Evidence of conditional conservatism: fact or artifact? Panos N. Patatoukas Yale University

Transcription

1 Evidence of conditional conservatism: fact or artifact? Panos N. Patatoukas Yale University Jacob Thomas Yale University Current Version: October 5, 2009 We received many helpful comments and suggestions from Sudipta Basu, two anonymous referees and Sanjay Kallapur (Editor). Dick Dietrich provided invaluable assistance with specification and statistical issues. We also thank Anwer Ahmed, Jeff Callen, John Core, Dan Givoly, S.P. Kothari, Wayne Landsman, K. Ramesh, Bill Rees, Stephen Ryan, Senyo Tse, Jenny Tucker, Martin Walker, Paul Zarowin, and seminar participants at the AAA Annual meetings, AAA FARS Doctoral Consortium, Florida, George Washington, Hong Kong UST, and Yale for their comments.

2 Evidence of conditional conservatism: fact or artifact? Abstract The differential timeliness measure proposed in Basu (1997), which estimates the fraction of observed bad news reported in contemporaneous earnings minus the corresponding fraction for good news, has been used widely to study conditional accounting conservatism. Timeliness is measured as the slope from a regression of earnings, scaled by lagged price, on returns. We find that differential timeliness estimates are biased by two empirical regularities related to lagged price, the deflator in the timeliness regressions: it is negatively related to a) the variance of returns, and b) the probability of a loss, and the magnitude of price-deflated losses. Even though these regularities are unrelated to conditional conservatism, their effects are substantial and pervasive. Also, prior findings regarding time-series and cross-sectional variation in differential timeliness are confounded by corresponding variation in these regularities. Keywords: conservatism; asymmetric timeliness; losses; scaling by share price.

3 Evidence of conditional conservatism: fact or artifact? 1. Introduction Basu (1997) has had a substantial impact on accounting research. Among other innovations, it offers an intuitive and convenient way to measure ex post conservatism, conditional on news observed. The measure captures differential timeliness, or the difference between the fraction of bad and good news reported in contemporaneous earnings. Timeliness is measured as the slope from regressions of earnings, scaled by lagged price, on returns. A large body of subsequent research has used this measure as the dependent or independent variable in analyses, and many papers investigate determinants of cross-sectional and time-series variation in the Basu measure. 1 We show that bias associated with the Basu measure is so pervasive and substantial that the measure is rendered unreliable. More generally, our results illustrate how bias can arise from completely unexpected sources; in this case it is due to two empirical regularities: lagged price is negatively related to a) return variance and b) the probability and magnitude of price-deflated losses. Importantly, cross-sectional and time-series variation in the Basu measure is affected substantially by corresponding variation in the two regularities. Although the Basu measure has considerable conceptual appeal, concerns have been raised before about measurement error as well as potential biases. Givoly et al. (2007) identifies a variety of conditions under which the Basu measure is likely to contain considerable measurement error. More relevant to this study, Dietrich et al. (2007) identifies conditions under which the Basu measure is biased because returns, which is the endogenous variable that describes market reaction to earnings and non-earnings information, is used as the regressor and also to partition firm-years into good and bad news subsamples. 1 Examples of studies investigating time-series variation include Basu, (1997), Givoly and Hayn (2000), Holthausen and Watts (2001), Ryan and Zarowin (2003), and Sivakumar and Waymire (2003) whereas those investigating cross-sectional variation include Giner and Rees (2001), Pae, Thorton and Welker (2005), Roychowdhury and Watts (2007), Ball et al. (2009), Easton et al. (2009), and Beaver and Ryan (2009).

4 Ball et al. (2009) seek to alleviate concerns raised in the literature. The Ball et al. (2009) counterarguments appear to have been persuasive since many studies continue to use the Basu measure. We hope our results will convince researchers to consider alternative measures of conditional conservatism. Whereas there could be disagreement about the conditions required for the Dietrich et al. (2007) biases, (e.g., returns are determined substantially by reported earnings), the conditions required for our bias are easily verified. As described in Section 4.4, despite appearances of similarity between the two sets of conditions, the two biases are unrelated; i.e., doubts about the validity of the Dietrich et al (2007) concerns are not relevant here. 2 To anchor the discussion, we review the Basu (1997) regression model. Regression models are one of many alternative methods proposed in Basu (1997) and the regression we describe below in equation (1) is one of many versions considered in Basu (1997) and the related literature. Equation (1) describes a regression of annual earnings levels (scaled by lagged price) on annual returns, estimated separately for good and bad news partitions, represented by positive and negative returns, respectively. The coefficient β 1, which is the incremental slope for bad news over that for good news, is the Basu measure of conditional conservatism. X P it it 1 = α + α D + β R + β R * D + ε 0 1 it 0 it 1 it it it (1) where X it = earnings per share reported by firm i in year t, P it-1 = price per share for firm i at the end of year t-1, R it = stock return for firm i in year t and D it = 1 if R it < 0, which represents bad news, and =0 otherwise. The empirical regularities and associated bias are illustrated below. The first empirical regularity, labeled the return variance effect, describes the negative relation between share price 2 Beaver and Ryan (2009) document biases in the Basu measure that increase with the value of the equityholders option to put the firm to debtholders. This source of bias is also unrelated to the bias documented in this paper. 2

