RE-EXAMINING ACCOUNTING CONSERVATISM: THE IMPORTANCE OF ADJUSTING FOR FIRM HETEROGENEITY ALAN G. HUANG School of Accounting and Finance, University of Waterloo, Ontario, Canada N2L 3G1 Email: [email protected] YAO TIAN School of Business, University of Alberta, Alberta, Canada T6G 2R6 Email: [email protected] TONY S. WIRJANTO School of Accounting and Finance and Department of Statistics and Actuarial Science, University of Waterloo, Ontario, Canada N2L 3G1 Email: [email protected] Version: March 2011 Abstract In this paper, we examine the role of firm heterogeneity of earnings in measuring accounting conservatism. Prior studies have documented the extent of accounting conservatism (in the form of the asymmetric response of accountings earnings to news) and its tendency to increase over time. However, many of these studies implicitly assume that unobserved firm-specific characteristics (known as firm heterogeneity or firm-specific fixed-effects) are unimportant to earnings determination. We find that after allowing for firm-specific fixed-effects in the earnings determination, (i) the level of accounting conservatism is smaller in magnitude than previously documented, and (ii) conservatism does not increase monotonically over time as has been claimed in prior studies. Our results echo recent studies that call for firm-level measures of conservatism, and emphasize the importance of allowing for firm heterogeneity in measuring accounting conservatism. Keywords: returns; earnings; firm-specific fixed-effects; asymmetry; conservatism. * We are solely responsible for all remaining errors. 1. INTRODUCTION Accounting conservatism is an important characteristic of the Generally Accepted Accounting Principles (GAAP). It refers to “the differential verifiability required for recognition of profits versus losses” (Watts, 2003). Many empirical studies have attempted to quantify the extent of accounting conservatism. In a seminal publication, Basu (1997) studies the timeliness of earnings recognition with respect to stock returns (as a measure for news) in a piecewise linear regression model. He finds that earnings reflect “bad news” faster than “good news”, providing empirical evidence of accounting conservatism. Employing Basu’s (1997) measure as well as a number of other measures, Givoly and Hayn (2000) show that accounting conservatism has increased over time.1 In this paper, we reexamine the appropriateness of the model specifications used in Basu (1997) in the context of firm heterogeneity. We focus on the Basu (1997) measure, since it is the most widely used conservatism measure in the literature (Ryan, 2006). Basu (1997) and many subsequent papers on accounting conservatism typically use a pooled ordinary least squares (OLS) approach to estimate their empirical models of accounting conservatism, where earnings is regressed on stock returns and negative/positive return regimes. This approach inadvertently neglects the panel structure of the data and treats unobserved and unobservable firm-specific characteristics as homogeneous and, thus, unimportant to the determination of earnings. In particular, the OLS measure obscures the cross-sectional variation in conservatism by assuming that all firms are homogeneous. 2 Our findings show that there are significant cross-sectional variations (due to unobserved firm heterogeneity or firm-specific fixed-effects) in earnings. By ignoring firm heterogeneity, the pooled OLS regression model essentially forces return to be the sole determining factor of the cross-sectional variations in earnings. This can result in an omitted 1 Studies of accounting conservatism that focus on high-tech versus low-tech firms include Kwon, Qin, and Han (2006), and studies that focus on the banking industry include Alali and Jaggi (2011) and Anandarajan, Francis, Hasan, and John (2011). In addition, non-US studies of conservatism include Elbannan (2011) for an emerging market, and Anandarajan, Francis, Hasan, and John (2011) for a large number of international markets. 2 Some studies control for industry fixed-effects (for example, Ahmed et al. 2002, Ahmed and Duellman 2007, Francis et al. 2009, Khan and Watts, 2009). Doing so ameliorates the cross-sectional variation problem but does not remove it. Controlling for industry fixed-effects assumes that firms within an industry are homogenous and thus ignores the cross-sectional variation of conservatism within the industry. The adjustment of firm- and time-fixedeffects is also used in other settings in accounting. For example, Beaver and Ryan (2000) use fixed-effects to model the persistent bias in book-to-market ratios due primarily to unconditional conservatism, and use time-specific effects to capture economy wide temporal variation. 1 variable bias in the parameter estimates of the regression model, and thus can lead to an erroneous inference about the extent of and trend in accounting conservatism. Our study is motived by recent studies that call for controlling for firm heterogeneity in estimating measures of conservatism. The literature suggests that conservatism is associated with firm- and economy-specific factors such as contracts (including debt and compensation contracts), litigation, taxation and regulation (Watts, 2003). Adjusting for these firm-level heterogeneities yields a firm-level measure of conservatism. However, traditional conservatism measures are not designed to capture these firm-level variations. Notably, Ryan (2006) comments that “(s)uch measures are currently not available and are desperately needed in order to address many research questions empirically.” It is not until recently that firm-level conservatism measures were introduced in the literature. 3 For example, based on the return decomposition model of Vuolteenaho (2002), Callen, Segal, and Hope (2010) define the ratio of unexpected current earnings to total earnings shocks as a firm-level conservatism measure. Khan and Watts (2009) specify the asymmetric earnings timeliness coefficient in Basu (1997) as a linear function of firm-specific characteristics of size, market-to-book and leverage, and substitute these characteristics into the Basu regression to arrive at a firm-level conservatism measure. In this paper we provide another refinement of Basu (1997) for firm-level conservatism. Since there are potentially many factors that affect variation in conservatism,4 and since the literature is inconclusive on which factors to use, we use a parsimonious setting of firm fixed-effects to control for firm-specific heterogeneity. The literature suggests that fixed-effects provide a mechanism to control for unobservable firm-specific effect in a number of settings. For example, observing that leverage is stable over time for individual firms, Lemmon, Roberts and Zender (2008) find that the unobserved firm-specific effects (as measured by fixed-effects) are responsible for the majority of cross-sectional variation in capital structure. In particular, related 3 Controlling for fixed-effects is a popular way to derive firm-specific measures. For example, Ramirez and Hachiya (2006) use fixed-effects to measure firm-specific organizational capital. 4 For example, Watts’s (2003) survey paper suggests that conservatism is associated with contracts, litigation, taxation and regulation, and Khan and Watts (2009) choose size, market-to-book and leverage as the primary factors that affect conservatism. 2 to this paper, Ball, Kothari and Nikolaev (2011) recently recommend that researchers interested in measuring conservatism in the Basu (1997) framework should control for firm-specific effects in order to avoid potentially spurious inferences. To provide evidence of firm heterogeneity in the earnings regression, we show that a substantial part of the total variation in earnings is attributed to cross-sectional differences. In fact, the unidentified cross-sectional fixed-effects account for more variation in earnings as the traditional returns-earnings specification. The adjusted R-square from a regression of earnings on just firmspecific fixed-effects is 13% in our constant sample of 691 firms from 1976 to 2005. In contrast, the adjusted R-square from the traditional earnings regression (Basu, 1997) is 10% only.5 These results are in line with Lemmon, Roberts and Zender (2008), who call for a fixed-effects study, and provide empirical support for Ball, Kothari and Nikolaev’s (2011) recommendation to control for fixed-effects in estimating the Basu (1997) asymmetric earnings timeliness. Due to the strength of the cross-sectional variations, and hence the importance of firm heterogeneity in earnings determination, we explicitly incorporate firm-specific fixed-effects into Basu’s regression model. We then use this refined model to reexamine the extent of accounting conservatism and its trend over time. We find that, after allowing for firm-specific fixed-effects in the earnings regression, the level of accounting conservatism is much smaller in magnitude than previously documented. In addition, contrary to Givoly and Hayn (2000), we do not find clear evidence to support the assertion that, over time, there is a monotonic increase in accounting conservatism. We also conduct a number of sensitivity analyses to ensure that our results are robust to different model specifications and alternative measures of accounting conservatism, such as employing different samples, and using cash flows from operations as a proxy for news (Ball and Shivakumar 2006). Overall, our results highlight the importance of allowing for firm heterogeneity in the regression models that use asymmetric response of accounting earnings to news as a conservatism measure. 