The Association Between Inter-Segment Profit Smoothing and the Conservatism of Accounting Earnings

Transcription

1 The Association Between Inter-Segment Profit Smoothing and the Conservatism of Accounting Earnings Daniel A. Bens University of Arizona Eller College of Management 1130 E. Helen Street, P.O. Box Tucson, AZ Steven J. Monahan INSEAD Accounting and Control Area Boulevard de Constance PMLS No F-7705 Fontainebleau Cedex, France Logan Steele University of Arizona Eller College of Management 1130 E. Helen Street, P.O. Box Tucson, AZ Latest Draft as of: December 26, 2008 Preliminary and Incomplete Please do not Cite or quote without the Authors permission Corresponding author. The authors gratefully acknowledge financial support from the Ratoff Family Fellowship at the University of Arizona and the INSEAD Alumni fund. The authors are solely responsible for any errors.

2 The Association Between Inter-Segment Profit Smoothing and the Conservatism of Accounting Earnings Abstract We devise a measure of inter-segment profit smoothing for multi-segment firms. This measure equals the ratio of the variance of profits for a portfolio of single segment firms to the variance of profits across a firm s multiple business units. The single segment firms are matched to the firm s segments by industry and size. We interpret increases in this ratio as capturing the smoothing of inter-segment profit. Specifically, as the ratio increases the underlying profit volatility (as captured by the control group) has increased while reported volatility stays constant. We find that our inter-segment profit smoothing measure is positively correlated with the level of information asymmetry at the firm (as measured by the bid-ask spread and the probability of an informed equity trade). As our main contribution, we document that increased inter-segment profit smoothing reduces the asymmetric timeliness (or conservatism) in firm level accounting earnings by as much as 24%. We use a Basu [1997] reverse regression framework for this analysis. We also demonstrate that inter-segment profit smoothing reduces the likelihood that the firm records a non-recurring negative accrual, namely an asset impairment. These accruals are generally used to communicate bad economic news, and thus our evidence suggests they are less likely to be recorded when inter-segment smoothing is high.

3 I. Introduction In this paper we examine whether multi-segment firms that publicly report relatively smooth operating performance across their segments exhibit a low association between bad economic news and its recognition in accounting earnings. That is, we test for a negative association between inter-segment profit smoothing and accounting conservatism 1. We measure inter-segment profit smoothing for publicly traded U.S. firms that report more than one business segment in their annual financial reports. Our measure is a decile rank variable within the population of multi-segment firms; the higher the rank within the sample, the higher the classification of smoothing. We first measure the variance of profit margins across a portfolio of single segment firms that are matched on size and industry to the multi-segment firm s business segments; this serves as the numerator of our ranked smoothing variable. We then measure the variance of profit margins across the firm s segments; this serves as the denominator of our ranked smoothing variable. The higher this ratio, the more the firm is using either industry aggregation schemes or cost allocation strategies to make their underlying income appear less volatile across business units than the underlying fundamentals (as captured by the single segment firms) suggest. Our proxy for conservatism is the measure of the asymmetric timeliness of accounting earnings first used by Basu [1997]. Basu s intuition, also articulated by Ball [1998] and Watts [2003] among others, is that U.S. generally accepted accounting principles (GAAP) exhibit an asymmetric bias towards reporting bad news on a more timely basis than good news. These standards restrict the recognition of good news in accounting earnings while encouraging the 1 As we discuss later in this section, our definition of conservatism follows Basu [1997], which is often referred to as conditional conservatism. Watts [2003] and Ryan [2006], among others differentiate this from unconditional conservatism, which consists of an understatement of all assets irrespective of the economic condition of the company. For simplicity, we use the term conservatism to mean conditional conservatism and the asymmetric timeliness of earnings 1

4 recognition of bad news. Further, in the judgmental application of GAAP accountants generally require less verification for the recognition of bad news vis-à-vis good news. Basu uses a reverse regression specification where accounting earnings are the dependent variable and stock returns the independent variable; the coefficient on stock returns is then allowed to vary by whether the measure for the year is positive (i.e., good economic news) or negative (i.e., bad economic news). Basu documents that the coefficient on the bad news is larger than that on good news, in both statistical and economic terms. This evidence confirms, empirically, the presence of an asymmetric bias towards accelerated bad news recognition in applied GAAP. We conjecture that the smoothing of income across segments is a form of information asymmetry in the capital markets. We estimate univariate correlations between our smoothing measure and other publicly available measures of information asymmetry to confirm this prediction. We document that the level of inter-segment smoothing is positively associated with the bid-ask spread and the probability of an informed trade (PIN), both of which have been interpreted in the literature as capturing information asymmetry. We next examine whether inter-segment smoothing is associated with a lower level of conservatism. Given our initial results that inter-segment smoothing is associated with greater information asymmetry, we conjecture that this smoothing obscures the underlying value of firm assets. This obfuscation complicates the task of identifying unrecognized losses (such as asset impairments) and creates uncertainty as to the extent of these losses. Therefore, we expect that inter-segment smoothing reduces the ability of the accounting system to communicate bad news in a timely fashion. Our empirical approach for testing this conjecture is based on the multivariate regression specified by Basu [1997]. In a pooled time-series cross-sectional regression, the dependent variable is annual earnings per share divided by beginning stock price. 2