5 and return variance. That is, a scatter plot of the inverse of lagged price (1/P it-1 ) on returns (R it ) resembles an inverted cone (see Panel A of Figure 1). High price firms with low values of 1/P t-1 have positive and negative returns of smaller magnitudes that are clustered close to the y-axis. The magnitudes of returns on either side of the y-axis increase on average for larger values of 1/P it-1. Consequently, regressions of 1/P it-1 on returns estimated separately for negative and positive returns representing bad and good news in the left and right halves of the inverted cone will result in negative (line AC) and positive (line BD) slopes, respectively. The next regularity, labeled the loss effect, is that the probability of reporting a loss and the magnitude of that loss (scaled by lagged price) are both negatively related to share price. That is, the frequency and magnitude of price-deflated losses will be greater for firms with lower share prices near the top of the inverted cone in Panel A of Figure 1 (indicated by the horizontal slice that includes A and B), relative to firms with higher share prices near the bottom of the inverted cone (indicated by the horizontal slice that includes C and D). Note that the loss effect is a general effect, not conditional on this period s news, and it describes losses in any year, not just the current year. That is, the loss effect is not caused by the current year s loss reflecting bad news which results in lower share prices. Since conditional conservatism predicts that losses reported this year may be in response to bad news disclosed this year, it is important to eliminate the possibility that the loss effect is partially reflecting conditional conservatism. To allay concerns, we illustrate the loss effect using lagged earnings (X it-1 ), rather than current earnings. Last year s earnings must be unrelated to this year s news, since it is reported before this year s news is revealed, and any patterns we document must therefore be unrelated to conditional conservatism. Panel B1 illustrates the joint impact of the return variance and loss effects, by showing regression lines we expect to observe when we replace the inverse of lagged price (1/P it-1 ) in the 3

6 scatter plot in Panel A with price-deflated lagged earnings (X it-1 /P it-1 ). Absent the two effects, we expect lagged earnings to be unrelated to current returns; i.e., the regression lines estimated separately for bad and good news (AC and DB, respectively) will be flat lines corresponding to the mean value of X it-1 /P it-1. However, when the two effects are incorporated, we expect the regression line for bad news (A C ) to be upward sloping and the regression line for good news (D B ) to be downward sloping. This is because firms with extreme positive and negative returns are more likely to be low price firms, relative to firms with less extreme returns, and those firms should exhibit lower mean values of X it-1 /P it-1, because of the loss effect. In essence, the return variance and loss effects convert the symmetric relation for good and bad news described by AC and DB into the asymmetric relation described by A C and B D, where slopes for the bad and good news subsamples are biased upward and downward, respectively. The excess of the slope of A C over that for B D represents spurious indication of differential timeliness. Panel B2 illustrates how a relation between price-deflated current earnings and returns that is symmetric for bad and good news subsamples is affected by return variance and loss effects. Assume that there is no conditional conservatism and the same fraction of news is reported as earnings for both good and bad news. That symmetric relation is described by the regressions lines AC and BD, where the slopes of the regression lines represent timeliness or the fraction of news reported in contemporaneous earnings. Introducing the two effects will cause the regression lines to be biased toward A C and B D. As with lagged earnings, the excess of the slope of A C over that for B D will erroneously be interpreted as evidence of conditional conservatism. These predicted patterns are observed in actual data. We confirm that the return variance and loss effects are substantial. The spurious evidence of differential timeliness predicted for 4

7 lagged earnings (as in Panel B1) is clearly evident in our results, and the magnitude of that differential timeliness is a large fraction of the differential timeliness observed for current earnings (as in Panel B2). We also show that time-series and cross-sectional variation in the Basu measure documented in the prior literature is confounded by corresponding variation in the return variance and loss effects. In our time-series analysis, we show how variation in the two regularities results in similar variation in differential timeliness for both lagged and current earnings regressions. Our cross-sectional analysis focuses on how variation in the two regularities across two attributes that have been studied in the prior literature firm size and book-to-market ratios results in similar variation in differential timeliness for both lagged and current earnings. We consider alternative ways to mitigate biases caused by the return variance and loss effects. For example, we include lagged price as an additional regressor and estimate regressions separately for partitions based on share price (e.g., Easton, Nikolaev, and van Lent, 2009). We also investigate unscaled regressions and scaling by deflators other than share price. Unfortunately, none of those efforts successfully eliminate our bias. We emphasize that our results do not provide estimates of the true underlying level of conditional conservatism, since the asymmetric timeliness we document for lagged earnings is a lower bound for our bias. Also, we have not investigated other potential sources of bias, such as those suggested in Dietrich et al. (2007). Despite the potential for Basu measures to be upward biased, however, we should also emphasize that our results do not suggest that there is no underlying conservatism in the data. 5