5 Consistent with these results, model specification tests conducted in this paper reject the assumption of firm homogeneity. 3 Our econometric refinements of the Basu (1997) specifications are intended to provide an agnostic firm-level measure of conservatism. As such, this paper joins several recent studies in addressing the efficacy of the Basu measure. Dietrich, Muller and Riedl (2007) show that the Basu specification gives rise to evidence consistent with accounting conservatism even in the absence of accounting conservatism; that is, Basu (1997) overstates the occurrence of accounting conservatism. Patatoukas and Thomas (2009, 2010) further support the views espoused by Dietrich, Muller and Riedl (2007) by arguing that the scale variable used in the Basu (1997) regressions entails an effect that biases the Basu estimator. Contrasting these views, Givoly, Hayn and Natarajan (2007) claim that Basu’s (1997) measure understates accounting conservatism. In addition, Ball, Kothari and Nikolaev (2011) challenge the views of Patatoukas and Thomas (2010) by arguing that the bias in the Basu estimator is in fact caused by the correlation between the expected values of earnings and return. Ball, Kothari and Nikolaev (2011) advocate for using fixed-effects in the Basu (1997) regression to correct for this bias. Our paper takes this suggestion seriously: we explicitly control for firm fixed-effects. We also contribute to the literature by showing that controlling for fixed-effects considerably weakens the trend of conservatism. The remainder of the paper is organized as follows. In section 2, we briefly review studies on accounting conservatism and present the pooled OLS earnings regression model originally employed in Basu (1997) to measure the extent of accounting conservatism. In section 3, we discuss the potential shortcomings of the pooled OLS regression model, and introduce the fixedeffect model to overcome some of these problems. In Section 4, we first demonstrate the importance of firm heterogeneity in earnings determination, and then report the findings on the extent of accounting conservatism and its trend over time after allowing for firm-specific fixedeffects in the earnings regression. We conduct four sets of sensitivity analysis in Section 5 to demonstrate the robustness of our results. Finally, Section 6 concludes. 2. ACCOUNTING CONSERVATISM Over the years, a considerable amount of research has investigated the extent of accounting conservatism, its impact on the quality of financial reporting, and its trend over time. In these 4 studies, researchers have introduced a number of definitions of conservatism. For instance, Belkaoui (1985) broadly defines conservatism as “reporting the lowest values of assets and revenues and the highest values of liabilities and expenses.” Taking a balance-sheet perspective, Penman and Zhang (2002) define conservative accounting as “choosing accounting methods and estimates that keep the book values of net assets relatively low.” Basu (1997) argues that “conservatism in the balance sheet is of dubious value if attained at the expense of conservatism in the income statement, which is far more significant”. Taking an income-statement perspective, he defines accounting conservatism as “the practice of reducing earnings in response to ‘bad news’, but not increasing earnings in response to ‘good news’.” In this paper, we adopt Basu’s convention and define accounting conservatism in terms of the asymmetric timeliness of earnings to reflect “good news” and “bad news”. To measure the extent of accounting conservatism, Basu (1997) proposes a piecewise linear regression: EPSit / Pit 1 0 1Rit 2 Dit 3 ( Dit Rit ) uit (1) In the above regression, EPS it is the earnings per share of firm i in fiscal year t, Pit 1 is the price per share of firm i at the beginning of the fiscal year t, Rit is the discretely compounded annual return of firm i in fiscal year t, and Dit is a dummy variable set equal to 1 when Rit 0 (bad news) and to 0 when Rit 0 (good news); and u it is the unobserved zero-mean error term. In (1), the slope coefficient, 3 , measures the incremental responsiveness of earnings to bad news over earnings to good news. It is expected to be positive and significant under a conservative reporting system. Basu (1997) estimates this model as a pooled cross-sectional regression and finds that earnings is about four and a half times ( ( 1 3 ) / 1 4.66 ) as sensitive to negative returns as it is to positive returns. These results provide strong evidence for accounting conservatism. Since the publication of Basu (1997), the above framework has been widely adopted to measure the extent of accounting conservatism. It is, in fact, the “most widely used conservatism measure 5 in the literature” (Ryan, 2006). Givoly and Hayn (2000) use Basu’s (1997) measure as well as other measures to examine the trend in conservatism over time. Estimating the model in a pooled OLS regression model using the intersection of firms in Standard and Poor’s Compustat and the Center for Research in Security Prices (CRSP) from 1950 to 1998, these authors find that the timely recognition of “bad news” relative to “good news” increased over time. In other words, the authors document that there is an increasing trend in accounting conservatism. Like most studies on earnings-returns relation, Basu (1997) and Givoly and Hayn (2000) estimate the returns-earnings relation by using the standard pooled OLS approach. It is important to note that the pooled OLS estimator in equation (1) treats sample observations as being serially uncorrelated for a given firm with homoskedastic errors across firms and time periods. However, as we will show empirically in Section 4 of this paper, this assumption is rather restrictive. In fact, we find there are significant cross-sectional variations (i.e., firm heterogeneity) in earnings determination. In the presence of this firm heterogeneity, the pooled OLS estimates of ( 0 , 1 , 2 , 3 ) ' in (1) are biased and inconsistent. This is our underlying motivation to refine Basu’s (1997) model by explicitly allowing for firm-specific fixed-effects. We then use the refined model to reexamine the extent of accounting conservatism and its trend over time. 3. MODEL DEVELOPMENT—ACCOUNTING FOR FIRM HETEROGENEITY There are two methods to account for firm heterogeneity: the fixed-effects model and the random-effects model. On one hand, the fixed-effects model can be estimated using three alternative approaches: between-firms, within-firms, and Least Squares Dummy Variable (LSDV) estimations. This model is particularly suitable when the differences between individual firms can be reasonably viewed as parametric shifts in the regression function itself. For instance, an example for this is, when the cross-sections used in the estimation represent a broadly exhaustive sample of the population of firms. On the other hand, if the cross-sections are drawn from a larger population, so that the sample firms may not be reasonably considered exhaustive, then it is more appropriate to view firm-specific terms in the sample as randomly distributed over the full cross-sections of firms, and to instead apply the random-effects model.6 6 See, for example, Greene’s (2003) textbook for an illustration of fixed-effects and random-effects models. 