5 The proxy for economic news is the concurrent stock return. A dummy variable that equals one whenever returns are negative for the year is then interacted with the returns variable, allowing for a differential coefficient for bad news. Our hypothesis variable is an additional interaction term with the negative stock returns variable. This hypothesis variable is the decile ranking of the ratio of the variance of matched single-segment firm profits to the variance of the firm s profits across its multiple segments. We interpret the higher ranked decile as containing firms that are engaging in relatively more inter-segment profit smoothing. Our main regression results are consistent with our prediction that when the smoothing variable is high then the sensitivity of accounting earnings to bad economic news is reduced. We conclude that inter-segment smoothing results in a reduction of the informativeness of earnings, and that this manifests itself via the delayed accounting recognition of economic losses. This is the main contribution of our study. Our results are strengthened by the inclusion of other cross-sectional determinants of asymmetric timeliness identified by Khan and Watts [2007]: firm size, book-to-market, and leverage. We also examine a set of specific accounting accruals designed to communicate bad news and test whether a higher level of inter-segment smoothing reduces their use. We focus on asset write-downs (both tangible and intangible). In a probit regression, we find that the smoothing variable is negatively associated with the likelihood of recording one of these charges, after conditioning on several market and accounting return variables. This provides further evidence that inter-segment smoothing reduces the timely recognition of bad news via firm-level accounting earnings. Our paper contributes to the growing body of literature on the conservatism of accounting earnings. Ball [1998] and Watts [2003] posit that the contracting and governance demands on 3

6 accounting logically lead to the asymmetric timeliness of accounting earnings documented by Basu [1997]. We examine how the combination of organizational structure (i.e., the diversification into multiple business segments) and financial reporting choice (i.e., the smoothing of accounting profits across segments) can reduce the accounting system s ability to fulfill this contracting and governance role. The next section of this paper reviews the existing smoothing and conservatism literatures while placing our hypotheses in this context. The third section describes our intersegment smoothing variable in more detail. The fourth section describes our sample selection process and presents descriptive statistics. The fifth section empirically documents results related to our hypothesis that the inter-segment smoothing variable is associated with information asymmetry. The sixth section presents empirical results related to our main contribution that inter-segment smoothing reduces accounting conservatism. The seventh section presents additional findings that certain non-recurring negative accruals used to communicate bad news are less likely when inter-segment smoothing is high. The final section summarizes the paper. II. Literature Review and Hypothesis Development Income smoothing has been the subject of a number of accounting research studies, both analytical and empirical, over the past three decades. A comprehensive review of that literature is beyond the scope of this study, but others have periodically reviewed this research (see Ronen and Saden [1981]; Schipper [1989]; and Dechow and Skinner [2000], for examples). The traditional income smoothing study examines the inter-temporal recognition of firm-level profit figures. These models or empirical specifications seek to identify situations where revenues or 4

7 expenses are shifted by the firm from one period to another in an attempt to present a time-series of earnings with a lower variance. Yet this definition is problematic because GAAP, when neutrally applied, requires firms to smooth cash flows via the accrual process. Dechow and Skinner [2000] discuss the problems a researcher faces when trying to determine whether there is too much smoothing and thus earnings management. With any earnings management study, the challenge for the researcher is in identifying what the unmanaged portion of earnings is before testing whether the managed piece corresponds to the economic forces of their hypothesis. In the smoothing literature, for example, DeFond and Park [1997] study the descriptive validity of Fudenberg and Tirole s [1995] theory that managers smooth income across time to prevent the intervention of the principal/owner in the manager s division (e.g., by sacking him). DeFond and Park interpret their empirical evidence as consistent with the theory. But in a follow up paper, Elgers, Pfeiffer and Porter [2003] point to methodological problems in measuring expected accruals via the modified-jones model, and how this biases the DeFond and Park conclusions in favor of the smoothing hypothesis being tested. In our study, we take a different tack from much of the prior literature and focus on intersegment income smoothing by the firm across segments as opposed to across time. The smoothing of income across segments can take two forms. The first approach is to aggregate different business activities into segments such that the variance of the underlying profits across units is reduced. This is the form of smoothing described analytically by Hayes and Lundholm [1996]. In their setting, firms choose to aggregate business segments with dissimilar profitability so as to hide information from competitors. The second approach to smoothing income involves the manipulation of the cost allocation of expenses that are components of segment profits. 5

