# «Leif Atle Beisland University of Agder Dissertation submitted to the Department of Accounting, Auditing and Law at the Norwegian School of Economics ...»

The regression model is PRICE = α0 · IND + α1 · BOOK + α2 · IFRS + α31 · EARN’ + α32 · LOSS + α33 · INTAN + α34 · TRAN + α35 · BETA + α36 · SIZE + α4 · BOOK · IFRS + α51 · BOOK · EARN’ + α52 · BOOK · LOSS + α53 · BOOK · INTAN + α54 · BOOK · TRAN + α55 · BOOK · BETA + α56 · BOOK · SIZE + ε; see (1) The variables are defined in Panel B of Table 1, except IND, which is a vector of dummy variables for each industry. This means that there is one constant term for each industry, meaning that fixed industry effects are controlled for. The coefficients of IND are not reported. The set of control variables does not include BTM and MOM. BTM is already represented by BOOK. Including the lagged price (MOM) in the regression would change the specification to a regression of the price change on BOOK. Since the Breusch-Pagan test for heteroskedasticity (H) and the Arallano-Bond test of autocorrelation (AC) detect significant HAC, we employ Newey-West standard deviations when calculating the t- and p-values; see White (1980) and Newey and West (1987). The coefficient estimates are based on OLS, unless in the last regression model that utilizes feasible GLS in which HAC is taken into account in the coefficient estimates; it allows firm-specific heteroskedasticity and first-order autocorrelation (45 observations are lost because of only one observation in the panel). One asterisk * means statistical significance at the 10% level, two asterisks ** means significance at the 5% level and three asterisks *** means significance at the 1% level, tested two-sided. The condition number is a measure of multicollinearity. If it is above 20, there is some troublesome multicollinearity; if it is above 30, there is severe multicollinearity; see Belsley, Kuh and Welsch (1980). The condition number with control variables equals 29.11 or 28.73, which indicates problematic multicollinearity. However, further analyses of the variance-decomposition proportions suggest that the test variable BOOK · IFRS is not severely collinear with any other variable – and therefore not an important source of the identified multicollinearity. Furthermore, the variance inflation factor of the test variable is only 2.72. Due to lacking observations, the sample is reduced from 725 to 635 when control variables are employed in the model.

value of equity BOOK is 0.896; it is highly significant.33 The interaction effect between the BOOK and IFRS yields a coefficient of 0.381, which also is highly significant. Thus, the difference of 0.381 between the BRC of the IFRS sample and the NGAAP sample firms is large enough to reject the null hypothesis of equal BRC in favour of Hypothesis 1. But to conclude in this way is premature, as we have not yet controlled for other drivers of differences in the response coefficient of the balance sheet between the two samples, for example earnings per share EARN’.

The second regression model in Table 4 presents the results when appropriate control variables are utilized – both variables related to risk and attributes of earnings.34 Observe that the condition number is 29.11, suggesting some problematic multicollinearity. Fortunately, the collinearity is not related to the test variable BOOK · IFRS, as its variance inflation factor is only 2.72 and is thereby far below the often emphasized cut-off value of 10.35 Whether there is severe multicollinearity among some of the control variables does not matter – as we are not much concerned with their regression coefficients. The coefficient of the test variable is estimated at 0.161, which is only weakly significant.36 We apply the concept significant when the significance level is below 5%, tested two-sided. If the significance level is below 1%, it is termed highly significant. If it is below 10%, it is weakly significant – but not emphasized. We have tested the extent of heteroskedasticity (H) and autocorrelation (AC) in the error terms by the Breusch-Pagan test and the Arellano-Bond test, respectively. The result is that the null hypothesis of homoskedasticity and no autocorrelation should be rejected. We choose to stick to OLS but adjust the standard deviation for arbitrary HAC when we calculate t- and p-values; see White (1980) and Newey and West (1987). However, we use, as a robustness test, feasible GLS to estimate the regression coefficients allowing for a general covariance matrix.