6 In untabulated results, 7 we perform a series of model specification tests in our sample to determine the most appropriate panel-data regression model for equation (1). Our test results confirm the presence of firm-specific effects and suggest that estimates of the parameters in the linear piece-wise earnings regression in equation (1) obtained by the pooled OLS are biased and inconsistent. Furthermore, our test results favor the fixed-effects model with the LSDV estimation approach over the RE model and other fixed-effects models. Consequently, our subsequent analyses in the remainder of the paper will focus on estimates obtained from the LSDV/FE specification. Let i=1,…,N denote firms, and t 1,..., T denote time periods. The fixed-effects model is specified as follows: N EPS it / Pit 1 = 0 1 Rit 2 Dit 3 ( Dit Rit ) j F jit + uit , (2) j 2 where Fijt is the firm dummy variable (i.e., Fijt 1 for j=i and Fijt 0 for i j ; i,j=2,…N), j is the individual firm-specific effect which is assumed to be time invariant, and u it is the unobserved zero-mean error term. In the context of our study, j includes firm-specific characteristics of a company that are candidates for additional explanatory variables in the model. If there is no firm heterogeneity in the data, this model would reduce to the original Basu model as j 0 for all j. Note that the regression model in equation (2) can be extended to include time fixed-effects as follows: N T j 2 s 2 EPSit / Pit 1 = 0 1 Rit 2 Dit 3 ( Dit Rit ) j F jit sYsti + uit , (3) where Ysti is the time dummy ( Ysti = 1 for s = t and Ysti = 0 for s t ; s,t = 2,…T). Consistent with the literature that calls for firm-level conservatism measures (e.g., Ball, Kothari and Nikolaev, 2011), in this paper we choose to present results from regression (2) to emphasize the importance of allowing for firm heterogeneity in measuring accounting conservatism. We note that adding 7 The results are available from the authors upon request. 7 time fixed-effects does not change the qualitative results appreciably. The estimation results are available upon request. 4. EMPIRICAL RESULTS In this section, we will first describe the sample, and provide empirical evidence on the importance of firm heterogeneity. We will then use the refined model in equation (2) to account for firm-specific fixed-effects and reexamine the extent of accounting conservatism and its trend over time. 4.1 Data Our sample covers NYSE/AMEX/NASDAQ firms for the period of 1976-2005. We first consider a constant sample where firms have non-missing observations in all years of the entire sample period. In robustness checks, we use an alternate sample that allows for new listings and de-listings of firms. The accounting data used in this study are retrieved from Compustat, and the return data are from CRSP. The dependent variable in equation (1), EPSit / Pit 1 , is calculated as earnings per share (Compustatvariable EPSPX) divided by the lagged-one fiscal-year-end closing price (Compustat variable PRCC_F). Firm’s annual stock return is defined as the cumulative monthly returns for the period from 9 months before the fiscal year end to 3 months after the fiscal year end. Finally, as in Basu (1997), we winsorize the top and bottom 1% of the earnings and returns variables to exclude outliers. This results in a final sample of 691 firms for the time period from 1976 to 2005. In addition, in robustness checks, we use cash flows from operation (Compustat variable OANCF) to proxy for economic gains/losses. Since cash flows from operation is only available after 1988, we construct a constant sample of firms over the period from 1988 to 2005 by requiring firms to have a non-missing value in accruals and cash flows from operations every year. This alternative sample has a total of 995 firms. 4.2 Evidence of the Importance of Firm Heterogeneity 8 We now show that the variations in earnings and returns for the pooled sample are attributed mainly to firm-specific cross-sectional variations. Table 1 reports the full-sample standard deviations, the time-series mean of cross-sectional standard deviations, and the cross-sectional mean of time-series standard deviations of individual firms on earnings and stock returns. We observe that the means of the cross-sectional volatilities for both variables are almost as large as the pooled volatility, whereas the means of the time-series volatilities of the individual firms are much smaller. As a further illustration of this point, Figure 1 plots the ratios of the crosssectional volatility of each variable to its pooled counterpart over time. As shown in the figure, the high ratios between the means of the cross-sectional sample volatilities and pooled sample volatility are not due to any particular sub-periods, as the ratios remain high across time. Collectively, Table 1 and Figure 1 show that it is the between-firm differences instead of withinfirm differences that drive the pooled sample variations. [Table 1 about here.] [Figure 1 about here.] The fact that the sample variation of return is mainly driven by cross-sectional variation justifies the use of LSDV/FE.8 However, it is important to point out that the original Basu model does not account for the between-firm differences. By ignoring the firm-specific fixed-effects, the pooled OLS model essentially forces the returns to be the sole factor determining the crosssectional variations (firm heterogeneity) in earnings. This can lead to an omitted variable bias in the returns-earnings regression, and consequently an erroneous interpretation of the extent of accounting conservatism. In our refined model, we capture this unobserved firm heterogeneity using the firm-specific fixed-effects in equation (2). Before we estimate the fixed-effects regression of equation (2), it is necessary to examine how much of the documented “conservatism” effect (as captured by the returns variable) in Basu’s pooled OLS model is attributable to the unobserved firm-specific fixed-effects. To this end, we 8 If there is little cross-sectional variation in return, then a between-firms regression, which averages variables over time across firms, would also be appropriate, since in this case what matters is the time-series changes in variables. This is the approach taken by Grambovas, Giner and Christodoulou (2006). 9 use the analysis of covariance (ANCOVA) to decompose the variations in the earnings due to the different factors based on the following piecewise linear fixed-effects regressions in (a) to (f).9 N EPS it / Pit 1 = 0 j F jit + u it (a) j 2 T EPS it / Pit 1 = 0 sYsti + u it (b) s2 N T j 2 s2 EPSit / Pit 1 = 0 j F jit sYsti + u it (c) EPSit / Pit 1 = 0 1Rit 2 Dit 3 ( Dit Rit ) + uit (d) N EPS it / Pit 1 = 0 1 Rit 2 Dit 3 ( Dit Rit ) j F jit u it (e) j 2 N T j 2 s 2 EPS it / Pit 1 = 0 1 Rit 2 Dit 3 ( Dit Rit ) j F jit s Ysti + uit (f) The ANCOVA results are reported in Table 2. Each column in Table 2 corresponds to a different specification in (a)-(f). The numbers in each column, excluding the last row, are fractions of total Type III partial sum of squares for each model. That is, the partial sum of squares for each effect is scaled by the aggregate partial sum of squares of all effects in the model, to sum each column to unity. As a result, the values in the table correspond to the fractions of the model’s sum of squares due to particular effects. A special case arises when there is only one effect in the model. In this case, the total explained sum of squares is due to that effect only. For instance, in column (a), when only firm-specific fixed-effects are included in the model, the estimate takes on a value of 1.00 (or 100%). [Table 2 about here.] The last row of Table 2 reports the value of the adjusted R-square for each specification. Column (a) in the table shows that firm-specific fixed-effects alone account for 13% of the observed 9 Lemmon, Roberts and Zender (2008) use a similar method in the context of capital structure determinants and find that the traditional determinants for capital structure explain little of firms’ actual leverage once firm fixed-effects are controlled for. 10 variation in earnings. Column (d) represents the results from the pooled OLS model in Basu (1997). The value of the adjusted R-square for this particular specification is 10%, lower than that from the model with only firm-specific fixed-effects. Column (e) shows our proposed specification, which adds the firm-specific fixed-effects to the Basu model in column (d). The adjusted R-square rises from 10% in column (d) to 20% after including the firm-specific fixedeffects. In addition, a substantial 82% of the explanatory power in this model is captured solely by the firm-specific fixed-effects.10 In summary, the results of the variance decomposition show that earnings contain a nontrivial firm-specific component that is not captured by the pooled OLS specification in Basu (1997). Since a substantial part of the total variation in earnings emanates from the cross-sectional differences, there is a need to control for firm fixed-effects in the Basu (1997) model. 4.3 Empirical Results on the Extent of Accounting Conservatism and its Trends 4.3.1 The Extent of Accounting Conservatism Given the presence of firm-specific fixed-effects in the returns-earnings relation, we adopt a panel-data methodology to estimate the refined Basu model to account for firm-specific fixedeffects and reexamine the extent of accounting conservatism. For comparison purposes, we also estimate the original Basu model using the pooled OLS approach, and report the estimation results from both models in Table 3.11 [Table 3 about here] As Table 3 shows, the refined model has a much larger adjusted R-square than the original model (i.e., 0.196 versus 0.097), suggesting a better fit for the refined model. Also, both models produce statistically significant estimates for the parameters ( 0 , 1 , and 3 ), all with expected 10 We also observe in Table 2 that the contribution from the year fixed-effects is significant. In this paper, we follow the literature and focus on the effect of firm heterogeneity in earnings determination. As previously noted, including the year fixed-effects in the earnings regression does not change the qualitative results. 