8 Hann and Lu [2008] find that the incidence of segment losses appears to be unusually low, and they attribute this result to strategic cost allocation decisions by managers. With the exception of Hann and Lu [2008], most past research that examines management s distortion of segment information focuses on the aggregation of dissimilar segments into a common segment(s). The early literature in this area focused on the proprietary cost incentive as the source of demand for this behavior. For example, Harris [1998] finds that firms with segments in highly concentrated industries tend to aggregate those segments with others in external reports; she also finds that segments in industries where abnormal profits tend to persist are aggregated by firms. She concludes that these managers engage in this behavior to hide information from competitors, as predicted by the Hayes and Lundholm [1996] model. More recent work in this area focuses on agency cost incentives to withhold information on disparate segments. Berger and Hann [2007] examine the revelation of previously hidden segments that was mandated by a change in accounting rules effective in They conclude that the newly revealed segments were previously withheld for agency cost reasons; managers were attempting to prevent shareholders and other stakeholders from pressuring the firm to divest unprofitable segments by hiding the information about these segments In this paper, we examine the smoothing of internal profits across segments via either aggregation or cost allocation. 3 Our measure, which will be more fully described in the next section, captures the variance of profits reported across a firm s multiple segments and compares this to a benchmark of the variance of profits reported across a sample of single segment firms 2 The Financial Accounting Standards Board (FASB) issued Statement of Financial Accounting Standard (SFAS) No. 131, Disclosures about Segments of an Enterprise and Related Information effective for fiscal years beginning after December 15, It is beyond the scope of this project to determine exactly how the profits are smoothed across the segments. However, we present evidence in Section VI that the autocorrelation in our smoothing measures is This suggests that the smoothing measure we use is fairly consistent through time, and thus likely due to business aggregation or cost allocation decisions that cannot be easily changed from year to year. 6

9 matched by size and industry to the segments. Our interpretation is that when the benchmark matched sample has a high variance, but the treatment firm has a low variance, then the treatment firm is engaging in income smoothing across its segments. The advantage to capturing smoothing this way, as opposed to the traditional inter-temporal smoothing measures, is that it alleviates our need to specify a level of expected earnings (or accruals) in a time-series for a sample firm. This is the methodological problem that DeFond and Park [1997] face per the analysis of Elgers et al. [2003]. In a sense, our smoothing measure is akin to Beaver s [1968] research design of examining the information content of earnings not by modeling the first moment expectations of the stock market, but rather the historical pattern in the second moment of variance in stock returns, and how that variance changes around earnings announcements. Our goal in this paper is to examine how inter-segment profit smoothing affects the ability of the overall firm earnings figure to communicate timely information. Our first hypothesis is that inter-segment smoothing suppresses the information in the components of firm earnings because investors are now less aware of the underlying source of firm profits. We predict that inter-segment smoothing will thus be positively associated with other measures of information asymmetry that are publicly observable, namely, bid-ask spreads and the probability of informed (PIN) trading as defined by Easley, Kiefer and O Hara [1997]. Conditional on inter-segment smoothing increasing information asymmetry, we focus on a specific manifestation of how accounting information may be garbled with this smoothing. We focus on the conservatism inherent in GAAP, and how inter-segment smoothing might affect the application of this principle. Basu [1997] defines conservatism as accountants tendency to require a higher degree of verification for recognizing good news than bad news in financial statements (p. 4). This 7

10 definition leads to the natural prediction that bad news will be reported on a more timely basis than good news, since the accountant will require less corroborating evidence to recognize a loss than a gain. Basu empirically documents the application of this principle in the U.S. over a period of roughly four decades ( ); he also cites sources that claim it had been applied in Europe over the past six centuries. Ball [2001] and Watts [2003] explain the demand for conservatism as coming from the contracting process where, largely due to governance needs and lending arrangements, conservatism arises endogenously. An exogenous force that has been studied as a source of conservatism is the legal nature of a corporation s national environment. Ball, Kothari and Robin [2000] find that firms in common-law countries, where accounting standards are more likely to arise from the bottomup via commercial transactions, contracts and court precedent, are likely to exhibit more conservative earnings numbers than those in code-law countries, where accounting standards come from the top-down via government action. Bushman and Piotroski [2006] examine additional macro factors in a cross-country study, including legal origin, securities laws and other political economy measures. Less research has examined conservatism at the firm level within the U.S. to determine forces that might lead to cross-sectional variation in its application. Ryan [2006] identifies many of the problems associated with applying the Basu framework generally (see especially pp ), as does Givoly, Hayn and Natarajan [2006]. Khan and Watts [2007] suggest that three firm fundamentals are natural determinants of a firm s conservatism measure: the market-tobook ratio, firm size, and firm leverage. They use these variables to form a C_Score that they validate with tests that are direct (e.g., higher C_Score firms exhibit more asymmetric timeliness as defined by Basu) and indirect (e.g., higher C_Score firms tend to experience underlying 8