8 The remainder of this paper is organized as follows. The sample is described in Section 2, and Section 3 contains results describing the bias due to the two regularities. Section 4 describes additional analyses, and Section 5 concludes. 2. Sample and descriptive statistics We collect our sample from the intersection of Compustat annual files (including the research file) and the CRSP monthly returns file. We scan the Compustat files for firm-years ending during the period 1963 to 2006 that have non-missing values in the current and prior year for earnings per share before extraordinary items (X it ), common shares outstanding, and fiscal year-end stock price (P it,). See the Appendix for further details of variables. We impose a minimum lagged price filter (P it-1 > $1) because the return variance and loss effects are even stronger for very low price firms. Eliminating such firm-year confirms that the results are not driven by a small set of very low price firms. Next, we match these observations to compounded inter-announcement returns, which begin with the fourth month of the fiscal year. We require non-missing returns on the CRSP monthly returns file for the 12 months beginning with the fourth month of the current fiscal year and ending with the third month of the next fiscal year. Compounding these annual returns generates a buy and hold return (R it ) that proxies for the return from holding the stock between last year s earnings announcement and this year s announcement. We compute corresponding market returns based on the CRSP equally weighted market portfolio (including distributions), which are then subtracted from firm returns to generate abnormal returns (AR it ). Basu (1997) considers different specifications for both the dependent variable (e.g., unadjusted and market-adjusted earnings) and the independent variable (e.g., unadjusted and market-adjusted returns, R it and AR it, respectively, computed over different 12-month windows). 6

9 To improve comparability with key results reported in the literature we report results based on unadjusted earnings and market-adjusted returns (AR it ) cumulated over the inter-announcement window. We confirm that our main findings are not affected when we use the other dependent and independent variable specifications. To be consistent with the common practice in this literature, we exclude firm-years falling in the top or bottom 1% of the annual cross-sections of any of the following variables: 1/P it-1, X it /P it-1, X it-1 /P it-1, and AR it. The resulting sample contains 124,562 firm-year observations. Table 1, Panel A, provides some statistics for the pooled distributions of the different variables. Observing a higher median than mean for X it /P it-1 is consistent with the general findings that price-deflated earnings are left-skewed and firms reporting large losses tend to have lower prices (see, for example, Durtschi and Easton, 2005). Observing a higher mean than median for returns (R it ) indicates a right-skewed distribution, also consistent with general findings regarding the distribution of returns. 3 The negative mean and median abnormal returns (AR it ) are also expected, since a) our sample firms are larger on average than the firms included in the equally-weighted CRSP market index and b) large firm returns tend to be lower than small firm returns on average. The mean value of for the bad news dummy (D it ) indicates that 57.2 percent of firm-years are classified as bad news firms in our sample. Similarly, the mean value of for the loss dummy (LD it ) suggests that 17.9 percent of our sample firm-years report losses. Panel B of Table 1 reports the Pearson and Spearman correlations. Two key correlations that underlie the loss effect are as follows. First, the loss dummy (LD it ) is strongly positively related to 1/P it-1 ; i.e., low price firms are more likely to report losses. Second, X it /P it-1 is 3 Right skewness in returns and normally distributed error terms in the traditional linear regression model of returns on earnings would imply that earnings also be right-skewed. The fact that earnings are left-skewed is therefore anomalous, especially given that operating cash flows are also right-skewed. Conditional conservatism is one explanation for this anomaly (Basu, 1995). 7

10 negatively related to 1/P it-1 ; i.e., low price firms are more likely to have lower values of pricedeflated earnings, mainly because the losses they report are larger when deflated by price. Panel C of Table 1 provides some information about the incidence of loss firm-years in our sample (additional details are introduced later). Briefly, of the 17.9 percent of our sample reporting losses, the good news subsample reports losses only about 11 percent of the time, whereas the bad news subsample reports losses about 23 percent of the time. The bottom two rows of Panel C describe loss behavior for news partitions based on unadjusted or raw returns, rather than market-adjusted or abnormal returns. This alternative partition for news increases the fraction of good news firms from 43 to 60 percent. Those results show that good news firms report about 11 percent losses, similar to the news partitions based on abnormal returns. The smaller bad news subsample, however, reports losses almost 28 percent of the time. 3. Results 3.1 Replicating pooled regressions estimated in Basu (1997) Although we emphasize results based on annual cross-sectional regressions, given that coefficient estimates vary substantially over time, we first report results based on pooled regressions to allow comparisons to prior research. Panel A of Table 2 contains the results of estimating the Basu regression for unadjusted and market-adjusted price-deflated earnings levels for the full sample of 124,562 firm-years. The independent variable is market-adjusted abnormal returns (AR it ), which is also used to create the good and bad news partitions. The first row reports the coefficient estimates and associated t-statistics for unadjusted price-deflated earnings (X it /P it-1 ). The coefficient β 1, which is the Basu measure of differential timeliness or conditional conservatism, is It is slightly lower than estimates reported in prior research because of the additional minimum price filter we impose when selecting our sample. 8

