Stale and Scale Effects in Markets-Based Accounting Research: Evidence from the Valuation of Dividends

Abstract

This study revisits prior research on the valuation of dividends in an accounting-based valuation framework. Using a battery of tests, we show that market value deflation is essential in market-based tests of dividend displacement and signalling because it controls for ‘stale’ information in addition to scale (size) differences across firms. For US firms, we show that after controlling for ‘stale’ information, the empirical association between dividends and market values switches from positive to negative. This switch is not explained by scale differences across firms. Further, we show that after controlling for staleness, the valuation of dividends remains positive for European firms. This result is explained by the relatively stronger association of dividends with future earnings in these settings (i.e. signalling). Lastly, our country-specific estimates of dividend valuation provide a potentially valuable index for studies aimed at examining the effects of accounting and securities regulation on information asymmetries in an international context.

JEL Classification:

• M41
• C10
• G12
• G14

View correction statement:

Erratum

Notes

^†Supplemental materials are available in an online Appendix at: http://dx.doi.org/10.1080/09638180.2013.795870

1 See Rees (1997), Fama and French (1998), Giner and Rees (1999), Lo and Lys (2000), Akbar and Stark (2003), Hand and Landsman (2005), Rees (2005), Pinkowitz et al. (2006), Dittmar and Mahrt-Smith (2007), Dedman et al. (2009), Shah et al. (2009), Dedman et al. (2010, 2012), and Rees and Valentincic (2012).

2 See Christie (1987), Landsman and Magliolo (1988), Kothari and Zimmerman (1995), Barth and Clinch (1998), Easton (1998, 1999), Brown et al. (1999), Lo and Lys (2000), Easton and Sommers (2003), Gu (2005), Lo (2005), and Barth and Clinch (2009).

3 Although stock splits do not affect total market value of a firm, Gu (2005) provides an analogous argument with regard to firm-level (undeflated) analyses. Gu (2005, p. 71) argues that ‘ … [o]ne can view the observed variations in firm sizes as firms’ choices about growth, mergers and acquisitions, and divestitures, which are similar choices about the number of shares, stock splits, and dividends’. The online supplemental material provides results of a simulation showing how stock splits and the resulting scale bias can affect the relationship between dividends and stock prices.

4 While we focus on coefficient estimates, scale effects can also lead to spurious inferences when the researcher is interested in assessing (variation in) the explanatory power of Equation (3). Brown et al. (1999) show that scale effects in price levels regressions bias explanatory power (R²) upwards and that prior evidence on the increase in value relevance of accounting information (as measured by R²) is spuriously driven by time variation in scale.

5 See Rees (1997), Fama and French (1998), Giner and Rees (1999), Lo and Lys (2000), Akbar and Stark (2003), Hand and Landsman (2005), Rees (2005), Pinkowitz et al. (2006), Dittmar and Mahrt-Smith (2007), Dedman et al. (2009), Shah et al. (2009), Dedman et al. (2010, 2012), and Rees and Valentincic (2012).

6 The concern is only with the coefficient on cash dividends, while the coefficient of NETCAP (or special dividends) is frequently found to be close to displacement (Dedman et al., 2010; Rees and Valentincic, 2012).

7 For example, Hand and Landsman (2005) propose that the positive coefficient of dividends is induced by mispricing of accounting earnings and/or book value of equity. Pinkowitz et al. (2006) suggest that agency costs of free cash flows explain why firms that return free cash flow to shareholders in the form of dividends may enjoy a higher valuation.

8 See Fama and Babiak (1968) for empirical modelling and validation, Jagannathan et al. (2000) and Brav et al. (2005) for recent application in the US, and Goergen et al. (2005) for application outside of the US market.

9 Note that the original (empirical) Lintner model includes an error term that captures dividend determinants not systematically reflected by the other factors. Assuming the expected value of the error term equals zero, we find that the inferences gained from our modelling are not affected by the inclusion of this error term. Thus, we ignore the error term without loss of generality.

10 In Section 4.2, we provide empirical evidence on the descriptive validity of our model and assumptions.

11 Because in our sample the lower tails of market value of equity, book value of equity, and dividends are restricted by zero, we only eliminate extreme observations in the upper 1% of the distribution. These variables cannot take on extreme negative values that could spuriously influence our estimations. Eliminating the bottom 1% for those variables would therefore unnecessarily eliminate relevant observations. We replicate our main analyses after also eliminating the smallest observations or after winsorizing the test variables. Results are qualitatively unchanged. As the net capital contributions variable equals zero in a disproportionate part of the sample, we identify extreme observations in the lower and upper one percentiles of the distribution by ignoring zero values. Not doing so would result in the elimination of potentially important cross-sectional variation in this variable.

12 Untabulated results discussed in this section are not materially affected by the use and choice of deflator. Because we are partly interested in explanatory powers in this section and because explanatory powers are affected by scale (Brown et al. 1999), we do believe it is important to present estimation results based on variables deflated by a proxy for scale.

13 Adding the current earnings term which is part of our model in Equation (8) has limited consequences for the explanatory power (adjusted R² increases to 0.384).

14 In contrast, the residual value of dividends (i.e. the component of dividends not explained by past earnings realizations) hardly explains variation in current and lagged market values (adjusted R² of 0.0014 and 0.0010, respectively).

15 Petersen (2009) provides guidance on the correction of standard errors for correlation across observations in finance panel data sets. He distinguishes between a firm effect (correlation across years for a given firm, or time-series dependence) and a time effect (correlation across firms for a given year, or cross-sectional dependence), and finds that clustering standard errors by firm or over time produces unbiased standard errors when the number of clusters is sufficiently large. Gow et al. (2010), however, show that clustering standard errors on both firm and time is warranted in the accounting literature, since many accounting variables exhibit great dependence over time and across firms. In our case, untabulated tests reveal that the method of standard error correction has no qualitative impact on the inferences drawn.