11 fundamentals that lead to a higher demand for conservatism, such as greater information asymmetry). Our focus is whether a discretionary reporting decision, the smoothing of profits across the segments of a firm, impairs the ability of the accounting system to fulfill an important objective, the communication of bad news in a relatively timely manner. Our maintained assumption is that bad news events, such as revenue or cost shocks that reduce profits, increase the underlying volatility of the income stream. Yet a reporting system that smoothes these shocks across the business units of a firm will obscure theses shocks and cause firm level accounting numbers to be less responsive to bad economic news. Thus, our second hypothesis predicts that firm-years classified as higher smoothers will exhibit less conservatism. That is, the incremental bad news coefficient in the Basu regression will be attenuated in firm-years that are classified as having high levels of smoothing of their internal profits. One of the ways in which bad news is recognized in accounting earnings is through the use of negative, non-recurring accruals. Compustat uses the non-gaap term special items to describe these accruals. We believe that either asset write-downs or restructuring charges are the accounting technology frequently used to actually record the bad news. We refer to these items as negative non-recurring accruals. Our third hypothesis is that these negative non-recurring accruals will be recorded with a lower frequency for smoothing firms. Although this prediction may seem trivial in light of our second hypothesis, the role of these accruals in the application of conservatism is unclear. Callen, Hope and Segal [2009] conclude that special items per Compustat are a noisy manifestation of asymmetric timeliness. Further, Frankel and Roychowdhury [2007] find that restructuring charges appear to be particularly abused. Specifically, Frankel and Roychowdhury 9

12 [2007] find that restructuring charges are frequently more persistent (i.e., not as non-recurring as they are purported to be) for firms that score low on the Khan and Watts [2007] C_Score. As such, these charges may reflect big bath or cookie jar reserve behavior more than they do the asymmetric timeliness feature of earnings. Accordingly, there is tension regarding whether we can observe an effect of inter-segment smoothing on the likelihood of negative non-recurring accrual recognition. III. Measurement of Inter-Segment Smoothing Our main hypothesis variable in this study is a within sample decile ranking of a ratio of variances, denoted RVR (for ranked variance ratio). In the denominator of this ratio is the variance of profits across a firm s segments. 4 In the numerator is the variance of profits across a sample of single segment firms that are matched by size and industry to the multi-segment firm s business segments. The higher the ratio, the more smoothing that can be inferred. We assume that the variance in profit across the matched single segment firms represents the expected level of variation in profits within a multi-segment firm. As RVR increases above one, the more the expected performance variance exceeds the reported performance variance. We interpret this as inter-segment profit smoothing. We use ranks as opposed to a continuous measure of the ratio to control for extreme values that can result when measuring a variance across a small number of data points. Both profit figures in the raw variance ratio are deflated by the respective sales of the unit of analysis (i.e., segment sales for the multi-segment firms and firm sales for the single segment group). We match on industry at the most precise level possible using SIC classification, limiting ourselves to the two digit level. That is, where possible, we match first on the four-digit 4 Only multi-segment firms are included in our sample. We elaborate on the sample selection in Section IV. 10

13 level, second on the three-digit level if a four-digit match is not possible, and third on the twodigit level if no match occurs in the first two attempts. After matching at the industry level, we then choose the single segment firm with the value of sales closest to the business segment s. Because the definition of profit at the segment level may vary across the sample firms, we use an algorithm to infer how the firms are measuring this figure at the segment. This is necessary because we wish to remain consistent in our definition of profit with our benchmark single-segment firms. We sum the profit figures across the segments, and then compare it to several likely firm level earnings figures. We identify the minimum difference between the sum of segment profit and the firm level measures. We then use that firm level measure for our single segment firms when calculating the variance of profit for this matched portfolio. Our candidates for firm level earnings in this process include (all identifying numbers refer to Compustat fields): net income before extraordinary items (#18); net income before extraordinary items plus depreciation expense (#18 + #14); net income before extraordinary items plus interest expense (#18 + #15); net income before extraordinary items plus tax expense (#18 + #16); net income before extraordinary items plus depreciation expense plus interest expense (#18 + #14 + #15); net income before extraordinary items plus depreciation expense plus tax expense (#18 + #14 + #16); net income before extraordinary items plus interest expense plus tax expense (#18 + #15 + #16); 11

14 and, net income before extraordinary items plus depreciation expense plus tax expense plus interest expense (#18 + #14 + #16 + #15). Our measure of smoothing is similar to one used by Ettredge, Kwon, Smith and Stone [2006], however there are significant differences. Ettredge et al. use the range of return on sales (ROS) across a firm s segments as the measure of variability. 5 This serves as a dependent variable in their analyses. The primary research question they ask is whether this variability increased following SFAS 131. They construct a benchmark inherent variability measure by looking at the range of mean profits of the single segment firms operating in the same industries as the multi-segment firms various segments. This inherent variability is then used as a control in an OLS regression where the aforementioned range of ROS across firm segments serves as the dependent variable. While our ranked variance ratio (RVR) measure differs significantly from the Ettredge et al. approach, the variables are similar. But more importantly, the research question we ask is considerably different from this other study. Ettredge et al. are concerned with how a new accounting standard, SFAS 131, affected the inter-segment smoothing behavior of multi-segment firms. Our research question is whether the inter-segment smoothing behavior of multi-segment firms can affect the asymmetric timeliness of earnings. Thus our contribution is to assess whether the accounting system s role of communicating bad news on a timely basis is compromised via inter-segment profit smoothing; this is a fundamentally different research question than Ettredge et al. address. 6 5 In footnote 11 (p. 98) they report that their results are robust to using the standard deviation of ROS rather than the range. 6 We do explore the role of SFAS 131 in our tests in Section VI. 12