11 The standard R-squared value is not an appropriate goodness-of-fit measure for panel-data regressions since it can yield numerical values outside the interval [0,1]. Instead, a generalization of the R-squared value due to Theil (1961) is reported for LSDV/FE regressions. The Theil R-squared value is defined as one minus the sum of squared errors divided by the sum of squares of the (transformed) dependent variable. 11 signs. Recall that 0 captures the current recognition of unrealized gains from the prior period. The estimated coefficients of 0 are positive and highly significant in both models, suggesting a delayed recognition of gains in a conservatism accounting system. The estimates of 2 in both models are very small in magnitude and statistically insignificant, suggesting that the current recognition of prior-period gains (as captured by the constant term, 0 ) does not change with the type of news received in the current period. The estimates of the coefficient 3 (which measures the incremental responsiveness of earnings to bad news) are positive and statistically significant in both models, providing evidence for accounting conservatism. However, the estimated 3 from the refined model is much smaller in magnitude than that from the original model (0.139 versus 0.205). A test of the hypothesis that 3 = 0.205 in the refined model is strongly rejected at below the 1% significance level, indicating that the refined model gives different levels of accounting conservatism than the original model does. In particular, according to the original model, earnings is about 8.59 (= (0.027 + 0.205)/0.027) times as sensitive to negative returns as it is to positive returns, while the refined model suggests that earnings is about 4.48 ( = (0.040 + 0.139) /0.040) times as sensitive to negative returns as it is to positive returns. A comparison of the estimate of 3 and the ratio of ( 1 3 ) / 1 between the pooled OLS model and their counterparts in the LSDV/FE model reveals that they are significantly different at the 1% level. Overall, these results suggest that after correcting for firm heterogeneity, the actual extent of accounting conservatism is only about one half of the level as previously documented. 4.3.2 The Trend in Accounting Conservatism In examining the trend in conservatism over time, Givoly and Hayn (2000) estimate the Basu (1997) model using pooled OLS and find that the timely recognition of “bad news” relative to “good news” has increased over time. Their results provide evidence for an increasing trend in accounting conservatism.12 12 In the Appendix, we replicate Givoly and Hayn’s (2000) analysis based on Basu’s measure in our sample. Using their sample period, our replication results are very close to what they have reported. However, their findings appear 12 Given the ubiquity of the Basu model to measure conservatism (Ryan, 2006), and given the significant difference in our findings regarding the extent of accounting conservatism after allowing for firm-specific fixed-effects, it is worthwhile to reexamine the trend in accounting conservatism by using the refined Basu model that explicitly accounts for firm heterogeneity. To this end, we divide the full sample period into sub-periods with a five year interval, as well as sub-periods corresponding to business cycles defined as periods from NBER business recession to NBER business peak. We estimate both the original pooled OLS model and our refined LSDV/FE model over these various sub-periods. We report the estimation results in Table 4. [Table 4 about here.] A number of observations warrant discussion. First, contrary to the findings in Givoly and Hayn (2000), the refined model does not show an increasing trend in conservatism, as evidenced in both the estimate of 3 and the ratio of ( 1 3 ) / 1 . As for 3 , its estimates from the refined model are positive and significant in each of the five-year intervals from 1976 to 2005, providing evidence of accounting conservatism in each sub-period. However, the estimated values of 3 from the refined model show an overall declining trend over time. This indicates that, the first half of the sample period has a larger estimate of 3 than the second half. In particular, the estimate of 3 drops from 0.132 in the sub-period of 1976-1980 to 0.033 in the sub-period of 1996-2000. This indicates a significant reduction in asymmetry during this period. An exception is the sub-periods of 1986-1990 and 2001-2005, during which the estimate rose from 0.072 in 1981-1985 to 0.147 in 1986-1990, and from 0.033 in 1996-2000 to 0.065 in 2001-2005 respectively. As for ( 1 3 ) / 1 , the conservatism trend as measured by this ratio in the refined model is relatively stable across these sub-periods. The ratio, which measures the sensitivity of earnings to bad news relative to their sensitivity to good news, always exceeds unity, suggesting a greater tendency of firms to recognize bad news in a more timely fashion than good news. Once firm-specific fixed-effects are incorporated into the regression specification, there is no to be both sample specific and estimation-method specific. In particular their main findings are reversed once the sample period is extended to cover the period of 1995-2004. 13 clear evidence that the incremental responsiveness of earnings to bad news is stronger in later years than in early years. Overall, our results cast serious doubts on the assertion that the trend of conservatism is increasing over time. Second, similar to the results reported in Table 3, the estimated values of 3 and ( 1 3 ) / 1 in the refined model are consistently smaller than their counterparts in the original model in all subperiods, providing evidence that the results reported in Table 3 for the whole sample is not likely caused by a particular sub-period. The estimated values in the former are usually one half or a third of those in the original model, confirming that after incorporating firm heterogeneity in the earnings regression, the measured degree of accounting conservatism is weaker than previously reported. Finally, the estimate value of 1 , which measures earnings’ responsiveness to good news, in the refined model is larger than that in the original Basu model. A comparison between the first half and the second half of the sample period reveals that the estimate of 1 tends to decrease over time in both the original pooled OLS and the refined LSDV/FE models. These results indicate that (i) controlling for firm heterogeneity results in earnings being more responsive to good news than is the original Basu (1997) setting and (ii) not only is earnings less responsive to bad news over time, it is also less responsive to good news over time—that is, earnings is less sensitive to news. Overall, these results further weaken the argument of asymmetric timeliness of earnings of Basu (1997), in that earnings’ responsiveness to news is now less asymmetric. When the sample period is partitioned according to the NBER business cycles, the results are similar. The estimated value of 1 in the refined model shows a steady decline from 0.076 in 1976-1979 to 0.023 in 2001-2005, which is again consistent with the idea that earnings reflect positive returns on a less timely basis over the business cycles. The corresponding estimates of 3 show a similar pattern to those obtained for the five-year intervals. In addition, the estimate of ( 1 3 ) / 1 over the business cycles shows a similar pattern to the ratio obtained over the subperiods with the five-year interval. Again, the evidence corroborates that earnings are also less timely in reflecting bad news over time. 14 5. 5.1 SENSITIVITY ANALYSES An Alternate Sample Allowing for New Listings and Delistings Our previous results use the sample filtering procedure of Givoly and Hayn (2000), which require all firm-year observations to be non-missing. While this sample selection procedure makes our study more comparable to Givoly and Hayn (2000), we note that Basu (1997) does not impose this condition. We now confirm the robustness of our results by using the Basu (1997) sample selection procedure that allows for new listings and delistings of firms. Specifically, to allow for meaningful fixed-effects regression, we require that each firm has at least 10 observations in our sample period of 30 years. This sample covers 6,041 firms, with an average cross section of 3,623 firms. Compared with the 691 firms in Tables 3 and 4, this sample is about four times larger. We then repeat the exercises of Tables 3 and 4 with this sample. The results are presented in Table 5. We observe that the results are similar to, if not better than, those presented in Tables 3 and 4—the percentage reduction in the estimate of 3 in Table 5 is generally larger than that in Tables 3 and 4. In other words, for the full sample, after controlling for firm fixed-effects, (1) the degree of conservatism is much smaller in both the full-sample (Panel A) and each of the five-year sub-periods (Panel B) as indicated by the estimated values of 3 and ( 1 3 ) / 1 ,13 and (2) conservatism is decreasing over time as suggested by the estimates of 3 or remains relatively stable as suggested by the estimates of ( 1 3 ) / 1 . In sum, these results reiterate the findings in Tables 3 and 4. [Table 5 about here.] 5.2 The Returns-Cash Flow Regression The basic reasoning in Basu (1997) is that the parameter 3 in equation (2) or (3) captures a more sensitive relationship between earnings and returns when returns are negative. However, it 13 For each of the five-year sub-periods, we further require that firms have three non-missing observations in that sub-period for the fixed-effects adjustment to be meaningful. 