbling and informational components, we follow the procedure of Dou et al. (2013). They modify an approach by Tucker and Zarowin (2006) which builds on the prices-leadearnings framework (i.e., the extent of information about future earnings that is reflected in current stock returns). Tucker and Zarowin (2006) implement this model to analyze whether income smoothing enables investors to better predict future earnings. The underlying idea is that the informational role of income smoothing is to provide private, additional information about the firm s performance (e.g., Sankar and Subramanyam, 2001; Kirschenheiter and Melumad, 2002). Income smoothing in the current period is public information the investors will use to predict future earnings and thus price the firm accordingly. However, when managers use the garbling role of income smoothing for opportunistic reasons, they do not provide sustainable information about future earnings. To measure the extent to which prices lead earnings, we use the following model (Tucker and Zarowin, 2006; Dou et al., 2013): 9 We address the critical points in these assumptions by additionally applying a median split and omitting the industry separation. Untabulated results show that our main inferences remain unchanged.

12 12 (5) and are earnings per share (EPS) for years and 1 and is the sum of EPS for 1 to 3. All EPS variables are basic EPS excluding extraordinary items, adjusted for stock splits and stock dividends, scaled by lagged stock prices. are annual stock returns, adjusted for stock splits and dividends, which incorporate all publicly available information. Additionally, denotes the aggregate future stock return in years 1 to 3 with annual compounding. Current and past EPS control for unexpected earnings in the current period. However, pertinent to this analysis are the future earnings response coefficients (FERC) and that measure revisions in investors expectations about future earnings. Following Dou et al. (2013), we differentiate between cumulative three-year EPS being positive ( ) or negative ( ), assuming that earnings of profitable companies are easier to predict. Since we employ firm- level data, whereas Dou et al. (2013) operate with countryindustry units, the decomposition based on the presented framework is slightly modified. We initially run regression (5) for all firms with sufficient firm-year and particularly lossyear observations to estimate the firm-level FERCs and. 10 Next, within each twodigit SIC industry, with a minimum of 20 firms, we estimate regression (6) with the corresponding firm- level FERC variables. The predicted values of proxy for the firmlevel informational part of income smoothing ( _ ) while the residuals represent the corresponding garbling component ( _ ): (6) _ _, (6a) (6b) 10 Following the firm-level robustness test of Dou et al. (2013, 1650), we require at least 10 firm-year observations (see the sample selection process).

13 13 _ captures the income smoothing role to provide private information that enables investors to predict future earnings more precisely. On the contrary, _ reflects the smoothing component that is unrelated to information about future performance. Instead, the variable measures opportunistic motives for income smoothing. Cross-Sectional Measure of Cost Stickiness To examine H1 and H2, we test for cost stickiness through the cross-sectional model introduced by Anderson et al. (2003). The model analyzes the relationship between changes in costs and simultaneous changes in sales in order to proxy for asymmetric cost behavior. Equation (7) represents the basic model:,,,, _,,, (7) Subscripts ( ) again reflect firm (time) indices, respectively., and, denote operating costs of firm in year and 1. We employ operating costs instead of SG&A costs, used by Anderson et al. (2003), because they capture a broader extent of managerial cost choices. For example, operating costs include costs of manufacturing goods, providing services and costs of marketing and distribution. Following Kama and Weiss (2013), we compute operating costs as annual sales revenue less income from operations., and, are sales revenues. In order to distinguish between periods with a positive and negative change in activity, the indicator variable Decrease_Dummy, is integrated as an interaction term with the change in sales. Accordingly, Decrease_Dummy, equals 1 if sales of firm decrease between fiscal years 1 and, and 0 otherwise. Applying ratios and log-specifications enhances the comparability across firms and moderates potential heteroscedasticity (Anderson et al., 2003). Furthermore, logspecifications enable easier interpretation of coefficients as percentage changes in operating costs and sales. measures the average percentage growth in operating costs when