15 IV. Sample Selection, Variable Definitions and Descriptive Statistics Firm-level data is gathered from the Compustat Annual Industrial Files. Firms must have data on sales (#12), diluted EPS including extraordinary items (#169), net income before extraordinary items (#18), the current and lagged fiscal year-end stock price (#199), the current and lagged fiscal year-end shares outstanding (#25), the book value of equity (#60), the book value of long-term debt including the current portion (#9 + #34), and assets (#6) to be included in our sample. Firms must also have a continuous string of stock returns on the Center for Research in Security Pricing (CRSP) database from April of the sample year to March after the sample year-end. We gather segment information from the Compustat Segment database requiring segment sales (SALE), an income variable (OPS, OIBD, or NI), and an SIC code (SSIC1 or SSIBC1). Segments are deleted if segment sales are equal to or less than zero. Compustat includes singlesegment firms as segments in the segment database, therefore segments are also deleted if segment sales are equal to or greater than aggregate firm sales. We identify multi-segment firms as those firms that are successfully matched to at-least two segments on SPC Permanent Number and year-end date. Multi-segment firms are deleted from the sample if the sum of segment sales for segments successfully matched to the multisegment firm is not within 1% of aggregate firm sales. We identify single-segment firms as those firms that are successfully matched to zero or one segments on SPC Permanent Number and year-end date. We do this matching before any segments are deleted from the segment file to reduce the likelihood that multi-segment firms are mistakenly identified as single-segment firms. 13

16 Table 1 presents two panels of descriptive statistics. Panel A includes all of the observations (17,367) that will be included in our main Basu regression tests. That panel presents the descriptive statistics regarding the distributional characteristics of the variables that will be used in the regressions. This sample is truncated by deleting observations in the top and bottom 1% of deflated earnings per share (DEPS), annual returns (R), book-to-market (BTM), and the top 1% of leverage (LEV). The variables are defined as follows: DEPS: Diluted earnings per share including extraordinary items (#169) divided by the stock price at the beginning of the period (#199). RVR: A ranking of the firm s variance ratio into deciles, higher deciles indicate higher levels of segment performance variance suppression. Neg: A dummy variable equal to one if returns for the 12 month period beginning in April of the sample year are less than zero; else the dummy variable takes on a value of zero. SIZE: The natural log of total firm assets (#6) BTM: The book value of equity (#60) divided by the market value of equity (#199 * #25) LEV: The book value of long-term debt including the current portion (#9 + #34) divided by the market value of equity (#199 * #25) Panel B of Table 1 presents the univariate correlations for all of the variables above. Pearson correlations are presented above the diagonal, with Spearman below. 14

17 V. Correlations between Inter-Segment Profit Smoothing and Information Asymmetry Measures In our first set of analyses we examine the univariate correlations between RVR and other publicly observed measures of information asymmetry. Our objective is to gain comfort that the smoothing RVR measure does indeed measure a phenomenon that is associated with information asymmetry. We focus on univariate correlations between RVR and both the bid-ask spread (BA_S) and the probability of informed trading (PIN). The bid-ask spread is a common measure of information asymmetry in the equity market. Moreover, it is one that more transparent disclosure or accounting policies are thought to reduce. For example, Healy, Hutton and Palepu [1999] find that sustained increases to voluntary disclosure (as measured by a qualitative rating of disclosure quality by sell-side equity analysts) are associated with a decline in the spread. Leuz and Verrecchia [2000] document that German companies that voluntarily commit to a more transparent accounting system (as measured by a switch from German GAAP to the more transparent IFRS or U.S. GAAP in the mid-1990s) experience a decline in the spread. The PIN measure, as developed by Easley, Kiefer and O Hara [1997] is another variable used to capture information asymmetry. It is an estimate of the probability that a trade in an equity security has been initiated by an informed trader. Hence, it is increasing in the information asymmetry between informed and uninformed investors. Like Healy et al. s [1999] analysis of disclosure and bid-ask spreads, Brown and Hillegeist [2007] find that voluntary disclosure quality is also inversely related to the PIN, the implication being that a firm s reporting strategy can reduce this information asymmetry measure that raises the cost of capital for a firm. 15