15 is possible that a significant estimate of 3 is driven instead by the changes in returns. Abarbanell and Bernard (1992), Sloan (1996) and Ball and Bartov (1996), for instance, argue that the market may not fully understand the earnings process. In particular, Abarbanell and Bernard (1992) reason that the market tends to over-react to bad news. Therefore, it is possible that the market believes that negative earnings changes are more permanent than positive earnings changes. This would give rise to the estimate of 3 being driven by changes in returns rather than being a measure of earnings conservatism as suggested by Basu (1997). The key question here is whether the market properly interprets earnings changes. To facilitate this interpretation, we follow Basu (1997) and decompose earnings into its cash flow and accrual components. This leads to a piecewise linear cash flows regression with firm-specific fixedeffects: N CFOit / Pit 1 = 0 1 Rit 2 Dit 3 ( Dit Rit ) j F jit + uit , (4) j 2 where CFO denotes cash flows from operation. If the significance of the coefficient 3 in equation (2) is due to accruals anticipating bad news, then when cash flows is used as the dependent variable in the regression as in equation (4), the significance of 3 in equation (4) should be diminished or, better, vanished. The results in Table 6 confirm that this is the case. [Table 6 about here.] Table 6 shows that the magnitudes of the estimates of 3 are invariably much smaller than those in the earnings regression. This is particularly the case for the LSDV/FE estimates reported in the table. Moreover, for the sub-periods of 1994-2000 and 2001-2005, the LSDV/FE estimates of 3 are not significant. This is in line with the above accruals hypothesis of earnings conservatism: Current period cash flow is more likely to be affected by poor performance of firms, and, thus, the observed increased sensitivity between current period cash flow and current returns are, to some extent, driven by the changing nature of the returns.14 14 We note that the estimate of 3 for the sub-period of 1988-1993 is significant (with a t-statistic of 3.20). However, at 0.070, the magnitude of the estimate is small. 16 We further provide an alternative measure of earnings conservatism: the implied effect of accruals, defined as the difference between the estimated coefficient of 3 from the earnings regression and the estimated coefficient of 3 from the cash flows regression. If the accrual hypothesis of conservatism is true, the implied effect of accruals should be small and a large effect would indicate the existence of conservatism. The results are reported in the Table 7. [Table 7 about here.] The estimates of the pooled OLS model provide evidence of accounting conservatism, as illustrated by a large and significant difference between the estimates in columns 1 and 2. However, the LSDV/FE estimates indicate that the differences between the two estimates are small in magnitude and not statistically significant at the 5% level. Interestingly, the difference in the sub-period of 1988-1993 is negative, implying that earnings may have been boosted by accruals when firms’ poor performance is anticipated by returns. However, this difference has a t-statistic of 0.45 only. Thus, after controlling for the firm-specific fixed-effects, the evidence of accounting conservatism diminishes. Tables 6 and 7 are consistent with our previous findings that accounting conservatism is much weaker than has been documented. 5.3 Including Lagged Returns in Basu’s (1997) Model Kothari and Sloan (1992), Collins, Kothari, Shanken and Sloan (1994), and Beaver and Ryan (2000), among others, examine how well current and past returns explain current earnings. They find that returns in all three previous periods contain information about current earnings. Based on this argument, we extend the specification in equation (2) to allow for past returns to explain current earnings.15 The inclusion of the firm-specific fixed-effects leads to the following paneldata specification: EPSit / Pit 1 = 0 10 Rit 11Rit 1 12 Rit 2 13 Rit 3 20 Dit 21Dit 1 22 Dit 2 23 Dit 3 N 30 ( Dit Rit ) 31 ( Dit 1 Rit 1 ) 32 ( Dit 2 Rit 2 ) 33 ( Dit 3 Rit 3 ) i Fijt + uit j 2 15 Ryan and Zarowin (2003) study a similar specification without incorporating firm-specific fixed-effects. 17 (5) In equation (5), the parameters ( 10 , 11 , 12 , 13 ) measure the impacts of current and past returns on current earnings. The parameters ( 20 , 21, 22 , 23 ) are the shifting coefficients when there is bad news in returns in the period, and the parameters ( 30 , 31, 32 , 33) are the shifting coefficients on returns when there is bad return news in the period. If the accounting system is conservative and earnings reflect news for up to three prior periods, we should expect each of the estimates in ( 30 , 31, 32 , 33) to be positive. Moreover, the sum of coefficients, 30 31 32 33 , measures the aggregate degree of conservatism (i.e., the incremental responsiveness of earnings to negative news above that to positive news in the three years after the news occurs). The estimation results of equation (5) are provided in Table 8. A comparison between the pooled OLS and LSDV/FE results again re-emphasizes the importance of controlling for firm-specific fixed-effects in the earnings regression. In particular, the extent of accounting conservatism as measured by the pooled OLS estimates 30 , 31 , 32 , 33 is uniformly greater than their LSDV/FE counterparts, both in the full sample (Panel A) and in sub-periods of 1980-1989, 1990-2000, and 2001-2005 (Panel B). Furthermore, there is no evidence of an increasing trend of conservatism from the LSDV/FE regression. First, recall that the estimate of the coefficient 30 measures the shift in the association between returns and earnings when returns are negative. Between 1980 and 2000, there is a decreasing trend in the incremental responsiveness of earnings to bad news; the coefficient estimate then stabilizes in the period of 2001-2005. Second, the responsiveness of earnings to lagged returns displays a similar declining trend, as shown in the smaller estimates of 31 , 32 and 33 over time. Consequently, the aggregate indicator of accounting conservatism, 30 31 32 33 , also declines over time. In sum, the results in Table 8 corroborate our previous findings that conservatism is weaker than has been documented once firm fixed-effects are controlled for. 5.4 Alternative measures of Accounting Conservatism 18 Dietrich, Muller and Riedl (2007) argue that Basu’s (1997) approach to assess asymmetric timeliness can yield biased estimates when earnings information affects returns. In light of such criticisms, we follow Ball and Shivakumar (2006) to use alternative models to measure accounting conservatism. These models do not rely on the returns-earnings relation to measure conditional conservatism. Rather, they rely on the relationship between accruals and cash flows. Operating cash flows are used to determine good news versus bad news. The asymmetrically timely recognition of bad news would result in a greater association between accruals and cash flows relative to good news. This argument suggests that we can use Ait (the total accruals of firm i in fiscal year t scaled by total assets) as the dependent variable in the piecewise linear panel-data regression. The independent variable, CFOit , is the operating cash flows of firm i in fiscal year t (Compustat variable OANCF) scaled by total assets (Compustat variable AT). This leads to the following three piecewise linear accruals panel-data regressions. The first one is based on the cash flow model: N Ait 0 1CFOit 2 Dit 3 ( Dit CFOit ) j F jit uit (6) j 2 where Dit is a dummy variable set equal to 1 if CFOit is negative and 0 otherwise. In the second and third models, we control for other variables that could potentially affect accruals. In particular, the second model is based on the linear accrual model proposed by Dechow and Dichev (2002): N Ait 0 1CFOit 2 Dit 3 ( Dit CFOit ) 1CFOit 1 2CFOit 1 j F jit uit , (7) j 2 and the third model is based on the linear accrual model proposed by Jones (1991): N Ait 0 1CFOit 2 Dit 3 ( Dit CFOit ) 1REVit 2GPPEit j F jit uit , (8) j 2 where REVit is the change in revenue ( REVit , measured as Compustat variable SALE) in year t, REVit REVit 1 , and GPPEit is the gross property, plant, and equipment (Compustat variable PPEGT). Each variable in equations (6), (7) and (8) is scaled by total assets. 