14 14 sales increase by 1%, while the sum of and represents the average cost decline for a 1% decrease in sales. Therefore, a significantly negative value of indicates cost stickiness, given a positive value of. Firm-Specific Measure of Cost Stickiness A main disadvantage of the cross-sectional regression model is that it cannot serve as an explanatory variable to examine implications of cost stickiness on income smoothing. However, such a variable is necessary to investigate H3a and H3b. Therefore, we additionally employ a firm-specific measure of cost stickiness adopted from Weiss (2010). The variable is premised on quarterly data but conceptualized to generate annual firm observations. Equation (8) illustrates the slightly modified cost asymmetry measure:,,,,, 3 (8), Working with quarterly data, denotes the change in sales revenue between two quarters (,, ) and the corresponding change in costs (,, ). Consistent with Weiss (2010) and the preceding cross-sectional model, we employ operating costs. Subscript signifies the most recent of the last four quarters in fiscal year with a decrease in sales and the most recent of the last four quarters with an increase in sales. The variable compares the slope of a cost function between the two most recent quarters in fiscal year, one with an increase and the other with a decrease in sales. If costs increase more when activity rises than they decrease when activity declines by an equivalent amount, they are sticky and the term in square brackets exhibits a negative sign. Although this sign corresponds to the stickiness measure in the cross-sectional model, we multiply the term by minus one to create consistency with the aforementioned smoothing measures. As a consequence, higher values of imply more sticky cost behavior and indicate that there are more managerial incentives not to reduce idle capacity.

15 15 The approach entails the serious disadvantage of a significant data loss. For example, we follow Weiss (2010) and restrict the quarterly sample to observations for which sales and costs change in the same direction. Moreover, the measure only identifies values for firm-years which entail both a sales increase and decrease in at least one of the four quarters. Nevertheless, the benefits of the approach outweigh these data restrictions. As has been pointed out, it is possible to apply the variable as an independent variable to examine the effects of cost stickiness. Furthermore, it provides an evident reference point of a linear cost function ( 0) to distinguish between sticky ( 0) and anti-sticky ( 0) firms. Based on this, we supplementary create an indicator variable ( _ ) that equals 1 if the value for is positive, and 0 otherwise. Multivariate Cost Stickiness Model To estimate the hypothesized effect of income smoothing activities on cost stickiness (H2), we specify the basic model (7). Initially, we include the firm-year level aggregate income smoothing score (, ) as a three-way interaction term. Furthermore, in line with prior literature (e.g., Anderson et al., 2003; Chen et al., 2012), the measure is supplementary integrated as a singular variable to control for its general impact on cost changes. 11 In the next specification, we add economic control variables (Anderson et al., 2003). These determine monetary as well as indirect components of adjustment costs. _ is calculated as the natural logarithm of the number of employees divided by sales. Similarly, the asset ratio ( _ ) is computed as the natural logarithm of the ratio of total current assets to sales (Anderson et al., 2003). Furthermore, we control for successive sales decreases since managers pessimism, as they might consider negative demand shocks to be more permanent after two consecutive downturns, is expected to mitigate cost stickiness (Anderson et al., 2003; Banker et al., 2014). We create an indicator variable _ that equals 1 if sales decrease in two consecutive 11 Moreover, we will add a two-way interaction term between the smoothing measure and the natural logarithm of change in sales as a robustness test, following Kama and Weiss (2013) or Dierynck et al. (2012).

16 16 years, and 0 otherwise. All economic controls are again included as three-way interaction variables and main terms. 12 Since the presented asset or employee ratios should vary by industry, and further industry-specific factors such as the level of competition or special regulation may also influence adjustment costs, industry classifications can also play an important role for the occurrence of cost stickiness (Subramaniam and Weidenmier, 2003). Therefore, we additionally include industry-fixed effects based on two-digit SIC codes. Finally, since the cost stickiness model requires pooled cross-sectional regressions, we add year- fixed effects to control for potential unobserved macroeconomic factors that may change over time but affect all firms in a similar fashion (Chen et al., 2013). Taking all these adjustments of the basic model (7) into account, equation (9) represents the extended cost stickiness model:,,,, _,, _,,, _,, _, _,, _, _, _,,, _, _, _, _ _, (9) If is positive and significant, we find evidence that income smoothing measures mitigate cost stickiness further supporting H1 and particularly H2. Employing the smoothing indicator variable, we supplementary compare the degree of cost asymmetry among smoothing 12 Prior research, particularly Dierynck et al. (2012, 1227), shows that these variables do not affect increases in sales and are, therefore, not included as two-way interaction terms.