18 We predict with our first hypothesis that the smoothing measure RVR is positively associated with both BA_S and PIN. Our logic is that when firms smooth profits across their segments, this increases the level of information asymmetry as captured by these proxies. This prediction is similar to the conjectures of Healy et al. [1999] and Brown and Hillegeist [2007]. BA_S is calculated as the average closing bid-ask spread per CRSP throughout the year that RVR is measured. We use calendar year PIN scores calculated by Assistant Professor Soeren Hvidkjaer for the years 1983 to 2001, posted on his webpage; to proxy for information asymmetry. The PIN score measures abnormal trading activity by inferring imbalances in the volume of buy versus sell orders in a given stock. The PIN score employs the assumption that the market maker should incorporate new public information directly into prices for subsequent trades of a given stock, thus creating no systematic order imbalance. In contrast, the market maker is not capable of incorporating private information into stock prices for subsequent trades. Thus, an order imbalance will exist whereby bad private information generates an excess of sell-orders and good private information generates an excess of buy orders. Firms with relatively greater abnormal trading volume (a higher PIN score) likely have more information asymmetry between outside investors and insiders. Easley & O Hara (1992), Easley, Hvidkjaer, & O Hara (2002 & 2004) provide an extensive treatment of the PIN score and its relation to information asymmetry. We limit our analysis for our first hypothesis to univariate correlations because the ultimate goal of this analysis is to gain comfort that RVR actually captures a measure of information asymmetry. This provides us with a cleaner interpretation of our primary analyses in the next section where we examine whether RVR inhibits the asymmetric timeliness of earnings. 16

19 Panels A and B of Table 2 present all of the regression variables defined in Section IV for smaller sub-samples where the BA_S and PIN are available. We have only 11,785 observations with a usable bid-ask spread, and 3,801 with a PIN. The use of the PIN score imposes several sample size limitations on our study, which account for the dramatic drop in the number of viable observations. The PIN score is calculated on a calendar year basis only, so firms that don t have a December fiscal year-end are lost. Further, the PIN score is only available for firms on the NYSE/AMEX stock exchange for the years 1983 through In Panel C of Table 2 we present the correlations between our smoothing measure RVR, and BA_S and PIN. Pearson correlations are presented in the upper diagonal, Spearman in the lower. The correlations across all variables are positive and significant, as predicted. The magnitudes of the correlations between PIN and BA_S are quite high, which is to not surprising as they are both designed to capture the primitive information asymmetry construct. Our variable of interest, RVR, is positively correlated with each of the other measures, but the magnitudes of the correlations are modest. This is to be expected, however, as the primitive measures should encompass all of the features that lead to information asymmetry (e.g., disclosure policy, organizational structure, other accounting choices, etc.) whereas RVR captures a single aspect. VI. Methodology and Results of Testing Main Hypothesis The basic Basu regression uses earnings as the dependent variable, with contemporaneous stock returns as an explanatory variable that proxies for economic news. Returns are interacted with a dummy variable that takes on a value of one if the return is 17

20 negative, zero otherwise. In equation form, the model is presented below (all variables are defined in Section IV): DEPS it = β 0 + β 1 Neg it + β 2 R it + β 3 Neg it * R it [1] The evidence of conservatism, or asymmetric timeliness, is provided by a positive coefficient β 3. We augment the basic Basu model by including our proxy for inter-segment profit smoothing RVR as an additional interactive variable per model [2] below: DEPS it = β 0 + β 1 Neg it + β 2 R it + β 3 Neg it * R it + β 4 RVR it + β 5 Neg it * RVR it + β 6 RVR it * R it + β 7 Neg it * RVR it * R it [2] Our prediction for our second hypothesis is that the β 7 coefficient is negative. This would suggest that inter-segment smoothing reduces the ability of the accounting system to communicate bad news in a timely manner. Table 3 presents estimates of several variants of model [2]. In the first column of results, the model above is estimated for the entire pool of 17,367 observations. The results of this estimation support our second hypothesis. The β 7 coefficient is negative and statistically significant at the 1% level. The magnitude of the coefficient is Thus, if a firm moved from the lowest decile to the highest decile of smoothing, then that would suggest the Basu asymmetric timeliness coefficient would fall by (or nine multiplied by ); this reflects a decline of 24.0% of the baseline estimated β 3 coefficient of Even a more conservative change of a one standard deviation increase in RVR of would suggest a decline in the asymmetric timeliness coefficient of (or multiplied by ); a decline of 7.6% in the magnitude of the coefficient. As an extension of model [2], we use the lagged value of RVR. We expect that smoothing should represent a fairly static decision by the firm, especially if it is driven by the 18

21 choice in how business activities are aggregated into reportable segments; we do not expect this decision to change much from year to year. Indeed, when we estimate a univariate regression where current RVR is the dependent variable and lagged RVR the explanatory variable, the adjusted R 2 of the model is and the estimated coefficient for the lagged value is 0.394, which is highly statistically significant. Thus, inter-segment profit smoothing is a persistent phenomenon. In the third column of Table 3 we replace the contemporaneous RVR with its lagged values. The statistical significance and economic magnitude of the coefficient do not change in this specification. In the second and fourth columns of Table 3 we estimate model [2] annually, using either the contemporaneous (second column) or lagged (fourth column) RVR as our main variable of interest. Then, we take the average β 7 coefficient across the 24 or 23 years and test for its statistical significance using the Fama-MacBeth method. This approach controls for crosssectional dependence in the financial data that likely exists. The results presented in columns two and four of Table 3 suggest that the β 7 coefficient is statistically insignificant when using the Fama-MacBeth approach. To examine this result further, we examined the time series of the β 7 coefficient from the annual model [2] estimations using the contemporaneous RVR value interacted with R and Neg. We noticed that it was after the application of SFAS 131 in 1998 that the stability of the β 7 coefficient began to deteriorate. Recall that SFAS 131 was designed to improve segment footnotes by aligning the external reporting with internal management tracking of business units. Ettredge et al. [2006] find that following the mandatory adoption of this standard the range of reported profits across business segments increases, on average; they conclude that this implies that SFAS 131 reduced the smoothing by management of profits across their segments. In our 19