Note that we have chosen to exclude BTM and MOM from the price regression. BTM is already represented by BOOK. If the momentum MOM, equal to the lagged value of the price, is included as a variable in this specification, the book value will explain the change in price, not the price itself. Changes are analysed by the return regression.

The variance inflation factor VIF = 1/(1 - R2), in which R2 is the explained variation in a regression of one explanatory variable on all the other explanatory variables. A VIF of 2.72 means that the regression of the test variable on all the other explanatory variables yields an R2 of 63.2%. The collinearity is too low to be detrimental. The threshold for this is 90%.

Notice that earnings EARN’ is a separate control variable and a variable that moderates BOOK in the second regression model of Table 4. If EARN’ is also moderated by IFRS, the estimated coefficient of the test variable

ticity and first order autocorrelation into account when estimating the coefficients; see e.g.

Green (2008, pp. 154-158). We observe that the coefficient of the test variable BOOK · IFRS is estimated at 0.206 and is highly significant. Accordingly, the results suggested by the two OLS models are confirmed and the significance level is strengthened, meaning that the result is robust for changes in the statistical estimation technique.

If we, as an untabulated robustness test, use a constant sample of 113 identical firms each with two IFRS and two NGAAP observations, the estimated test variable coefficient is 0.110 by OLS and 0.177 by GLS. Only the latter coefficient is highly significant (t-value = 7.64).

This suggests that the finding of a higher balance sheet response coefficient according to IFRS than according to NGAAP is robust for an extended control in which the firms are identical in the two samples, in addition to controlling for differences in risk and earnings attributes related to each firm over time.

As a second untabulated robustness test, we reintroduce the full sample without removal of ‘extreme’ observations; see Panel A of Table 1. We focus on the OLS regression model with control variables. Now the regression coefficient of the test variable is estimated at 0.312, which is highly significant (t-value = 3.00). Thus, the removal of ‘extreme’ observations reduces the differences in the response coefficient of the book value between IFRS and NGAAP from 0.312 to 0.161.

BOOK · IFRS is estimated at 0.248, but insignificant due to increased multicollinearity. The VIF increases from

2.72 to 4.70.

ing outliers by simple truncation, we perform an initial screening based on Cook’s distances larger than one to eliminate gross outliers before calculating starting values and then perform Huber iterations followed by biweight iterations as suggested by Li (1985). Then the coefficient of the test variable is estimated at only 0.084, but still it is highly significant (t-value = 5.88).

The null hypothesis cannot be rejected in favour of Hypothesis 1 by our main test, which is the OLS regression with control variables and statistical inferences based on HAC standard deviations. However, all tests, the main as well as the robustness tests, indicate that IFRS leads to a higher response coefficient of the book value of equity – and hence to a higher balance sheet response. All the employed robustness tests, except one, find that the difference in coefficients is statistical significant. All in all, we interpret the evidence as consistent with Hypothesis 1, and conclude that the response coefficient of the equity book value is larger under IFRS than under NGAAP.

4.2 Test of Hypothesis 2 Hypothesis 2 says that earnings response coefficients are different when financial statements are prepared according to NGAAP than when prepared according to IFRS. The test methodology (4) - (6), i.e. the return regression, is applied to test the hypothesis. The results of the tests are presented in Table 5.