11 The second row in Panel A is based on market-adjusted earnings (X it /P it-1 mean X t /P t-1 ) also considered in Basu (1997). Since subtracting the cross-sectional mean each year is roughly akin to subtracting a constant, the main difference between this regression and the earlier one is that the intercept α 0 is smaller by 0.050, which corresponds to the mean level of price-deflated earnings of reported in Panel A of Table 1. The main finding, however, remains unchanged: there is evidence of substantial conservatism, indicated by a β 1 value of The results in Panel B of Table 2 confirm that the choice of market-adjusted returns versus unadjusted returns for the independent variable does not alter the results in a substantive way. Even though switching from abnormal returns to unadjusted returns alters substantially the good and bad news partitions (see Panel C of Table 1), the main coefficient of interest (β 1 ) and the associated t-statistics are similar to results reported in Panel A. Panel C of Table 2 provides the results of our efforts to replicate the results in Basu (1997) using our data sources and software code but the same sample period (from 1963 to 1990) and sample selection procedures (e.g., listed on NYSE or AMEX) as in Basu (1997). The first and third rows contain the results from Panels A and B of Table 1 in Basu (1997), which correspond to regressions of unadjusted earnings on unadjusted returns and market-adjusted earnings on market-adjusted returns, respectively. The second and fourth rows present the results of our efforts to replicate those two specifications. Since we find no major differences between the first and second rows or between the third and fourth rows, we conclude that there are no unintended differences in sample selection or analyses between Basu (1997) and our paper. 4 4 The biggest difference we find is that our sample size is about 12 percent larger. One possible reason for this difference is that we use current exchange membership which may differ from the prevailing membership when the Basu (1997) sample was collected. 9

12 3.2 Could estimates of β 1 be biased by the two effects related to lagged price, the deflator? As described in Figure 1, estimates of β 1 could be biased upward because lagged share price is negatively related to return variance and the probability and magnitude of price-deflated losses. To investigate this thesis, we report in Panel A of Figure 2 the extent to which the variance of abnormal returns, or Var(AR t ), and the fraction of firms reporting losses in the prior year, or LD t-1, increase with the inverse of lagged price (1/P t-1 ). We form deciles of lagged price each year (group together all fiscal years ending in the same calendar year), and then compute the time-series mean of the annual mean values of Var(AR t ) and LD t-1 for each decile. The sharply increasing trend for Var(AR t ) confirms the presence of a substantial return variance effect: returns tend to be more extreme for low price firms. The sharply increasing trend for LD t-1 in Panel A of Figure 2 confirms the first part of the loss effect: low price firms are more likely to report losses. To investigate whether magnitudes of price-deflated losses are larger for low price firms, we report in Panel B the time-series means of annual cross-sectional means for lagged price-deflated earnings (X t-1 /P t-1 ) for each decile of lagged price. We report these means separately for negative (X t-1 <0) and positive (X t-1 0) values of lagged earnings. The results in Panel B confirm the second part of the loss effect: while there is little difference in magnitudes of X t-1 /P t-1 across price deciles for profitable firm-years, low price firms report considerably larger magnitudes of X t-1 /P t-1 for firm-years reporting losses. To link the return variance and loss effects to bias in coefficient estimates from the Basu regression, we split the sample each year into negative and positive values of abnormal return and form deciles within each bad and good news partition. We then compute the mean values of three variables related to lagged prices and plot in Figure 2, Panel C, time-series means of those variables versus the time-series mean abnormal returns for each of the twenty deciles. The first variable we consider is the inverse of lagged price (1/P t-1 ), indicated by the middle, dashed line. 10

13 Consistent with the return variance effect, firms with extreme positive and negative abnormal returns (AR t ) have lower share prices indicated by higher values for 1/P t-1. The first row in Table 3 offers an alternative description of this regularity by reporting the results of replacing the dependent variable in the Basu regression (X it /P it-1 ) with 1/P it-1. Unlike the pooled regressions in Table 2, the Table 3 regressions report the mean values from 44 annual regressions. The reported t-statistics are derived from the time-series distribution of those coefficients. Consistent with the dashed line in Figure 2, Panel B, the good news slope is positive (β 0 =0.073) and the slope for bad news firms is negative (β 0 +β 1 = = ). The second variable described in Panel C of Figure 2 is the fraction of loss firms (LD t -1 ), indicated by the solid line labeled LD t-1. We focus on the fraction of loss firms in the prior year (t-1) rather than the current year (t) to eliminate any possibility that the loss effect is driven in part by firms following conditional conservatism and reporting losses after observing bad news in year t. 5 Consistent with the loss effect, we find in Panel C of Figure 2 that firms with extreme positive and negative abnormal returns this year are more likely to have reported a loss last year. The third variable we consider, SIGN it-1, combines the probability of reporting a loss with the return variance effect; it is a dummy variable set to -1 if a loss is reported last year, and set to +1 otherwise. Panel C of Figure 2 describes how the sign of earnings reported last year scaled by lagged price (SIGN it-1 /P it-1 ) varies across the 20 bad and good news deciles. As predicted in Figure 1, the greater prevalence of losses for extreme good and bad news causes the V-shaped relation observed for 1/P t-1 to be mildly inverted for SIGN t-1 /P t-1 (the dotted line). The second row in Table 3 confirms that analysis by regressing SIGN it-1 /P it-1 on AR it ; the slope for good news is close to zero, but the slope for bad news is slightly positive. 5 Similar plots are obtained when we consider the contemporaneous fraction of losses, rather than the lagged fraction of losses. This is expected by the loss effect, since low price firms are expected to report more losses, both in the current and prior year, unconditional on the news observed this year. 11