22 sample, there appears to be some evidence that the higher pre-sfas 131 smoothing levels were accompanied by a greater reduction in the asymmetric timeliness of earnings. Specifically, the average value of β 7 over the 15 years from was ; the Fama-MacBeth t-statistic for this average coefficient is Moreover, the estimated coefficient is negative in 11 of the 15 years. Yet in the 9 years from , the average β 7 value is ; the Fama- MacBeth t-statistic for this is and the estimated coefficient is negative in 4 years. 7 We examine the effects of this accounting regime change further in our next set of analyses that incorporates more controls in the base model. We augment model [2] by adding the variables identified by Khan and Watts [2007] as cross-sectional determinants of asymmetric timeliness. These variables are as follows: SIZE: The natural log of total firm assets (#6) BTM: The book value of equity (#60) divided by the market value of equity (#199 * #25) LEV: The book value of long-term debt including the current portion (#9 + #34) divided by the market value of equity (#199 * #25) Including the asymmetric timeliness measure interactions with these firm fundamentals will insure that our smoothing measure is not driven by these other forces. The resulting model is presented below: DEPS it = β 0 + β 1 Neg it + β 2 R it + β 3 SIZE it * R it + β 4 BTM it *R it + β 5 LEV it * R it + β 6 RVR it * R it + β 7 Neg it * R it + β 8 SIZE it *Neg it * R it + β 9 BTM it *Neg it * R it + β 10 LEV it *Neg it * R it + β 11 RVR it *Neg it * R it [3] 7 One year in particular drives up the average in the post-sfas 131 era: in 2003 the estimated coefficient is Excluding that year the post-sfas 131 coefficient is essentially zero:

23 We present the results of estimating model [3] in Table 4. We specified model [3] with minimal intercept terms to reduce multi-colinearity problems consistent with Khan and Watts [2007]. Our results are quantitatively and quantitatively similar with and without the control variables, although SAS diagnostic output suggests that strong multi-colinearity problems exist when the full set of intercept terms is included. The unconditional asymmetric timeliness coefficient (β 7 ) is positive and significant, as expected. The Khan and Watts conditional variables also take on their expected signs. Our variable of interest is β 11, which conditions the asymmetric timeliness on the degree of inter-segment smoothing. Consistent with our earlier analyses, this estimated coefficient is negative and statistically significant in column one where the pooled regression model is presented. Thus, the results continue to suggest that intersegment smoothing reduces the asymmetric timeliness of earnings. The second column of the table presents the Fama-MacBeth coefficients and t-statistics. In this specification, the β 11 smoothing coefficient is also negative and statistically significant. As previously noted, we detect an attenuated effect of smoothing on the asymmetric timeliness of earnings in the years following the enactment of SFAS 131 ( ). If we try to control for this temporal effect via an additional dummy variable for the change in accounting regime, we end up with a model that suffers collinearity problems per the SAS diagnostic output. This is not surprising given the various interactions involved with such a model. That is, we have dummy variables for both time and negative returns, as well as our smoothing measure RVR. With all combinations of these variables interacted with the economic income proxy (i.e., stock returns, R), the model is difficult to interpret. In lieu of that approach, we present a simpler model in Table 5. In the first column of results, we present model [3] which includes the Khan and Watts determinants of asymmetric 21

24 timeliness, for the pre-sfas 131 period only. In this model, our smoothing interaction, coefficient β 11 is still negative and significant with a magnitude of In the post-sfas 131 era presented in the second column, we see that the magnitude of the coefficient has fallen by roughly 25% to , although it is still statistically significant. While this suggests that SFAS 131 has reduced the ability of inter-segment smoothing to compromise the asymmetric timeliness of earnings, the standard has not eliminated the effect. Overall, the results from Tables 3-5 suggest that inter-segment smoothing reduces the asymmetric timeliness of earnings, consistent with our second hypothesis. Additional analyses suggest that this effect has attenuated post-sfas 131, but is still present. VII. Inter-Segment Smoothing and the Likelihood of Asset Impairments The results above suggest that inter-segment profit smoothing reduces accounting conservatism as measured by Basu s asymmetric timeliness specification. In this section, we analyze a specific method in which conservatism is applied in order to re-enforce this conclusion. We analyze whether smoothing reduces the likelihood of accounting accruals that are used to communicate bad news, namely asset impairments or restructuring charges. If inter-segment smoothing allows firms to avoid undertaking real actions to address fundamental operational problems, then we expect a negative association between the measure RVR and the likelihood of an asset impairment or restructuring charge after conditioning on economic fundamentals. This is the third hypothesis of our paper. We examine only the time period as this is when Compustat fields for specific impairment and restructuring charges are more complete. Compustat has presented the annual data field Special Items (#17) as a record of impairment and restructuring charges since

