Our empirical model is estimated on a pooled basis, as pooled estimation increases the degrees of freedom and may improve the relative precision of the estimated parameters (Gujarati, 1995 D.N. Gujarati, Basic econometrics, 3/e, McGraw-Hill, New York (1995).Gujarati, 1995). A potential drawback is that the observations are not independent. To address this concern, we assess statistical significance using Huber–White robust standard errors ([Huber, 1967] and [White, 1980]).17
We estimate our model after performing a rank transformation of all the variables. This rank transformation is done for the following reasons. First, as noted above, the distribution of PS ratios is positively skewed. As a result, skewness and kurtosis tests indicate that the error term from using OLS with untransformed variables is not normally distributed. Using rank regression, the null hypothesis of normality of the residuals cannot be rejected. Second, the model is estimated on a pooled basis. An assumption of the pooled approach is that the residual variances are unchanged over time. However, as indicated in Table 1, the variability in 1999 PS ratios (standard deviation = 61.03) far exceeds that in 2000 (3.63). Rank regression controls for these differences across time periods. Third, rank regression does not assume a linear relation between the dependent and explanatory variables, while a specific nonlinear form need not be specified. Based on the Ramsey RESET test, linearity is rejected (not rejected) using untransformed (rank transformed) variables. To avoid parameter estimates that may be partially dependent on the number of observations, the ranks are scaled by the number of observations (Cheng, Hopwood, & McKeown, 1992). Overall, rank regression produces ‘well behaved’ estimates, as nearly all studentized residuals have an absolute value less than 2 (none greater than 3). In addition, all condition indices are less than 15, below the rule-of-thumb critical value of 30 (Belsley, Kuh, & Welsch, 1980).
The third column in Table 3 presents coefficient estimates for our model based on the full sample. Statistical significance is assessed based on two-tailed tests, except for those coefficients with expected positive/negative signs. With respect to the primary variables of interest, the coefficient estimate for %BARTERAPB (0.127, t = 0.83) does not significantly differ from zero. This is consistent with investors valuing barter revenue reported under APB 29 at the same multiple as other sources of revenue. Similar results occur for the coefficient estimate for %BARTEREITF (0.009, t = 0.06).
With respect to the remaining coefficient estimates, several points are noteworthy. First, consistent with expectations, the coefficient on GROWTH (0.289, t = 4.00) is significantly greater than zero. Second, the coefficients on GM (0.056, t = 0.71), OPEXP (0.121, t = 1.37), and BETA (0.120, t = 1.19) are not significantly different than zero. The significance of GROWTH and insignificance GM and OPEXP are consistent with the importance of sales over profitability in valuing the sample firms. Finally, the coefficient on BURN (0.379, t = 4.42) is significantly different than zero in the expected direction.
We further examine the valuation of barter revenue by interacting the %BARTER terms with segment dummy variables. This specification is motivated by the possibility that barter revenue may be valued differently for firms in the content/communities and search/portal segments since barter transactions relate to advertising, which is the main source of revenue in these sectors. In other words, barter revenues reported by e-tailers may be viewed as possible accounting manipulations. The results are reported in the fourth column of Table 3. The coefficient estimate for %BARTERAPB DET (− 0.325, t = − 1.71, p = 0.095) is consistent with a discount being applied to barter revenue of e-tailers. In contrast, the insignificant coefficient estimate for %BARTERAPB (1 − DET) (0.185, t = 1.12) indicates that barter revenue recognized by firms in the content/communities and search/portal segments under APB 29 was valued at the same multiple as other sources of operating revenue. The results differ for barter revenue recognized under EITF 99-17. Specifically, the insignificant coefficient estimates for %BARTEREITF DET (− 0.191, t = − 0.69) and %BARTEREITF (1 − DET) (0.151, t = 0.93) imply that barter revenue recognized under EITF 99-17 was valued at a similar multiple as other sources of operating revenue for all segments. We attribute the presence of a discount on e-tailers' barter sales in 1999 and its subsequent absence in 2000 to investor skepticism regarding the quality of barter revenue reported by e-tailers under the general guidance of APB 29.18 The inferences with respect to the remaining coefficients are consistent with those in the third column.
Including all firms in the chosen Internet industry segments (regardless of whether or not barter revenue is disclosed) results in little cross-sectional variation in %BARTER along with numerous zero values, which may bias downward the coefficient estimates on these terms. Accordingly, the final column of Table 3 presents the results from re-estimating the previous model using the subsample of 24 firms that reported barter revenue in 1999 and/or 2000. With respect to the variables of interest, there is an increase in the statistical significance of %BARTERAPB DET (− 0.563, t = − 2.20, p = 0.038). We attribute this to increased power from removing the zero-valued observations, despite the overall decrease in firm-year observations. The coefficient estimates for the remaining %BARTER variables remain insignificant. Consistent with the full sample results, the coefficients on GROWTH (0.340, t = 4.25) and BURN (0.236, t = 2.46) are positive and significant.
The most notable differences from the full sample results are the significance of the overall intercept term (0.419, t = 4.12), D2000 (− 0.394, t = − 4.51), and DCC (− 0.108, t = − 3.01). We interpret this result to suggest that there is an overall average value for firms with barter revenues that is absent for non-barter firms. We investigate this possibility using the 21 firms with no reported barter revenue in either year (42 observations), and find that consistent with our conjecture, all intercepts are statistically insignificant. Moreover, for firms in the content and communities segment, the industry-specific factor becomes relevant for firm valuation in the presence of barter revenues. Furthermore, consistent with our conjecture, the coefficient on D2000 is significantly less than zero, reflecting the lower ISDEX level in April 2001. We also note that the full sample estimates for D2000 (− 0.196, t = − 1.51) and DCC (− 0.082, t =−1.60) are marginally significant.
Overall, the results are consistent with the perception of biased measurement under the lenient recognition standards of APB 29 and increased credibility (and value relevance) of reported barter revenue under EITF 99-17.