14 Panel D of Figure 2 incorporates the second part of the loss effect low price firms report larger magnitudes of price-deflated losses (X it-1 /P it-1 ) to the first part of the loss effect and the return variance effect. We begin first with lagged earnings, to remove any possibility that our results are affected by conditional conservatism. Comparing the dotted line labeled X t-1 /P t-1 in Panel D with the dotted line labeled SIGN t-1 /P t-1 in Panel C shows that incorporating the magnitude of losses increases the asymmetry between the slopes for bad and good news. As projected in Panel B1 of Figure 1, the joint effect of the return variance and loss effects causes a downward bias in the slope for good news and an upward bias in the slope for bad news firms. The third row in Table 3 confirms those biases by regressing (X it-1 /P it-1 ) on AR it. The large, positive estimate of β 1 (=0.116) suggests conditional conservatism, even though lagged earnings could not possibly reflect conditional conservatism based on news that has not yet been released (as of the date the lagged earnings are reported). Note that by measuring returns between earnings announcements, rather than over the fiscal year, current returns are less likely to contain news reported in last year s earnings. This bias in estimates of timeliness for bad and good news is substantial, and it should logically carry over to the Basu regression which is based on current earnings (X it /P it-1 ). As projected in Panel B2 of Figure 1 the bias created by the return variance and loss effects can convert a timeliness relation that is symmetric for bad and good news into an asymmetric one. The actual plot in Panel D of Figure 2 (indicated by the solid line labeled X t /P t-1 ) is consistent with that bias. And the results reported in the bottom row of Table 3 confirm that relationship based on a regression analysis: the slope for good news firm-years is mildly positive (β 0 =0.019) and the slop for bad news firm-years is substantially positive (β 0 +β 1 = =0.204). To provide some structure to the description provided above, we investigate simulated data created to represent the return variance and loss effects. We confirm that while the two 12

15 effects are necessary, neither is sufficient by itself. The return variance by itself actually biases downward the estimate of differential timeliness and the loss effect by itself creates no bias. 6 Our results so far suggest that a large fraction of the conditional conservatism observed for current earnings is due to potential biases caused by the two regularities because we are able to a) document the presence of the regularities, b) provide evidence consistent with projections implied by the regularities, and c) show results for lagged earnings that are similar to those observed for current earnings. Additional evidence consistent with that inference is provided in the next section. Note that the results observed for lagged earnings represent a lower bound on our bias. 4. Additional analyses 4.1 Time-series variation in estimated β 1 and the return variance and loss effects The mean coefficients from annual Basu regressions reported in the bottom row of Table 3 mask considerable time-series variation in coefficient estimates. The top (solid) line in Figure 3, labeled X t /P t-1, shows how the annual differential timeliness estimate (β 1 ) varies from values close to zero early in the sample period, to values above 0.3 around To be sure, it is possible that this variation is merely random variation around the mean value of reported in Table 3. We show next that the variation reported in Figure 3 is related strongly to time-series variation in the return variance and loss effects. 6 As a technical issue, we find that the loss effect is derived from an underlying skewness effect. That is, the second necessary condition for upward bias can be stated as: high price firms exhibit more right-skewness for price-deflated earnings than low price firms or low price firms exhibit more left-skewness than high price firms. Since price-deflated earnings do not in practice exhibit right-skewness and since observed left-skewness is due to the loss effect, we elect to describe the necessary condition as a loss effect rather than a skewness effect. Note that this skewness effect, which suggests that left-skewness increases as share price decreases, is not the same as the earnings skewness described in the literature (e.g., Dietrich et al., 2007) which refers to left skewness in the overall price-deflated earnings distribution. 13

16 The bottom, dotted line in Figure 3 (labeled 1/P t-1 ) describes time-series variation in the return variance effect by reporting the estimated values of β 1 for annual regressions when we replace the dependent variable in the Basu regression (X it /P it-1 ) with the inverse of lagged price (1/P it-1 ). The annual estimates that vary between 0 and -0.3 describe time-series variation around the mean estimate of reported in the first row of Table 3. The main finding is that years with greater return variance effects, indicated by more negative values of β 1 estimates from the 1/P it-1 regressions, have more positive values of β 1 estimates for the X it /P it-1 and X it-1 /P it-1 regressions; i.e., the solid and dotted lines at the top appear to be mirror images of the dotted line at the bottom (the correlation between the β 1 estimates for is 1/P it-1 and X it /P it-1 is -0.59). As with the analyses described in Section 3.2, we consider the impact of time-series variation in the loss effect in two stages. First, we report estimates of β 1 for the annual regressions based on SIGN t-1 /P t-1, which allows us to determine the impact of the fraction of loss firms. Then we report estimates of β 1 for the annual regressions based on X t-1 /P t-1, which allows us to determine the impact of the magnitude of price-deflated losses. As before, we focus on prior year s losses to ensure that our results are unrelated to conditional conservatism. The middle dashed line in Figure 3, labeled SIGN t-1 /P t-1, shows a pattern of time-series variation in estimates of β 1 that resembles the time-series variation exhibited by β 1 from annual Basu regressions (the top solid line in Figure 3); the correlation between the two series of β 1 estimates is The degree of resemblance is low in the early part of the sample period, but increases later in the sample period. The resemblance increases substantially when we also incorporate the magnitude of losses, represented by the middle dotted line in Figure 3, labeled X t- 1/P t-1 ; the correlation between the two series of β 1 estimates increases to 0.81 Overall, the high degree of comovement exhibited by estimates of timeliness (β 1 ) from regressions based on 14