17 17 and non-smoothing firm-year observations by conducting a subsample-analysis on the basic model (7) following the approach of Kama and Weiss (2013). In line with H1 and H2, we expect a more symmetric cost behavior for smoothing firm-year observations. We employ standard errors that are robust to autocorrelation as well as heteroscedasticity and clustered at the firm- level in all specifications. Multivariate Income Smoothing Model The hypothesized impact of cost stickiness on income smoothing (H3a and H3b) is examined by creating regression models with the income smoothing measures, _ and _ as dependent variables and the firm-level cost stickiness measure ( ) as an explanatory variable. All variables are computed at the firm-level implying only one observation per company. Consistent with prior research (e.g., Dou et al., 2013; Gassen and Fülbier, 2015), we include variables for firm size ( ), sales growth ( ) and profitability ( ) to control for firm-specific operational factors that inherently induce a naturally smoothed earnings path. Large and well-run firms have more stable and diversified operations and, thus, should exhibit a smoother income stream. is calculated as the natural logarithm of total assets, is the annual percentage change in sales revenue and is net income scaled by lagged total assets. We further control for the book-to-market ratio ( ), the debt ratio ( ) and industry litigation risk ( ) to proxy for managerial motives to discretionarily smooth earnings (e.g., Matsumoto, 2002; Dou et al., 2013; Lee et al., 2015). is a common proxy for growth opportunities. Managers of such firms might be subject to greater pressure from debt- and shareholders to smooth earnings in order to increase the options to finance future growth. Prior research also shows that firms with higher leverage are more likely to report smooth income to meet creditor s expectations and avoid debt covenant violations (e.g., De- Fond and Jiambalvo, 1994; Gassen and Fülbier, 2015). Moreover, we consider industry litiga-

18 18 tion risk since Matsumoto (2002) reports that managers of firms in industries facing a high litigation risk are also more inclined to smooth earnings. is a dummy variable coded as 1 if a firm operates in an industry with high litigation risk, and 0 otherwise (Francis et al., 1994). To create firm-level measures, we again follow Dou et al. (2013), and compute the median values of all metric variables. We control for industry-fixed effects by including binary two-digit SIC code indicators. Year-fixed effects are unnecessary attributable to the model s firm-level design. Furthermore, we again use standard errors robust to autocorrelation and heteroscedasticity, clustered by industry and/or firm. In the aggregate, we estimate the following model (10) to test the impact of cost stickiness on income smoothing: (10) _ The variable of interest is. If is negative and significant, the results support H3a that a higher degree of cost asymmetry reduces the potential for income smoothing. To test H3b, we further specify model (10) with the decomposed parts of income smoothing, as shown in equation (11): _ _ _ (11) Consistent with H3b, we particularly expect to be negative and significant for the _ component. More precisely, we expect _ _ meaning that a higher level of cost stickiness primarily constrains engagement in garbling earnings smoothing.

19 Data and Samples 19 Our initial sample comprises 36,481 firms based on 470,245 firm-year observations from the Compustat NA fundamentals file. Moreover, we extract data from the Compustat NA fundamentals quarterly files to compute Weiss (2010) firm- level cost stickiness measure. All samples are consistently designed to span the period 1980 to 2011 because we need extensive data to compute our variables. However, the different models require distinct sample selection processes. Table 1 summarizes these procedures. [Insert Table 1 about here] Panel A describes how we construct our final sample for the cost stickiness model. It requires annual data and firm-year level variables. The sample selection process mostly follows Anderson et al. (2003). Since we target a final period from 1980 to 2011, we start by only considering observations between 1974 and Six additional previous years are necessary because we require two prior years to compute some variables, such as the change in total accruals, and four further previous years in order to calculate the standard deviations and correlations for our firm-year level smoothing variables over a rolling fiveyear window. Next, we exclude financial institutions and public utilities (SIC codes and ) as their financial statements are incompatible to other firms (Subramaniam and Weidenmier, 2003). Moreover, we require operating costs and sales to be available in the current and previous fiscal year and operating costs to be less than sales revenue (Anderson et al., 2003). Furthermore, we delete observations with changes in sales and costs in the top and bottom 0.5 percent of the distribution (e.g., Chen et al., 2012). Cumulatively, this results in a final sample of 44,414 (6,242) firm-year (firm) observations, respectively. Panel B describes sample selection for the income smoothing model. Similarly to Dou et al. (2013), we define a unit of analysis at the firm- level. Therefore, the standard deviations and correlations are also computed at the firm- level and the analysis requires only

20 20 two previous years to calculate. Hence, we exclude observations prior to 1978 and after 2011, again resulting in a final period covering the years For consistency, we again exclude all financial and public utility firms. In addition, we merge the annual base data with the firm-level cost stickiness measure, based on quarterly data. For the remaining intersection, we require data availability to compute the smoothing variables. Finally, firms with less than 10 firm-year observations and industries with less than 100 firm-year observations are excluded (Dou et al., 2013). These selections are particularly necessary for the decomposition of income smoothing, but are already employed in this aggregate model for consistency. Moreover, firm- level results based on fewer observations would be rather unrepresentative. The final sample comprises 2,248 firms (based on 34,820 firm-year observations). The sample selection process for the decomposed income smoothing model shown in Panel C is similar but needs even stricter requirements. We again delete observations prior to However, we require three years of additional data through 2014 to compute cumulative EPS and stock returns. Again, this results in a sample period from 1980 to We further delete observations that are in the top or bottom 1% of the distribution of EPS and stock returns to minimize the effect of outliers on equation (5) (Tucker and Zarowin, 2006). Although we again exclude firms with less than 10 and industries with less than 100 observations, we additionally delete firms or industries with insufficient observations to perform the decomposition. This is for example the case if firms have no loss observations to