19 As a further robustness check, we follow Ball and Shivakumar (2006) and consider, instead of the level, the change in cash flow, CFOit CFOit CFOit 1 , as another financial reporting measure to proxy for good and bad news. We repeat the estimation of equations (6), (7) and (8) with CFOit in place of CFOit and define Dit based on CFOit . The results from estimating the above three models are reported in Table 9 for the pooled OLS and LSDV/FE specifications. In Panel A, the level of CFO is used to proxy for economic loss. In Panel B, the change in CFO is used to proxy for economic loss. The asymmetry implies that the coefficient on the interaction terms ( Dit CFOit in Panel A and Dit CFOit in Panel B) should be positive. This is confirmed in Table 9. That is, in panel A, the incremental loss coefficients are all positive, ranging from 0.627 to 0.735 across the three accrual models for the pooled OLS specification, and from 0.600 to 0.655 for the LSDV/FE specification. The LSDV/FE estimates of the asymmetry are consistently smaller than those from the pooled OLS model. However the F-test rejects the pooled OLS specification in favor of the LSDV/FE specification at the 5% level of significance. In Panel B, again all of the incremental loss coefficient estimates are positive, ranging from 0.071 to 0.199 for the pooled OLS specification, and from 0.032 to 0.104 for the LSDV/FE specification, with the LSDV/FE estimates uniformly smaller than the pooled OLS estimates. Once again, the pooled OLS specification is rejected at the 5% level of significance in favor of the LSD/FE specification. These results are consistent with the earlier results using stock return serving as a proxy for economic loss. [Table 9 about here.] Overall, the results obtained from Table 9 confirm the findings that a timely loss-recognition plays an important role in accounting for accruals when we use either the level of, or the change in, cash flows as a proxy for gains and losses. More importantly, the magnitude of asymmetric responsiveness of earnings to bad news declines once firm-specific fixed-effects are controlled for. The decline is particularly substantial with the change in cash flows serving as a proxy for economic gains or losses. For example, in Panel B, the Jones model has a coefficient estimate on 20 D CFO of 0.119 in the pooled OLS regression. The coefficient estimate reduces to only 0.032 in the LSDV/FE regression, amounting to a reduction of over 70%. In summary, the results emerging from our previous sensitivity analyses reinforce the importance of incorporating firm specific fixed-effects in the models intended for measuring accounting conservatism. 6. CONCLUSION In this paper, we demonstrate that a significant part of the total variation in the earnings is attributed to the cross-sectional differences. Given the importance of firm heterogeneity, we propose to refine the original Basu (1997) model of accounting conservatism to explicitly account for firm-specific fixed-effects. We then use the refined model to reexamine the extent of accounting conservatism and its trend over time. We find that the estimated incremental responsiveness of earnings to bad news is not as strong as previously documented, and that there is no clear trend of an increase of accounting conservatism over time. We conduct sensitivity analyses to demonstrate that our results are robust to different model specifications. Our results echo recent studies that call for firm-level measures of conservatism. Noting that conservatism is associated with a number of firm-specific factors (Watts, 2003), Ryan (2006) comments that firm-specific measures are “desperately needed” for empirical studies. In a response to Ryan (2006), Ball, Kothari and Nikolaev (2011) recommend that researchers interested in measuring conservatism in the Basu (1997) framework should control for firmspecific effects in order to avoid potentially spurious inferences. We provide empirical supports for the call in Ball, Kothari and Nikolaev (2011) by showing that a substantial part of the total variation in earnings is attributed to cross-sectional differences. Furthermore, focusing on the most widely used conservatism measure of Basu (1997) (Ryan, 2006), we show that accounting conservatism does not appear to show an increasing trend, contrary to what is previously believed. In sum, our study demonstrates the importance of adjusting for firm heterogeneity in examining accounting conservatism. 21 REFERENCES Abarbanell, J. S. & Bernard, V. L. (1992). Tests of analysts’ overreaction/underreaction to earnings information as an explanation for anomalous stock price behavior. Journal of Finance 47, 1181-1207. Ahmed, A.S., Billings, B.K., Morton, R.M. & Stanford-Harris, M. (2002). The role of accounting conservatism in mitigating bondholder-shareholder conflicts over dividend policy and in reducing debt costs. The Accounting Review 77, 867-890. Ahmed, A. S. & Duellman, S. (2007). Accounting conservatism and board of director characteristics: an empirical analysis. Journal of Accounting and Economics 43, 411-437. Alali, F. & Jaggi, B. (2011). Earnings versus capital ratios management: Role of bank types and SFAS 114. Review of Quantitative Finance and Accounting 36, 105-132. Anandarajan, A., Francis, G., Hasan, I. & Kose, J. (2011). Value relevance of banks: Global evidence. Review of Quantitative Finance and Accounting 36, 33-55. Ball, R. & Bartov, E. (1996). How naive is the stock market's use of earnings information? Journal of Accounting and Economics 21, 319 - 337. Ball, R., Kothari, S.P. & Nikolaev, V.A. (2011). Econometrics of the Basu asymmetric timeliness coefficient and accounting conservatism. Working Paper, University of Chicago Booth School of Business. Ball, R. & Shivakumar, L. (2006). The role of accruals in asymmetrically timely gain and loss recognition. Journal of Accounting Research 44, 207–242. Basu, S. (1997). The conservatism principle and the asymmetric timeliness of earnings. Journal of Accounting and Economics 24(1): 3–37. Beaver, W. & Ryan, S. (2000). Biases and lags in book value and their effects on the ability of the book-to-market ratio to predict book return on equity. Journal of Accounting Research 38, 127-148. Belkaoui, A. (1985). Accounting Theory. San Diego: Harcourt Brace Jovanovich International Edition. Callen, J., Segal, D. & Hope, O. (2010). The pricing of conservative accounting and the measurement of conservatism at the firm-year level. Review of Accounting Studies 15, 145178. Collins, D., Kothari, S. P., Shanken, J. & Sloan, R. (1994). Lack of timeliness and noise as explanations for the low contemporaneous return-earnings association. Journal of Accounting and Economics 18, 289-324. Dechow, P. M. & Dichev, I. D. (2002). The quality of accruals and earnings: The role of accrual estimation errors. The Accounting Review 77 (Supplement), 35-59. 22 Dietrich, J., Muller, K. & Riedl, E. (2007). Asymmetric timeliness tests of accounting conservatism. Review of Accounting Studies 12, 95-124. Elbannan, M. A. (2011). Accounting and stock market effects of international accounting standards adoption in an emerging economy. Review of Quantitative Finance and Accounting 36, 207-245. Francis, B., Hasan, I., Park, J.C. & Wu, Q. (2009). Gender differences in financial reporting decision-marking: evidence from accounting conservatism. Working paper, Rensselaer Polytechnic Institute. Givoly, D. & Hayn, C. (2000). The changing time-series properties of earnings, cashflows and accruals: Has financial reporting become more conservative? Journal of Accounting and Economics 29, 287-320. Givoly, D., Hayn, C. & Natarajan, A. (2007). Measuring Reporting Conservatism. The Accounting Review 82, 65-106 Grambovas, C.A., Giner, B. & Christodoulou, D. (2006). Earnings conservatism: Panel data evidence from the European Union and the United States. Abacus 42, 354–378. Greene, W. H. (2003). Econometric analysis. Prentice-Hall Inc., New Jersey. Jones, J. (1991). Earnings management during import relief investigations. Journal of Accounting Research 20(2), 193-228. Khan, M., & Watts, R. (2009). Estimation and empirical properties of a firm-year measure of accounting conservatism. Journal of Accounting and Economics 48, 132-150. Kothari, S. P. & Sloan, R. (1992). Information in prices about future earnings: Implications for earnings response coefficients. Journal of Accounting and Economics 15, 143–171. Kwon, S. S, Qin, Y. & Han, J. (2006). The effect of differential accounting conservatism on the “over-valuation” of high-tech firms relative to low-tech firms. Review of Quantitative Finance and Accounting 27, 143-173. Lemmon, M. L., Roberts, M. R. & Zender, J. F. (2008). Back to the beginning: persistence and the cross-section of corporate capital structure. Journal of Finance 63, 1575-1608. Patatoukas, P. N. & Thomas, J. (2009). Evidence of conditional conservatism: Fact or artifact? Working Paper. Patatoukas, P. & Thomas, J. (2010). More evidence of bias in the differential timeliness measure of conditional conservatism. The Accounting Review, Forthcoming. Penman, S. H. & Zhang, X. (2002). Accounting conservatism, the quality of earnings, and stock returns. The Accounting Review 77, 237-264. Ramirez, P. G. & Hachiya, T. (2006). Measuring firm-specific organizational capital and its impact on value and Productivity: Evidence from Japan. Review of Pacific Basin Financial Markets and Policies 9, 549-574. 