17 current and lagged earnings (X it /P it-1 and X it-1 /P it-1 ), suggests that much of the evidence of timeseries variation in the Basu measure is due to time-series variation in the return variance and loss effects. Additional confirmation of bias due to the loss effect is provided in Figure 4, where we report the time-series of estimates of timeliness from earnings/returns regressions and the corresponding fraction of loss firms separately for the bad and good news subsamples. 7 Timeliness estimates for the good and bad news subsamples are β 0 and β 0 +β 1, respectively, from equation (1). As predicted by the analysis in Figure 1, we find that time-series variation in the loss effect increases timeliness estimates for the bad news subsample but decreases timeliness estimates for the good news sample. The estimates of timeliness for the bad news subsample are shown by the dotted line with hash marks at the top of the figure, labeled β 0 +β 1, and the corresponding fraction of loss firms is shown by the solid line with hash marks, labeled %Loss (bad news). The two lines show a remarkable amount of positive comovement. The correlation coefficient between year-to-year changes in the bad news timeliness estimates and corresponding changes in the fraction of loss firms is That is, a larger fraction of losses in the bad news partition is associated with greater upward bias in the corresponding timeliness estimate. In contrast, we observe a negative comovement between estimates of timeliness for the good news subsample, shown by the dotted line without hash marks at the bottom of Figure 4 (labeled β 0 ), and the corresponding fraction of loss firms, shown by the solid line without hash marks, labeled %Loss (good news). That is, a larger fraction of loss firms in the good news partition is associated with greater downward bias in the corresponding timeliness estimate. The 7 Klein and Marquardt (2006) document a positive relation between β 1 and fraction of loss firms. We show here that the relation is more evident when we a) separately investigate timeliness and fraction of loss firms for good and bad news subsamples, and b) focus on changes in rather than levels of timeliness and fraction of loss firms. 15

18 correlation coefficient between year-to-year changes in the good news timeliness estimates and corresponding changes in the fraction of loss firms is over the sample period, and is particularly strong after The relation between timeliness estimates and fraction of losses for the bad and good news subsamples is quite vivid in the period from 1995 onward since a) year-to-year changes in timeliness estimates and the fraction of loss firms are unusually large, b) the pattern of sharp movements in the fraction of loss firms for bad news firms is exactly the reverse of that for good news firms, and c) these sharp patterns of changes for the fraction of loss firms for bad (good) news partitions are matched by equally sharp positive (negative) comovement in the corresponding timeliness estimates. To be sure, increased conditional conservatism should be reflected in more losses, and observing a positive relation between timeliness and the fraction of loss firms for bad news subsamples could be interpreted as prima facie evidence in support of the Basu measure. 8 Under this alternative explanation, time-series variation in conservatism is explained by time-series variation in factors such as extant accounting rules, how accountants tend to interpret and apply them, and audit characteristics such as auditor liability regimes (e.g., Basu et al., 2001) and whether or not quarterly numbers are audited (e.g., Basu et al, 2002). This variation in conservatism levels is associated with corresponding variation in the fraction of loss firms, because increased conservatism results in more bad news being recognized in contemporaneous earnings. It is also possible that increased conservatism results in a smaller fraction of good news being recognized currently. 8 For example, Givoly and Hayn (2000) posit that left skewness of the earnings distribution is consistent with conditional conservatism (i.e., a larger fraction of unfavorable events is recognized immediately in earnings causing large negative items versus favorable events being recognized in smaller amounts over many years) and the observed increase in left skewness of the earnings distribution over time is consistent with an increase in conditional conservatism. See also Basu (1995) and Ball et al. (2000) for the skewness/conservatism relation. 16