16 An analysis of variance inflation factors (VIFs) reveals that the mechanical interdependencies between the independent variable do not lead to multicollinearity problems. The maximum VIF equals 5.17 for the undeflated regressions, while for the deflated regressions the maximum observed VIF equals 2.33. Lagged market value deflation leads to the lowest average VIFs. All VIFs are substantially smaller than the critical value of 10.

17 As explained by Easton and Sommers (2003), simply removing observations in the top scale group with significant average studentised residuals does not solve the problem. After eliminating observations in scale group 40, the next-largest observations in group 39 automatically become the dominant and influential observations. Untabulated analyses reveal that if we eliminate observations in scale group 40 for the undeflated analysis, the mean studentised residual in group 39 jumps from 0.31 to 1.94. The mean studentised residual in group 39 jumps from 1.17 to 2.08 if we do the same for the per-share analyses. Hence, consistent with Easton and Sommers (2003), we find that the scale effect is different from the well known effect of outliers on regression estimates.

18 Because the theoretical valuation framework in Equation (1) does not have an intercept term, the unscaled intercept is added in the estimation after all variables are deflated (Barth and Kallapur, 1996; Lo and Lys, 2000). This approach is generally similar to Brown et al. (1999) and Kothari et al. (2005) and is expected to produce econometrically consistent estimates. In our case, the intercept is composed of unscaled year dummies.

19 Studies such as Rees and Valentincic (2012) use the current book value of equity as deflator in price-levels regressions. In contrast, we use the lagged (year t–1) value of book value because (i) earnings and dividends affect closing book value, thereby inducing a mechanical correlation between the independent variables and the deflator, (ii) the purpose of scaling is to ensure that firms are comparable at the start of the window over which the variable of interest (in this case, dividends) is measured, and (iii) scaling book value by itself results in an additional intercept term (a column of ones), thereby making it less straightforward to evaluate the regression results across different deflators. Nevertheless, untabulated test results for the coefficient on the dividends variables are qualitatively similar when using the current instead of lagged value of book equity as a deflator.

20 The differential effect of deflation on dividends versus net capital contributions can be explained by the fact that dividends exhibit higher correlation with past earnings than net capital contributions. Indeed, while the Spearman correlation between asset-deflated dividends and lagged earnings is high (untabulated ρ = 0.586), the correlation between lagged earnings and net capital contributions is weak (ρ = 0.048). At the same time, the untabulated Spearman correlation between lagged price and current dividends (deflated by lagged assets) is high and equals 0.638, while the correlation between net capital contributions and lagged price is weak (ρ = 0.025). These figures further support the argument that a failure to control for stale information explains mixed findings based on deflation by accounting (total assets or book equity) versus market proxies for scale.

21 Using Equation (1), the historical cost of capital (r = 12%), and a residual income persistence parameter ω = 0.62 (Dechow et al., 1999), the expected coefficient on earnings equals 1.39. This theoretical value is most consistent with the price-deflated estimation, where the earnings coefficient equals 2.16. On the other hand, the coefficient on book equity is substantially understated based on price-deflation compared to the coefficients obtained using other deflators and a theoretical value of 0.85. We argue this is likely the result of the fact that market value deflation purges out relevant information from the ‘stock’ variable BVE. This is a drawback of market value deflation when one is interested in the association between market values and stock variables such as book equity.

22 Note that the coefficient on net capital contributions is approximately one unit higher in the per-share regressions compared to the asset-deflated regressions. This is because net capital contributions do not affect value per share.

23 Our previous analyses are based on the assumption that the valuation model in Equation (1) is empirically valid. However, this might not hold for some firms whose characteristics do not fit the assumptions of the Ohlson (1995) model. Specifically, the Ohlson model is not well specified for firms that have abnormal performance (that is, no mean reversion of residual income), when accounting is permanently conservative, or for firms that report (large) dirty surplus adjustments (Myers, 1999; Lo and Lys, 2000). Consequently, results for our market value-deflated valuation regressions may be driven by the deflation exacerbating model misspecification problems and biasing the dividend valuation coefficient (e.g. to a negative value as observed). In the online supplemental material, we show that our inferences are not affected by a subset of observations that are expected to violate the model's assumptions.

24 COMPUSTAT GLOBAL coverage of UK firms is very limited prior to 1989. Because we require lagged (t–1) data, the sample period starts in 1990.

25 We have ensured that all variables are denoted in terms of the countries' currencies in place on 1 January 2012 (either Euro or local currency).

26 A potential explanation for the variation in dividend valuation coefficients across countries is that the Ohlson (1995) model assumptions are differentially violated across countries and we therefore simply observe cross-country differences in the extent of bias. To alleviate this concern, we compute the absolute value of a firm's dirty surplus items (change in equity adjusted for earnings, dividends and net capital transactions) and generate three country-level measures that capture importance of dirty surplus flow: (i) the percentage of firms with dirty surplus items greater than 1% of total assets, (ii) the percentage of firms with dirty surplus items greater than 5% of total assets, and (iii) the median absolute value of dirty surplus items in the country. We find no statistically significant association between our 15 country-specific dividend valuation coefficients and each of the three country-level measures of dirty surplus accounting. This analysis is tabulated in the online supplemental material.

27 As can be seen in the appendix, deflating the variables in the earnings forecast equation by book value of equity (as in Rees and Valentincic, 2012) instead of assets produces an even stronger relation between dividend valuation and the predictive ability of dividends for future earnings (adjusted R² of 0.64 versus 0.45).

28 The online supplemental material lists country-specific valuation coefficients for a broader sample of 43 countries.