«Leif Atle Beisland University of Agder Dissertation submitted to the Department of Accounting, Auditing and Law at the Norwegian School of Economics ...»
Specification (1) implicitly assumes that there are no other effects on the book response coefficient BRC than the reporting regime IFRS. In order to control for other effects, (2) could be extended to (3) BRC = α1 + α4 · IFRS + α5 · CONT, Some may claim that stock prices do not respond to book values of equity. Rather, they respond to the value creation as measured by earnings. Thus, the concept of "book value response coefficient" could be replaced by "book value association coefficient". However, because the latter term does not appear to be standard in the literature, we choose to apply BRC throughout this paper (see e.g., Ghosh, Zhaoyang and Jain, 2005).
The regression coefficients are usually estimated through ordinary least squares OLS. However, the standard deviation of the coefficients, which are important when calculating t- and p-values, should be adjusted for heteroskedasticity and autocorrelation (HAC) – at least if tests show the presence of such empirical problems; see White (1980) and Newey and West (1987). If severe HAC is detected, feasible GLS should be considered, at least as a robustness test.
IFRS.15 In this way, other effects between the sample of firms reporting according to IFRS and those reporting according to NGAAP could be controlled for, including earnings per share EARN’; see Easton and Harris (1991). Another factor that should be controlled for is the intangible asset intensity of the two types of firms; see e.g. Lev and Zarowin (1999). Such a control would rule out that the difference in BRC is driven by difference in the intangible asset intensity between those who are reporting according to NGAAP and those who are reporting according to IFRS. Other control variables are presented after we have introduced the test methodology of Hypothesis 2. A supplementary approach to increase the level of control for other differences could be to focus on a constant sample of firms across reporting regime.
Hypothesis 2 could be tested by extending (1) to a price regression which includes both the book value per share BOOK and the earnings per share EARN’ as explanatory variables,16 or by running a regression of price on earnings alone. However, we will not test Hypothesis 2 in these ways (other than as a robustness test to (4) - (6) below). The reason is that price regressions suffer by scale problems in which the relation between two variables could be driven by different underlying scales – and not a ‘causal’ relationship between them; see Barth and Kallapur (1996), Brown, Lo and Lys (1999) and Easton and Sommers (2003).
The corresponding price regression consistent with (3) becomes PRICE = α0 + α1 · BOOK + α2 · IFRS + α3 · CONT + α4 · BOOK · IFRS + α5 · BOOK · CONT + ε. The regression could also be extended by including the interaction term BOOK · IFRS · CONT, but we choose not to do that in order to limit potential multicollinearity problems.
Notice that the period’s earnings per share EARN’ is included indirectly in the specified price regression (1) through the book value of equity per share BOOK. If EARN’ is a separate variable, the book value BOOK should be adjusted to BOOK’ = BOOK - EARN’ in order not to double account for earnings and thereby understate the earnings response coefficient. This adjustment is not considered necessary if EARN’ functions as a control and not as a test variable.
ing them by some scale measure, typically the previous period’s stock price. Thus, the remedy is to focus on the return regression.17 But since the book value of equity does not enter the return regression, we are stuck with (1) to analyze the response coefficient of the book value and hence the balance sheet.
The return regression is
where RET is the stock market return of firm i in period t, and EARN is the period’s earnings deflated by the previous period’s stock price, i.e. EARN = EARN’/PRICEt-1.18, 19 The betas are the regression coefficients and ε is the error term. The earnings response coefficient, i.e.
∂ RET/ ∂ EARN, consistent with (4) equals
The earnings response coefficient ERC equals the ‘core’ coefficient β1, i.e. the coefficient of NGAAP, plus a term β4 · IFRS depending on the dummy variable IFRS, indicating this reporting regime. Our focus is on the coefficient β4. If β4 is significantly different from zero, the The return regression also suffers from some scale problems, though less than the price regression. The scale of the return is the expected return – and this scale factor could be removed from the analysis by focusing on abnormal stock return. We indirectly do this through (6), since control variables related to risk make the residual equal to abnormal return.
The change in earnings ∆EARN is included indirectly in the specified return regression (4) through the level variable EARN. If ∆EARN is a separate variable as e.g. in Easton and Harris (1991), the level variable EARN should be adjusted to LEARN = EARN - ∆EARN in order not to double account for the change in earnings and thereby understate its response coefficient.
The regression model (4) could also be estimated on an excess stock return basis, in which excess return is return minus a proxy for the risk free rate of return. Thus, the excess return variable replaces the plain return variable in (4).
that IFRS moderates the ERC. If β4 0, stock market investors respond less to reported earnings when prepared according to IFRS than NGAAP.
Prior research has shown that several company specific characteristics may affect the value
relevance of accounting information, for instance measured by the earnings response coefficient given by (5). It is therefore important to control for these factors before making statistical inferences about whether and how the ERC is affected by reporting regime:
In (6) CONT is a vector of control variables possibly affecting ERC, in addition to the indicator variable IFRS.20 The vector of control variables CONT could be (BETA, SIZE, BTM, MOM; LOSS, INTAN, TRAN). We will now present each of these seven control variables.21 The first set of control variables is various risk proxies – systematic as well as firm specific.
When stock market returns are explained solely by various risk variables, the residuals become abnormal returns. When other variables, e.g. earnings, enter the return regression, they The corresponding return regression consistent with ERC given by (6) equals RET = β0 + β1 · EARN + β2 · IFRS + β3 · CONT + β4 · EARN · IFRS + β5 · EARN · CONT + ε. The return regression could also be extended by including the interaction term EARN · IFRS · CONT, but we choose not to do that in order to limit potential multicollinearity problems. Another extension is to control directly for changes in earnings ∆EARN; compare Easton and Harris (1991).