19 The evidence presented so far, however, has many features that are inconsistent with this alternative hypothesis. First, while increased conservatism might decrease the fraction of good news that is recognized currently as earnings, that fraction is unlikely to be zero or negative on average. The lower limit should be a positive fraction since there are some items of good news that would be recognized even under the most conditionally conservative methods. But the estimates of timeliness indicated by β 0 in Figure 4 are close to zero from 1985 onward. While estimates close to zero might be explained by measurement error in coefficient estimates, it is particularly hard to explain the negative estimates observed for years with higher fractions of loss firms in the good news partitions. Second, the level of conditional conservatism exhibits considerable year-to-year variation (especially the subperiod from 1995 onward) indicated by the line labeled X t /P t-1 in Figure 3. Why would the factors that determine conditional conservatism, such as auditor liability, exhibit so much volatility? More important, the pattern of variation in timeliness observed for bad news firms is opposite that for good news firms (see Figure 4). Why would the factors that determine timeliness impact good news and bad news firms in opposite ways? Finally, we reported in a prior version of this paper that annual estimates of timeliness, or β 0 +β 1 and β 0, for regressions based on current earnings are related strongly to corresponding estimates from lagged earnings. 9 That is, the fraction of current period losses for bad (good) news firms is not only positively (negatively) related to estimates of timeliness from regressions based on current earnings, but also positively (negatively) related to estimates from regressions based on lagged earnings. As hard as it is to explain how conditional conservatism can cause bad (good) news today to be reflected to a greater (lower) extent in lagged earnings, it is even harder 9 The positive relation between the estimates of β 1 for current and lagged earnings regressions (X t /P t-1 and X t-1 /P t-1 ) reported in Figure 3 reflect a positive relation between estimates of β 1 +β 0 for current and lagged earnings and a positive relation between estimates of β 0 for current and lagged earnings regressions. 17

20 to understand how time-series variation in current levels of conditional conservatism in bad (good) news partitions, which determine corresponding variation in the fraction of losses observed currently in those partitions, would explain similar (opposite) variation in the extent to which news is reflected in lagged earnings. Overall, we believe that our evidence is strongly inconsistent with the alternative explanation that changes in conditional conservatism cause changes in timeliness and the fraction of loss firms. The evidence is generally consistent with our thesis that time-series variation in the return variance and loss effects cause timeliness estimates of bad (good) news firms to be biased up (down), which in turn bias upward estimates of conditional conservatism (β 1 ). 4.2 Explaining cross-sectional variation in conditional conservatism Our next set of analyses considers the positive correlation documented in the literature between conditional conservatism estimates (β 1 ) and a) lagged book-to-market (B/M) ratios (e.g., Giner and Rees, 2001, Pae et al., 2005) and b) market value of equity (MV) or size (e.g., Giner and Rees, 2001, Givoly et al., 2007). We look for similarities in cross-sectional variation in estimates of β 1 from current and lagged earnings (X it /P it-1 and X it-1 /P it-1 ) regressions. Observing similar patterns confirms our view that both sets of estimates are biased upward by the return variance and loss effects, which in turn raises doubts about inferences based on variation in estimates of conditional conservatism from the Basu regressions. For example, prior evidence on variation in conditional conservatism across B/M ratios has created some disagreement in the literature. 10 To the extent that the evidence is due to effects unrelated to conditional conservatism, there may not in fact be a controversy. 10 Assuming that B/M ratios are negatively related to unconditional conservatism (e.g., Beaver and Ryan, 2000), does observing a positive relation between B/M ratios and conditional conservatism creates a puzzle? Some suggest that the positive relation observed between β 1 and B/M is due to measurement error in both earnings and returns (e.g., Givoly et al, 2007) and that error should decline as the measurement horizon increases (e.g., 18

21 Table 4, Panel A, provides the mean coefficients from annual Basu regressions, estimated separately for quintiles of lagged B/M ratios, calculated at the beginning of each year. We also provide results in the bottom row for all five quintiles combined. The sample with available B/M data (122,411 firm-years) is slightly smaller than our full sample. Estimated values of β 1 are positively related to B/M, since they increase monotonically from for the lowest quintile to for the highest quintile. These results are consistent with the findings of prior research, which suggest that low (high) B/M firms are associated with relatively low (high) levels of conditional conservatism. Panel B of Table 4 repeats the Basu regressions, but replaces current earnings (X t /P t-1 ) with lagged earnings (X t-1 /P t-1 ). The same monotonic trend is observed in Panel B: estimated values of β 1 increase from for the lowest B/M quintile to for the highest quintile. The levels of differential timeliness (β 1 ) in Panel B are closer to those in Panel A for lower B/M quintiles, but the gap increases for higher quintiles. Table 5 repeats the analysis for quintiles based on size, measured by market capitalization. The number of firm-years with available data is the same as our full sample (124,562 firm-years). Panel A indicates a monotonic negative relation between estimates of conditional conservatism (β 1 ) based on the original Basu regressions and firm size. The results reported in Panel B, based on the X t-1 /P t-1 regressions exhibit the same monotonic negative relation. As in Table 4, the levels of estimated β 1 in Panel B are similar to those in Panel A for large firms in the high MV quintiles, but the gap increases as size declines. While the cross-sectional patterns observed for lagged earnings in Tables 4 and 5 resemble the patterns observed for contemporaneous earnings, the extend of comovement is not Roychowdhury and Watts, 2007). In contrast, others (e.g., Basu, 2001, and Ball, Kothari, and Nikolaev, 2008) argue that the observed positive relation is expected as a property of income recognition in accounting, and need not be due to measurement error. 19