21 21 estimate regression (5) or if industries comprise less than 20 firms to run regression (6). The remaining number of firms is 431 (based on 6,514 firm-year observations). 13 [Insert Table 2 about here] Table 2 presents more sample characteristics by comparison. Panel A shows the industry distribution among the three models. For parsimony, we summarize the observations by Fama and French s (1997) 12 industry classification. Although there are slight differences among the three samples, we note that shops, manufacturing, business equipment and other are the most represented industries. Panel B presents the distribution of firm-year observations over the sample period. As intended, all samples span the period However, while the distributions of model [2] and [3] resemble each other quite closely, the distribution of the cost stickiness model [1] differs significantly. This can be traced back to the diverging data manipulation procedures as well as the distinct units of analysis implying the different computation of the smoothing variables. Descriptive Statistics RESULTS Distributional properties of the variables are presented in Table 3. For reasons of brevity, we focus on those variables which are pertinent for the analyses. Underlying variables such as total and discretionary accruals or operating cash flows are omitted. [Insert Table 3 about here] 13 The different requirements result in distinct characteristics of the three final samples. Sample sizes vary significantly between the models and are generally smaller than those of prior research, although we span a relatively longer period. The main reason for the relative small sample size of the cost stickiness model with 44,414 firm-year observations (compared to 64,663 in Anderson et al., 2003), is the rolling firm-year calculation of the smoothing variables that induce a sharp data loss. Furthermore, divergent sample selection and variable choices, such as the use of operating instead of SG&A costs, lead to further differences in sample size compared to previous research. The sample sizes of model [2] and [3], the aggregate (2,248 firms) and decomposed income smoothing model (431 firms), are also smaller than samples presented in Dou et al. (2013) with 2,628 or 1,523 country-industry observations, respectively. This is because Dou et al. (2013) use a larger, global data set and our analysis at the firm-level needs stricter requirements for the decomposition than the analysis with country-industry-units. Additionally, the application of the firm-level cost stickiness measure also entails significant data loss.

22 22 Panel A refers to univariate statistics of the cost stickiness model. Mean sales revenues, the frequency of sales declines, the mean ratio of operating costs to sales are in line with prior literature (e.g., Chen et al., 2012; Kama and Weiss, 2013). Although computed at the firmyear level, the aggregate income smoothing score and the underlying individual smoothing variables differ only insignificantly from the firm-level values presented in Panel B, or even the country-industry scores in prior literature (Dou et al., 2013). As assumed, 3 yields negative values, whereas 4 and generally exhibit positive values, indicating smoothing behavior across the sample. The indicator variable _ identifies 25% of the observations as smoothers, as a mechanical result of its computation. Employee and asset intensity are in line with prior research, as well. Panel B documents distributional statistics for the aggregate smoothing model. As mentioned above, the smoothing variables have the expected signs and exhibit distributions that are consistent with prior research ( Dou et al., 2013). The firm-level cost stickiness measure yields a positive mean (median) value of (0.047). As assumed, the sample is subject to sticky cost behavior, on average. The indicator variable _ shows that a large majority of sample firms (67.1%) exhibits cost stickiness. 14 The average firm in our income smoothing sample has a debt ratio of 25.9% and is growing and profitable. A mean book-to-market ratio below the value of one suggests growth opportunities. Furthermore, 26.6% of firms are subject to increased litigation risk. Panel C provides information about the distributional properties of the decomposed income smoothing variables _ and _. Moreover, based on the underlying 6,514 firm-year observations, we present descriptives for the EPS and stock return variables employed in the decomposition. Their values are in line with prior research in this area (Tucker and Zarowin, 2006; Cahan et al., 2008). On average, firms yield positive EPS and 14 Considering the untabulated underlying firm-year observations, 56.4% of the 34,820 firm-years display positive sticky values, which is similar to the ratio of 53.2% in the study of Weiss (2010).