We cannot rule out the possibility that our findings are attributable to changes in the economic environment rather than to changes in the financial reporting system. Our research design that includes control variables is constructed to mitigate the effects of the former. Obviously, we cannot disregard the possibility that there are more relevant omitted variables than the ones we have controlled for. For instance, the analysis could have been extended with a variable for investor sentiment; see e.g. Baker and Wurgler (2006). However, we have included the variables that prior research has found to affect value relevance. Even if investor sentiment is related to stock return, it would (by definition) be reasonable to assume that the variable would be unrelated to accounting fundamentals. Thus, the variable would leave regression coefficients on accounting values unaffected. A further control for differences in underlying economic conditions between IFRS and NGAAP is to test Hypothesis 1 and 2 on a sample consisting of exactly the same firms. This we will do as a robustness test.
stock market risk as in the Capital Asset Pricing Model. According to Fama and French (1992), firm size SIZE and the book-to-market ratio BTM are found to be relevant proxy risk
tential proxy risk factor if returns exhibit serial correlation; see Carhart (1997). In addition to entering directly into the return regression, these four risk factors might also influence the earnings response coefficient ERC, as suggested by (6).
The second set of control variables represents factors potentially influencing the informational content of earnings. Hayn (1995) finds that the response coefficient of negative earnings is far less than the response coefficient of positive earnings; see also Basu (1997). A dummy variable for losses LOSS should therefore enter (6) as a moderator variable for the ERC. Lev and Zarowin (1999) claim that the lack of intangible asset capitalization has been detrimental to the value relevance of financial reporting. Earnings become less informative for investors because expenditures on intangibles are not treated as investment expenditures and not matched with future revenues – and the periodic expenditures might create transitory noise.
The variable INTAN, which measures the degree of intangible asset intensity, should therefore be a control variable in (6). Finally, the findings of e.g. Elliot and Hanna (1996) suggest that transitory or non-recurring earnings are less value relevant than permanent or recurring earnings. The degree of transitory earnings TRAN should therefore be included in (6) as a control variable. Our empirical measures of the control variables are presented in detail in subsection 3.2.
Fama and French (1992) do not suggest that for instance SIZE and BTM are risk factors themselves. Instead, these variables may proxy for some underlying (unobservable) risk factors. Hence, the term proxy risk factor.
It is important to adjust the book-to-market ratio so that the book value does not contain the periods’ earnings;
the earnings variable is a separate variable in the return regression and should not be double accounted for. Furthermore, earnings are not known ex ante and are thereby not a risk factor, but a factor potentially contributing to explain excess return.
parts with control variables is that extensive use of interaction effects might lead to multicollinearity problems. Collinearity means that two explanatory variables are correlated, which in itself is no violation of the assumptions behind the regression model – only perfect collinearity is; see e.g. Wooldridge (2008, pp. 95-99). However, if some explanatory variables are highly correlated, their coefficient loadings might become somewhat ‘arbitrary’, which creates problems when evaluating the statistical significance of the regression coefficient of a test variable which is highly collinear with another (control) variable. Collinearity or multicollinearity between control variables CONT is no problem for the statistical inference of the emphasized test variable, which in our case is the interaction with the reporting regime IFRS.24 To evaluate the multicollinearity we focus on the condition number, i.e. the largest condition index. According to Belsley, Kuh and Welsch (1980), there is no multicollinearity problem if the condition number is below 20. A condition number between 20 and 30 indicates some multicollinearity problems, and the problem becomes severe if it is above 30. If a problem is detected, the variance inflation factors, VIF, could be used to determine whether the test variable is involved – or the problem is solely among control variables. No statistical inference results should be emphasized in which the test variable is severely affected by multicollinearity problems, i.e. their VIF should be less than 10. A potential remedy for severe multicollinearity is, however, to alter the empirical specification, e.g. by removing and adding control variables, to analyze coefficient stability. Because OLS standard deviations are inflated by their VIFs, within sample arbitrariness could be better captured by employing bootstrapped standard deviations when making statistical inferences (see subsection 5.1 for more details).
A problem with the control variables LOSS and TRAN is that they cannot be observed directly – and hence has to be estimated on the basis of accounting information. This means that they to some extent are becoming collinear with the test variable, the accounting variable IFRS, which might lead to build in collinearity. This design problem is difficult to circumvent.
statistics about the distributional properties of the sample and analyzes simple binary correlation between these variables. The regression results are presented and discussed in the next section.
3.1 Data and Sample Description We have collected market and accounting data for all firms listed on the Oslo Stock Exchange OSE from 2002 to 2006. Market and accounting data are measured at the end of the accounting year.25 Firms not reporting according to IFRS or NGAAP are excluded from the sample, mainly firms reporting according to USGAAP. Since the firms were required to report according to IFRS from 2005, IFRS was the prevailing accounting regime in 2005 and 2006.26 In order to study relatively equal sample sizes of IFRS and NGAAP observations, we also include two years of NGAAP financial statements, i.e. 2003 and 2004. The total number of observations in our sample is equal to 741, of which 391 are IFRS observations while 350 are NGAAP observations; see Panel A of Table 1. We do not have the previous year’s market values of firms that have been listed in 2003 - 2006. Therefore, the number of observations is lower in the return specification than in the price specification. The number of observations is reduced to 651.
We have considered the inefficiency-adjustment procedure proposed by Aboody, Hughes and Liu (2002) to account for possible slow stock market adjustment to disclosed accounting information. But since Gjerde, Knivsflå and Sættem (2008) do not find any significant differences in results due to the procedure, we continue without employing it.