23 Ryan, S. G. (2006). Identifying conditional conservatism. European Accounting Review 15, 511525. Ryan, S. G. & Zarowin, P. A. (2003). Why has the contemporaneous linear returns-earnings relation declined? The Accounting Review 78, 523-553. Sloan, R. G. (1996). Do stock prices fully reflect information in accruals and cash flow about future earnings? The Accounting Review 71, 289-316. Theil, H. (1961). Economic Forecast and Policy. Amsterdam: North-Holland. Vuolteenaho, T. (2002). What drives firm-level stock returns. Journal of Finance 57, 233–264. Watts, R. (2003). Conservatism in accounting, Part I: Explanations and implications. Accounting Horizons 17(3): 207-221. 24 Table 1 Full-sample volatility 0.105 Pooled and Cross-sectional Standard Deviations of Earnings and Returns Earnings Mean of cross-sectional volatilities 0.098 Mean of time-series volatilities 0.053 Full-sample volatility 0.392 Return Mean of cross-sectional volatilities 0.363 Mean of time-series volatilities 0.132 This table reports the full-sample standard deviations, the time-series means of cross-sectional standard deviations, and the cross-sectional means of time-series standard deviations of individual firms on earnings and stock returns. Earnings are computed as earnings per share (Compustat variable EPSPX) divided by lagged fiscal year-end closing price (Compustat variable PRCC_F), and returns are measured as the cumulative monthly returns for the period of 9 months before the fiscal year end to 3 months after the fiscal year end. Our constant sample consists of 691 firms from 1976 to 2005 from Compustat. A constant sample requires no missing observation in each of the sample year for each firm. 25 Table 2 Variance Decompositions of Returns-Earnings Relation Variable Firm FE ( j ) (a) 1.00 Year FE ( s ) Specification (c) (d) 0.53 (b) 1.00 (e) 0.82 0.47 (f) 0.44 0.47 Returns ( Rit ) 0.16 0.09 0.04 Dummy ( Dit ) 0.00 0.00 0.00 0.84 0.09 0.05 0.10 0.20 0.30 Returns*Dummy ( Dit Rit ) 2 0.13 Adjusted R 0.11 0.24 This table presents the variance decomposition for model specifications with or without firm/year fixed-effects. The regression equations are: N EPS it / Pit 1 = 0 j F jit + u it (a) j 2 T EPS it / Pit 1 = 0 sYsti + u it (b) s2 N T j 2 s2 EPSit / Pit 1 = 0 j F jit sYsti + u it (c) EPSit / Pit 1 = 0 1 Rit 2 Dit 3 ( Dit Rit ) + uit (d) N EPS it / Pit 1 = 0 1 Rit 2 Dit 3 ( Dit Rit ) j F jit u it (e) j 2 N T j 2 s 2 EPS it / Pit 1 = 0 1 Rit 2 Dit 3 ( Dit Rit ) j F jit s Ysti + uit where EPS = earnings per share, P = price, R = return, j (f) is the firm dummy, s is the year dummy, and Dit equals 1 if Rit< 0 and 0 otherwise. In each specification, we compute the type III partial sum of square for each variable and then normalize the square of each variable by the sum of the type III squares of the specification. Both firm and time effects are modeled as parametric shifts in the intercept terms of the regressions as in the LSDV framework. 26 Table 3 Extent of Accounting Conservatism Estimation Method Pooled OLS (1) LSDV/FE (2) 0 1 2 3 0.078 [64.01] 0.103 [5.87] 0.027 [10.93] 0.040 [16.29] 0.001 [0.42] 0.004 [1.59] 0.205 [25.25] 0.139 [16.94] R2 0.097 0.196 This table reports the estimation results from both the original and refined Basu’s model of accounting conservatism. Basu’s (1997) model is specified as: EPSit / Pit 1 = 0 1 Rit 2 Dit 3 ( Dit Rit ) uit The refined Basu model after accounting for firm-specific fixed-effects is specified as: N EPS it / Pit 1 = 0 1 Rit 2 Dit 3 ( Dit Rit ) j F jit u it j 2 In these specifications, EPS = earnings per share, P = price, R = return, Rit< 0 and 0 otherwise. In Pooled OLS, i is the firm dummy, and Dit equals 1 if i = 0 for all firms. LSDV/FE refers to least square dummy variable method with fixed-effects. R-square for the LSDV/FE model is calculated based on Theil (1961). The numbers in square brackets are the t-statistics. 27 Table 4 Trend in Accounting Conservatism Pooled OLS Sub-period 1976-1980 1981-1985 1986-1990 1991-1995 1996-2000 2001-2005 1 0.038 [6.91] 0.018 [3.29] 0.032 [4.39] 0.033 [5.58] 0.006 [1.28] -0.005 [-0.91] 3 ( 1 3 ) / 1 0.256 [8.39] 0.215 [10.84] 0.292 [16.44] 0.259 [10.98] 0.076 [6.46] 0.247 [13.27] LSDV/FE 7.74 Adj. R2 0.092 12.94 0.12 10.13 0.16 8.85 0.09 13.67 0.048 -48.40 0.088 3 1 0.051 [9.25] 0.021 [4.25] 0.055 [7.48] 0.051 [8.35] 0.008 [1.69] 0.023 [3.74] 0.132 [4.49] 0.072 [3.77] 0.147 [7.99] 0.089 [3.62] 0.033 [2.69] 0.065 [3.47] Pooled OLS By economic cycle 1976-1979 1980-1989 1990-2000 2001-2005 1 0.053 [7.72] 0.035 [8.97] 0.018 [5.00] -0.005 [-0.91] 3 ( 1 3 ) / 1 0.226 [6.98] 0.248 [16.98] 0.122 [11.58] 0.247 [13.27] ( 1 3 ) / 1 3.59 Theil’s R2 0.496 4.43 0.524 3.67 0.52 2.75 0.453 5.13 0.384 3.83 0.474 LSDV/FE 5.26 Adj. R2 0.105 8.09 0.139 7.78 0.065 -48.40 0.088 3 1 0.076 [10.82] 0.052 [13.50] 0.028 [7.55] 0.023 [3.74] ( 1 3 ) / 1 0.097 [3.10] 0.108 [7.44] 0.054 [5.05] 0.065 [3.47] 2.28 Theil’s R2 0.57 3.08 0.368 2.93 0.267 3.83 0.474 This table reports the estimation results on the trend in accounting conservatism over time using both the original and refined Basu’s model of accounting conservatism. The original Basu (1997)’s model is specified as: EPSit / Pit 1 = 0 1 Rit 2 Dit 3 ( Dit Rit ) uit The refined Basu’s model after accounting for firm-specific fixed-effects is specified as: N EPS it / Pit 1 = 0 1 Rit 2 Dit 3 ( Dit Rit ) j F jit u it j 2 where EPS = earnings per share, P = price, R = return, otherwise. In Pooled OLS, i is the firm dummy, and Dit equals 1 if Rit< 0 and 0 i = 0 for all firms. The numbers in square brackets are t-statistics. 28 Table 5 Estimating Conservatism with an Alternate Sample Allowing for New Listings and Delistings of Firms Panel A: Full-sample regression results Pooled OLS LSDV/FE 0 1 2 3 0.033 [22.31] 0.004 [0.05] -0.002 [-0.84] 0.023 [0.82] 0.023 [8.91] 0.028 [10.71] 0.457 [69.26] 0.281 [39.93] R2 0.073 0.244 Panel B: Regression results by sub-periods Pooled OLS Sub-period 1976-1980 1981-1985 1986-1990 1991-1995 1996-2000 2001-2005 LSDV/FE 1 3 ( 1 3 ) / 1 0.036 [9.18] 0.020 [4.57] 0.001 [0.11] -0.003 [-0.70] -0.017 [-4.29] -0.032 [-6.86] 0.446 [21.13] 0.344 [21.87] 0.468 [31.23] 0.502 [31.23] 0.361 [29.42] 0.426 [26.00] 13.39 Adj. R2 0.099 18.20 0.078 469.00 0.087 -166.33 0.074 -20.24 0.053 -12.31 0.065 1 0.053 [13.50] 0.032 [7.03] 0.025 [3.99] 0.027 [6.01] -0.001 [-0.21] 0.010 [2.10] 3 0.237 [11.24] 0.087 [12.62] 0.206 [12.62] 0.174 [9.99] 0.150 [10.99] 0.036 [2.04] ( 1 3 ) / 1 5.47 Theil’s R2 0.486 3.72 0.414 9.24 0.455 7.44 0.447 -149.00 0.387 4.60 0.424 This table reports the estimation results of accounting conservatism using both the original Basu model (Pooled OLS) and the refined model of fixed-effects (LSDV/FE) using an alternate sample that allows for firm new listings and delistings. We require that each firm has at least 10 observations in our sample period of 30 years. This sample covers 6,041 firms, with an average cross section of 3,623 firms. In Panel B, we further require that firms have three non-missing observations in each of the five-year sub-periods. The numbers in square brackets are t-statistics. 29 Table 6 The Returns-Cash Flows Relation Pooled OLS Sub-period 1988-1993 1994-1999 2000-2005 LSDV/FE 1 3 1 3 0.035 0.141 0.050 0.070 [5.85] [6.26] [9.00] [3.20] 0.005 [1.08] 0.125 [7.01] 0.036 [8.45] 0.011 [0.69] 0.050 [9.19] 0.053 [2.86] 0.050 [10.10] -0.010 [-0.59] Notes: The regression equation is: N CFOit / Pit 1 = 0 1 Rit 2 Dit 3 ( Dit Rit ) j F jit + uit , j 2 where CFO = cash flows from operations per share, P = price, R = return, Rit< 0 and 0 otherwise. In Pooled OLS, i = 0 for all firms. i is the firm dummy, and Dit equals 1 if The CFO variable is only available after 1988. The constant sample is between 1988 and 2005 and consists of 995 firms in total. The numbers in square brackets are the t-statistics. 