25 However, that field contains a myriad of other charges that are not impairments or restructurings. After reading detailed earnings announcements, Bens and Johnston [2008] screen out a number of special item charges that are unrelated to restructuring charges or asset impairments for their study of earnings management via these accruals (see their Table 1). Since 2001 Compustat has been recording specific charges that are either tangible asset impairments (#380), intangible asset impairments (#368) or restructuring charges (#376). 8 Frankel and Roychowdhury [2007] find that restructuring charges may reflect opportunistic behavior more than they do the asymmetric timeliness feature of earnings. Therefore, our primary test focuses on asset impairments alone. We code our dependent variable D_Spec as one if the firm has recorded an impairment charge to either tangible or intangible assets. In untabulated results we examine a dependent variable taking on a value of one if a charge is recorded in any of the three fields for the year, zero otherwise. We base most of our control variables on the study by Francis, Hanna and Vincent [1996], which identified several firm and industry level economic determinants of asset writeoffs and restructuring charges. We also include the change in GDP as a macro variable that is correlated with restructuring charges (Bens and Johnston [2008]). Finally, we include the three determinants of asymmetric timeliness identified by Khan and Watts: size, book-to-market, and leverage. Our variable of interest remains our ranked inter-segment profit smoothing variable, RVR, which we predict will be negatively associated with the likelihood of recording a write-off. We present our probit model below, and the specific definitions of the variables follow: 8 While some of these fields appear with non-zero values in 2000, the Compustat dataset appears to be incomplete for this year. Lee [2008] reaches the same conclusion and ignores this year (see footnote 15). 23

26 D_Spec it = ß 0 + ß 1 R it + ß 2 Sale it + ß 3 IND_Sale it + ß 4 ( Sale) it + ß 5 ( IND_Sale) it + ß 6 ROA it + ß 7 IND_ROA it + ß 8 CH_ROA it + ß 9 CH_IND_ROA it + ß 10 GDP it + ß 11 RVR it + ß 12 BTM it + ß 13 LVE it +ß 14 SIZE it [4] Consistent with Francis, Hanna and Vincent [1996] we include a variety of variables intended to control for the economic characteristics of the firm that likely impact the probability of an asset impairment. We remove special items (of which our dependent variable is a component) when determining firm financial performance so as to remove the possibility that the explanatory power of our variable of interest, RVR, is unduly sapped. These variables are defined as follows: ROA: Return on assets is calculated as (NIBE Special Items)/Lagged Total Assets or (#18 #17)/#6. IND_ROA: ROA: The average return on assets for the firm s 2-digit SIC code industry. The firm s change in return on assets from the previous year. IND_ROA: The average change in return on assets from the previous year for the firm s 2- digit SIC code industry. Sales: GDP: The firm s change in sales (#12) from the previous year. The year s change in chained GDP from the Federal Reserve Bank of St. Louis web-site: IND_Sales: The average change in sales from the previous year for the firm s 2-digit SIC code industry. ( Sales): The change in the firm s change in sales from the previous year. ( IND_Sales): The average change in the industry s change in sales from the previous year for the firm s 2-digit SIC code industry. 24

27 We present the results of estimating model [4] in Table 6. The first column of results used independent variables that are raw values, with the exception of RVR which is a ranked variable by definition. In the second column, all explanatory variables are decile ranked, except for GDP which is ranked from one to six, corresponding to the number of years in the panel. Firm specific data is ranked within 2-digit SIC code, whereas industry level data is ranked across industries. In both columns, we see a negative and significant [ß 11 ] association between the likelihood of impairments and inter-segment profit smoothing by multi-segment companies. The marginal effect from the probit model evaluated at sample means is Thus if a firm moved from lowest decile to the highest decile of smoothing, then that would suggest the probability of recording an asset impairment is 6.05% lower (or nine multiplied by.0067). Moving from one standard deviation below the average RVR to one standard deviation above decreases the probability of an asset impairment by 3.85%. In untabulated results we find that when the dependent variable D_Spec is formed using both asset impairments and restructuring charges the coefficient on our variable of interest, RVR, becomes insignificant but still negative. This is consistent with evidence in Frankel and Roychowdhury [2007] indicating that restructuring charges may reflect opportunistic behavior more than they do the asymmetric timeliness feature of earnings. This analysis provides complimentary evidence that inter-segment profit smoothing reduces the conservatism applied in accounting reports. Our analysis in Section VI demonstrates that the asymmetric timeliness coefficient declines as smoothing increases suggesting that recognition of bad economic news is delayed when smoothing is high. In this section we 25