The regression model is RET = β0 · IND + β1 · EARN + β2 · IFRS + β31 · LOSS + β32 · INTAN + β33 · TRAN + β34 · BETA + β35 · SIZE + β36 · BTM + β37 · MOM + β4 · EARN · IFRS + β51 · EARN · LOSS + β52 · EARN · INTAN + β53 · EARN · TRAN + β54 · EARN · BETA + β55 · EARN · SIZE + β56 · EARN · BTM + β57 · EARN · MOM + ε; see (4) - (6). The variables are defined in Panel B of Table 1, except IND, which is a vector of dummy variables for each industry. This means that there is one constant term for each industry, so that fixed industry effects are controlled for. The coefficients of IND are not reported. Since the Breusch-Pagan test for heteroskedasticity (H) and the Arallano-Bond test of autocorrelation (AC) detect significant HAC, we employ Newey-West standard deviations when calculating the t- and p-values; see White (1980) and Newey and West (1987). The coefficient estimates are based on OLS in the two first regression models. In the last model, coefficients are estimated based on feasible GLS, allowing for panel specific heteroskedasticity and first order autocorrelation (the number of observations is reduced by 26 because of only one observation in these panels). One asterisk * means statistical significance at the 10% level, two asterisks ** means significance at the 5% level and three asterisks *** means significance at the 1% level, tested two-sided. The condition number is a measure of multicollinearity. If it is above 20, there is some troublesome multicollinearity; if it is above 30, there is severe multicollinearity; see Belsley, Kuh and Welsch (1980). The condition number with control variables equals 33.70, which indicates severe multicollinearity. However, further analyses of the variance-decomposition proportions suggest that the test variable EARN · IFRS is not severely collinear with any other variable – and therefore not an important source of any problematic multicollinearity. Furthermore, the variance inflation factor of the test variable is only 2.34. The sample is reduced from 629 to 570 due to lacking observations of one control variable MOM. If MOM is removed, then the test variable obtains a coefficient of -0.509 with t-value -1.97, which is significant at the 5% level.

IFRS without control variables. The coefficient is estimated at 0.562, implying a difference in earnings response coefficient ERC of the same sign and magnitude between IFRS and NGAAP. The second regression model includes the full set of control variables. Now, the difference in ERC between IFRS and NGAAP is estimated at -0.354, but it is not statistical significant.37, 38 Notice that the number of observations falls from 629 to 570, when control variables are included in the regression model. The reason for this is missing observations of one control variable MOM; see also Panel C of Table 2. This variable is not statistically significant. If we drop MOM, the difference in ERC is estimated at -0.509, which is significant at the 5% level.

This suggests that there is some evidence from the OLS regression in support of rejecting the null hypothesis in favour of Hypothesis 2. This analysis also illustrates the importance of controlling for other changes than reporting regime between the two samples. Without controls, there is evidence that the ERC of the IFRS sample is larger than for the NGAAP sample. With controls, the opposite is found.

The third regression model in Table 5 employs GLS taking into account heteroskedasticity and first order autocorrelation instead of OLS with HAC standard deviations. The difference in ERC is -0.250 and highly significant. The first untabulated robustness test is to require an identical sample of firms each with two IFRS and two NGAAP observations. The coefficient If EARN is split in the level LEARN and the change ∆EARN and the analysis is performed by focusing on ∆EARN with LEARN as an additional control variable, the difference in ERC between IFRS and NGAAP is still insignificant.

If we instead use the price regression with control variables for earnings, including the book value interacting with earnings and accounting regime, and test whether the earnings response coefficient differs between IFRS and NGAAP, the coefficient of the test variable EARN’ · IFRS is estimated at -0.682, which is insignificantly different from zero.

cant).

As a second untabulated robustness test, we reintroduce the full sample without removal of ‘extreme’ observations; see Panel A of Table 1. The OLS coefficient of the test variable with the full set of control variables is estimated at -0.465, which is highly significant (t-value = If we instead, as a third untabulated robustness test, remove ‘extreme’ observations as suggested by Li (1985), the estimated coefficient is -0.538. The corresponding t-value is Again; the difference between the earnings response coefficient in the IFRS and NGAAP samples is highly significant.

We cannot reject the null hypothesis in favour of Hypothesis 2 by our main test, which is the controlled OLS regression with statistical inferences based on HAC standard deviations.

However, all tests with control variables, main as well as robustness tests, estimate the highest earnings response coefficient under NGAAP. In nearly all robustness tests, the difference in ERC is statistically significant. All in all, we choose to interpret the evidence consistent with Hypothesis 2. The earnings response coefficient ERC is higher under NGAAP than under IFRS.

5. On the Sources of the Structural Breaks in Response Coefficients In this section, we analyze the sources of the identified structural breaks in the response coefficients between the two reporting regimes. This is done by disaggregating the book value of equity and earnings into their underlying components. Then we analyze whether the response

NGAAP.