22 as strong as that observed in Section 4.1 for time-series variation. We emphasize again that the results observed for lagged earnings represent a lower bound for our bias. Also, our estimates for timeliness for bad and good news for the lagged earnings regressions are likely to be biased toward zero because the partitions based on B/M and size are correlated with X t /P t-1, the dependent variable in the returns/earnings regressions for the two subsamples. Given that β 0 is negative and β 0 +β 1 is positive for the lagged earnings specification, the effects that bias the timeliness estimates toward zero result in a downward bias for the conditional conservatism estimate (β 1 ). Our main point, however, is that a substantial portion of the cross-sectional variation in the Basu measure is likely to be driven by cross-sectional variation in our bias. 4.3 Efforts to mitigate biases created by the return variance and loss effects. We investigated a variety of approaches to eliminate the biases we document. Unfortunately, evidence of substantial bias remains in the different specifications we consider. The discussion below provides a summary of those efforts. Details of these results are not tabulated here, but are available from the authors. Even though our results indicate an indirect relation between share price and bias in the Basu measure, we investigate whether introducing simple controls for share price alters our results. For our first control, we include the inverse of lagged price as an additional regressor. Our results indicate evidence of spurious differential timeliness, measured by the coefficient β 1 in the lagged earnings version of equation (1). While the coefficient declines when 1/P t-1 is included, relative to the third row of Table 3, the t-statistic remains unchanged. Investigation of the separate annual regressions indicates a high correlation between the two sets of estimates of β 1, with and without 1/P t-1 as an additional regressor. We observe similar effects on the 20

23 coefficient β 1 when we introduce 1/P t-1 to the contemporaneous earnings version of equation (1), relative to the results reported in the fourth row of Table 3. We repeated the analyses in Tables 4 and 5 on partitions of our sample based on quintiles of share price. We confirm the results in Easton et al. (2009) that estimates of β 1 in the contemporaneous earnings version of equation (1) are significantly positive in all price quintiles, though the magnitudes of those estimates decrease with share price. More important, we show that corresponding estimates of β 1 for the lagged earnings version of equation (1) are also significantly positive in all price quintiles. 11 In addition to efforts to control for lagged price, we considered alternative deflators, such as lagged total assets, as well as unscaled versions of equation (1). 12 Again, we find spurious evidence of asymmetric timeliness, indicated by large and significant estimates of β 1 when the dependent variable is lagged earnings. Finally, to isolate the importance of the return variance effect on the bias we document, we replaced firm returns in equation (1) with corresponding industry returns over the same 12- month period. We consider industry returns based on equally-weighted average returns of firms in the same 2-digit and 4-digit SIC classifications. The industry returns appear to be reasonable proxies for firm-specific returns since the correlations are reasonably high: 0.51 (0.32) for the 4- digit (2-digit) industry classification. Our main finding is that the return variance effect is much lower for industry returns that is, industry return variances are similar for low and high price firms and evidence of asymmetric timeliness also decreases substantially for both the current We find that estimates of β 0, which are significantly negative for the lagged earnings regressions, are more negative for the lowest price (P t-1 ) quintile. This could be due to a combination of two effects: a) the lowest price quintile has higher mean returns in year t, and b) the lowest price quintile is slightly more likely to report losses in t-1. When considering alternative deflators, we replace lagged price in the dependent variable with alternative deflators such as lagged total assets. Unscaled versions of equation (1) are estimated as follows: X it (or X it-1 ) = α 0 P it-1 + α 1 D it * P it-1 + β 0 *AR it * P it-1 + β 1 *D it *AR it *P it- + ε it. 21

24 earnings and lagged earnings regressions. In fact, the 2-digit classification exhibits almost no evidence of the return variance effect and we observe insignificant estimates of β 1 for both the current and lagged earnings regressions. For the 4-digit classification, we observe some residual return variance effect and while the β 1 estimates are positive and significant they are substantially lower than the corresponding coefficients reported in the fourth and third rows of Table 3. While the insignificant estimates for the 2-digit classification are consistent with industry returns representing noisy proxies for firm-specific returns, they are also consistent with the view that eliminating the return variance effect eliminates spurious evidence of conditional conservatism (indicated by insignificant estimates of β 1 for the lagged earnings regressions). 4.4 Relation between our bias and biases noted in Dietrich et al. (2007). The return variance and loss effects that we link to lagged price appear similar to the two effects described in Dietrich et al. (2007) as causing biases in the Basu measure: heteroskedastic return errors and left skewness in earnings. We show below, however, that a) the two sets of biases are unrelated, and b) the two sets of effects are different. The main difference between the two sets of biases is that the Dietrich et al. (2007) analysis requires that returns are determined endogenously, partly from the information released in earnings (and the balance from non-earnings information). We do not require that return endogeneity condition. Because current returns, which reflect this period s unexpected news, should not be derived from last year s earnings information, the Dietrich et al. (2007) analysis cannot explain any of the results we observe for the lagged earnings regressions. 13 For example, neither the large estimates of spurious conditional conservatism (β 1 ) observed for lagged earnings regressions nor the high degree of comovement between those spurious measures and 13 We observe similar spurious results of differential timeliness when we replace current earnings in equation (1) with earnings from two years ago (X t-2 ). 22