30 Table 7 The Implied Effect of Accruals Pooled OLS, 3 LSDV/FE, 3 Earnings Cash Flows Earning Cash Flows regression regression Regression Regression Column Column Column Column Column Column Sub-period (1) (2) (1)-(2) (1) (2) (1)-(2) 1988-1993 0.441 0.141 0.300 0.043 0.070 -0.026 [9.35] [6.26] [5.73] [0.79] [3.20] [0.45] 0.188 0.125 0.062 0.062 0.011 0.051 [9.31] [7.01] [2.32] [2.70] [0.69] [1.83] 0.357 0.053 0.303 -0.009 -0.010 0.001 [4.15] [2.86] [3.45] [-0.10] [-0.59] [0.07] 1994-1999 2000-2006 Notes: For the earnings regression, the regression equation is: N EPSit / Pit 1 = 0 1 Rit 2 Dit 3 ( Dit Rit ) j F jit uit , j 2 and for the cash flows regression, the equation is: N CFOit / Pit 1 = 0 1 Rit 2 Dit 3 ( Dit Rit ) j F jit + uit , j 2 where EPS = earnings per share, CFO = cash flows from operations per share, P = price, R = return, dummy, and Dit equals 1 if Rit< 0 and 0 otherwise. In Pooled OLS, i = i is the firm 0 for all firms. The estimate of is reported in the table. The constant sample is between 1988 and 2005 and consists of 995 firms in total. The numbers in square brackets are the t-statistics. 31 Table 8 Estimation Results for the Extended Basu model with Lagged Returns Panel A: Full sample, 1979-2005 Method Pooled OLS LSDV/FE 0 10 11 12 13 20 21 22 23 30 31 32 33 30 31 32 33 0.08 [40.75] 0.09 [5.64] 0.04 [17.87] 0.05 [19.71] 0.02 [9.17] 0.03 [10.95] 0.01 [4.62] 0.02 [6.41] 0.01 [3.39] 0.01 [5.39] 0.01 [2.35] 0.01 [3.04] 0.01 [6.05] 0.01 [6.79] 0.01 [4.35] 0.01 [5.20] 0.01 [2.33] 0.01 [3.32] 0.15 [19.01] 0.12 [16.08] 0.21 [27.02] 0.19 [24.43] 0.13 [17.04] 0.11 [14.50] 0.1 [12.36] 0.08 [9.69] 0.59 {1907.8} 0.5 {1096.9} Panel B: By economic cycle, 1980-2005 1980-1989 Method Pooled OLS LSDV/FE 1990-2000 Pooled OLS LSDV/FE 2001-2005 Pooled OLS LSDV/FE 0 10 11 12 13 20 21 22 23 30 31 32 33 30 31 32 33 0.1 [33.17] 0.15 [5.67] 0.06 [21.67] 0.06 [[2.74] 0.06 [11.38] 0.04 [0.93] 0.05 [12.22] 0.05 [13.81] 0.03 [8.34] 0.04 [9.71] 0.05 [8.05] 0.06 [9.08] 0.01 [4.05] 0.02 [5.24] 0.02 [4.83] 0.02 [6.22] 0.03 [5.44] 0.04 [6.78] 0.01 [2.10] 0.01 [3.41] 0.01 [3.69] 0.02 [4.36] 0.01 [1.19] 0.02 [2.50] -0.01 [-1.90] 0 [0.09] 0.02 [4.56] 0.02 [4.96] 0.01 [2.02] 0.01 [1.18] 0 [-0.23] 0.01 [1.84] 0 [0.25] 0 [-0.95] 0.01 [1.52] 0.01 [1.76] 0.01 [1.77] 0.01 [2.99] 0.01 [4.35] 0.01 [3.22] 0.02 [4.15] 0.03 [4.66] 0 [1.15] 0.01 [2.07] 0.01 [4.15] 0.01 [3.51] 0.01 [1.36] 0.01 [1.56] 0 [-0.52] 0 [0.72] 0.01 [3.55] 0.01 [3.16] 0.01 [1.85] 0.01 [1.44] 0.19 [13.69] 0.14 [9.79] 0.1 [9.68] 0.06 [6.20] 0.16 [8.90] 0.06 [3.34] 0.23 [16.24] 0.18 [12.34] 0.14 [12.91] 0.11 [9.60] 0.25 [15.11] 0.17 [9.06] 0.14 [9.72] 0.09 [6.25] 0.1 [8.97] 0.07 [6.25] 0.14 [8.53] 0.05 [2.97] 0.11 [7.43] 0.05 [3.42] 0.08 [6.37] 0.05 [3.75] 0.07 [4.68] 0.01 [0.76] 0.67 {703.8} 0.46 {190.9} 0.42 {467.3} 0.29 {141.3} 0.62 {490.6} 0.29 {39.6} Notes: The regression equation is: EPSit / Pit 1 = 0 10 Rit 11Rit 1 12 Rit 2 13 Rit 3 20 Dit 21Dit 1 22 Dit 2 23 Dit 3 (Adj.) R2 0.2155 0.2892 (Adj.) R2 0.2572 0.412 0.145 0.3119 0.2452 0.5238 N 30 (Dit Rit ) 31 ( Dit 1 Rit 1 ) 32 ( Dit 2 Rit 2 ) 33 (Dit 3 Rit 3 ) i Fijt + uit , where EPS = earnings per share, P = price, R = return, i is the firm j 2 dummy, and Dit (Dit-1 /Dit-2/Dit-3) equals 1 if Rit, (Rit-1 /Rit-2 /Rit-3) is smaller than 0 and 0 otherwise. The data are from 1979 to 2005. In Pooled OLS, i = 0 for all firms. In the sub-period analysis, we drop year 1979 and report only three business cycles. Numbers in square brackets are t-statistics, and numbers in curly brackets are chi-statistics from the Wald tests. Table 9 Estimation Results from Alternative Models of Accounting Conservatism Panel A: Proxy for economic loss, the Level of Cash flow (CFO) < 0 CFO Model Pooled OLS Intercept CFOt LSDV/FE DD Model Pooled OLS Jones Model LSDV/FE -0.058 -0.009 -0.069 0.008 0.042 [-4.85] [-3.27] [-8.60] [-3.95] [6.83] [2.47] -0.477 -0.584 -0.558 -0.611 -0.471 -0.581 [-55.97] [-56.31] [-63.59] [-59.05] [-57.36] [-59.21] 0.067 0.057 [20.18] [16.84] 0.063 0.050 [18.05] [14.21] 0.095 0.099 [41.41] [43.82] -0.034 -0.068 [-27.73] [-24.25] CFOt+1 ∆REVt GPPEt DtCFOt R-squared (%) F-test for FE LSDV/FE -0.005 CFOt-1 Dt Pooled OLS 0.031 0.028 0.024 0.026 0.025 0.027 [14.33] [12.70] [11.52] [11.66] [12.15] [12.99] 0.735 0.655 0.627 0.600 0.714 0.610 [59.01] [41.52] [49.15] [37.84] [60.25] [40.90] 18.15 35.4 21.54 36.97 27.15 42.86 4.27 3.90 4.40 Table 9 cont’d Panel B: Proxy for economic loss, change in cash flow (∆CFO) < 0 DD Model CFO Model Intercept CFOt Pooled OLS LSDV/FE Pooled OLS -0.042 -0.104 -0.044 [-40.79] [-5.76] -0.176 [-33.39] Jones Model Pooled OLS LSDV/FE -0.110 -0.037 -0.025 [-43.98] [-6.13] [-31.80] [-1.36] -0.298 -0.334 -0.379 [-41.63] [-47.67] [-47.88] 0.088 0.064 [23.54] [16.70] 0.080 0.059 [21.68] [16.17] 0.089 0.098 [36.69] [40.48] -0.052 -0.087 [-42.20] [-29.25] CFOt-1 CFOt+1 LSDV/FE ∆REVt GPPEt ∆CFOit Dt Dt∆CFOt R-squared (%) -0.122 -0.060 -0.116 -0.075 [-25.20] [-12.46] [-25.34] [-15.88] 0.026 0.022 0.020 0.018 0.033 0.034 [21.01] [19.02] [16.39] [15.47] [28.94] [31.34] 0.199 0.104 0.071 0.047 0.119 0.032 [26.23] [13.14] [12.23] [8.18] [17.54] [4.16] 11.23 31.8 14.25 33.21 19.07 34.6 F-test for FE 4.83 4.52 3.80 In Panel A, the regression equation for the CFO model is: N Ait 0 1CFOit 2 Dit 3 ( Dit CFOit ) j F jit uit j 2 The equation for the DD model is: N Ait 0 1CFOit 2 Dit 3 ( Dit CFOit ) 1CFOit 1 2CFOit 1 j F jit uit j 2 And the equation for the Jones model is: N Ait 0 1CFOit 2 Dit 3 ( Dit CFOit ) 1REVit 2GPPEit j F jit uit j 2 In the above equations, A = accruals, CFO = cash flow from operations, REV is the change in revenue, GPPE is the gross property, plant, and equipment, and Dit equals 1 if CFOit< 0 and 0 otherwise. All these variables are scaled by lagged total assets. In Pooled OLS, i = 0 for all firms. In Panel B, CFO is replaced by the change in CFO (∆CFO), and Dit= 1 if ∆CFOit< 0 and 0 otherwise. The numbers in square brackets are the t-statistics. 34 Figure 1 The Ratios of Cross-sectional Volatility over Pooled Volatility for Earnings and Returns over Time 35 Appendix: Replication of Givoly and Hayn’s (2000) Results Based on Basu’s Measure and Discussion on Firm Composition of Full Sample Using pooled OLS regressions and the full sample of firms in the intersection of Compustat and CRSP from 1950 to 1998, Givoly and Hayn (2000)—hereafter “GH”—show that the asymmetric response of earnings to return using Basu’s (1997) measure has increased over time. As a reality check, Table A-1 presents Basu’s (1997) results for five-year sub-periods, using GH’s full sample derived from the intersection of Compustat and CRSP from 1963 to 2004 and estimation method. [Table A-1 here.] A general observation from the Table is that the asymmetric response of earnings to return ( 3 ) shows an increasing trend during the sample period studied in GH. This finding is consistent with GH. However, in the last two sub-periods of the sample (1995-2004), the asymmetric response is declining. Thus, the statement of an increase in the asymmetric response of earnings to return appears to be sample-specific, in addition to being estimation-method specific. In showing the increase in accounting conservatism, GH use not only Basu’s (1997) measure, but also accruals level to proxy for conservatism. GH present the accruals level results using a constant sample. Part of the reason for using a constant sample is the changing composition of firms in the full sample during the sample period, which is shown in Figure A-1. [Figure A-1 here.] From Figure A-1, we observe that the number of firms included in the sample varies substantially over time: the sample started with 803 firms in 1963, peaked to 5,570 firms in 1997, and dropped to 3,397 firms in 2005. A comparison of Table A-1 with Table 4, where the trend in the asymmetric response of earnings to return is shown for a constant sample, reveals that much of the increase in the asymmetric response documented in GH is attributed to new 36 listings. Given the degree of the changing composition of firms, we believe that focusing on a constant sample in studying the returns-earnings relation retains a consistent sample and, therefore, helps us to address the question of whether the returns-earnings relation has changed over time for the same firms. Therefore, we rely on a constant sample most of the time in this paper. 37 Table A-1 Replication of Givoly and Hayn (2000) Sub-period 0 1 2 3 Adj. R2 1963-1964 0.072 -0.004 0.023 0.137 0.144 1965-1969 0.068 -0.003 0.016 0.067 0.172 1970-1974 0.094 0.007 0.033 0.105 0.081 1975-1979 0.144 -0.017 0.057 0.346 0.123 1980-1984 0.073 -0.021 0.008 0.250 0.109 1985-1989 0.036 -0.004 -0.009 0.407 0.142 1990-1994 0.014 0.014 -0.015 0.449 0.083 1995-1999 0.021 0.007 -0.021 0.293 0.086 2000-2004 0.000 0.005 -0.033 0.363 0.057 Notes: The regression equation is: EPSit / Pit 1 0 1 Rit 2 Dit 3 ( Dit Rit ) uit , where EPS = earnings per share, P = price, R = return, and Dit equals 1 if Rit< 0 and 0 otherwise. We strictly follow Givoly and Hayn (2000) to derive a sample of firms from the intersection of Compustat and CRSP from 1963 to 2004. The numbers in square brackets are the t-statistics. 38 Figure A-1 The Number of Firms over Time in the Full Sample 39
© Copyright 2026 Paperzz