23 23 stock returns. More precisely, the indicator variable shows that almost three quarters (73.3%) of our observations yield positive cumulative EPS. Table 4 reports correlation matrices for our different models. Each matrix provides Spearman (Pearson) correlations above (below) the diagonal, respectively. Bold font indicates significant correlations at a 0.1 level (or higher). With respect to the cost stickiness model, the cross-sectional design to identify asymmetric cost behavior does not provide particularly meaningful pairwise correlations. Hence, the majority of cost stickiness-related research refrains from presenting correlations. However, we still believe this is valuable to identify potential problems with multicollinearity. [Insert Table 4 about here] The majority of correlations for the cost stickiness model presented in Panel A are significant but small in magnitude. As such, collinearity should not be a concern. Moreover, variance inflation factors (VIFs) are generally below the critical textbook level of 10 and, on average, even below 4. Panels B and C exhibit correlations for the income smoothing models. For brevity, both models are combined. Thus, all variables apart from the decomposed smoothing variables _ and _ have dual functions. The pairwise correlations among these variables represent the aggregate income smoothing sample. However, if correlated to the decomposed scores, they represent the decomposed sample data. Again, we are not concerned with collinearity. Furthermore, VIFs for these regression models are mostly even below 2. Unlike for the cost stickiness model, this correlation matrix further yields some interesting insights. The Pearson (Spearman) correlation coefficient between and amounts to ( 0.114) and is highly significant at the 0.01 level. This indicates the hypothesized negative relationship between cost stickiness and income smoothing (H1). With respect to the decomposition, the Spearman correlation is expectedly only negative and significant ( 0.137) for the garbling component (H3b). Equally support-

24 24 ing H3b, the Pearson correlation is significant and negative for both decomposed variables, but the garbling part exhibits the more negative and significant relationship ( vs ; p 0.01 vs. p 0.1). With respect to the decomposed variables, _ is significantly associated with factors which naturally affect a firm s earnings smoothness, such as firm size or profitability. Interestingly, _ exhibits significant correlations with variables proxying for managerial incentives to smooth earnings, such as leverage, book-tomarket ratio or industry litigation risk. Association of Income Smoothing with Sticky Cost Behavior Subsample Analysis In H2 we predict that cost behavior should be more symmetric (i.e., less sticky) when managers engage in income smoothing. To test this prediction, we follow Kama and Weiss (2013) and initially conduct a subsample analysis on the basic cost stickiness model by Anderson et al. (2003). First, we estimate the basic model (7) for the full sample to identify average cost behavior. Then, we partition the sample into smoothing and nonsmoothing firm-year observations, through the smoothing indicator ( _ ) and re-run equation (7) for both subsamples. The presented t-statistics are based on robust standard errors, clustered at the firm-level. Table 5 reports the results. [Insert Table 5 about here] In the full sample, both the adjusted and values exceed 0.90 implying that changes in operating costs are extensively explained by changes in sales revenue and that these changes are nearly proportional. As expected, the coefficient is positive and smaller than one and the sign of is negative and highly significant at the 0.01 level using a two-tailed test (t 7.28). Total operating costs increase, on average, by 0.942% per 1% increase in sales revenue, but decrease only by 0.881% ( ) when sales fall by an equivalent amount. This documents that our full sample is subject to cost stickiness.

25 25 Considering the two subsamples, the estimate of has a negative sign ( 0.066), significant at the 0.01 level, for non-smoothing observations ( 6.67). Consequently, operating costs are sticky for this subsample. However, for observations classified as smoother, the estimate of is smaller ( 0.015) and insignificant, implying no evidence for asymmetric cost behavior. The difference between the coefficient values of in both subsamples is positive ( ) and highly significant. This supports our hypothesis H2 that firms engaging in income smoothing activities exhibit a significantly lower degree of cost stickiness. More importantly, the results even suggest that operating costs of smoothing firms are not sticky at all. The extent of the slope for sales decreases ( ) shows how intensively managers react to changes in revenues in their cost decisions. The difference between the values of both subsamples is positive ( ) and highly significant. In line with our prediction in H1, managers with incentives to smooth earnings deliberately seem to cut costs more aggressively in periods with declining revenues to balance earnings paths. Comprehensive Regression Analyses To comprehensively test H1 and H2, controlling for other potentially influential factors, we expand the basic model (7). As has been pointed out, we report four specifications. Specification 1) starts with the baseline cost stickiness model including the aggregate income smoothing score as a three-way interaction term and stand-alone variable. In the second specification, we add economic control variables for cost behavior. In the third specification, we include year-fixed effects, and specification 4) additionally considers industry-fixed effects. Once more, we report robust t-statistics based on heteroscedasticity-consistent standard errors, clustered at the firm-level. We present these results in Table 6. [Insert Table 6 about here]