«Elizabeth Blankespoor University of Michigan Stephen M. Ross School of Business blankbe January 2012 Abstract This paper examines the ...»
Requiring valid output from the parsing procedure further reduces the sample to 13,969 10-K filings for 4,427 firms. Tier 1 firms’ first detailed XBRL filings began for fiscal periods on or after June 15,
2010. Of the original 397 detailed XBRL 10-K reports filed, I am able to obtain the necessary Compustat and tag PERL output for 323 filings. Table 1 Panel A provides details of the sample selection. As Panel B of Table 1 shows, the observations are fairly evenly distributed over the years.
3.2 Variable Definition Disclosure Measure To measure the amount of firm disclosure, I focus on the detailed, quantitative footnote disclosures, or the numbers in the footnotes. The number of quantitative disclosures in the footnotes is available post-XBRL for Tier 1 firms as the number of footnote XBRL tags. However, by definition, the XBRL information is not available for Tier 1 firms prior to XBRL implementation, or for my control group (Tier 2 and 3 firms) at any point during the time period. Therefore, I estimate 2006 was the first fiscal year that non-accelerated filers as well as accelerated filers were required to comply with Sarbanes Oxley disclosure requirements (http://www.sec.gov/news/press/2005-25.htm). Thus, I use 2006 through 2010 filings to maximize the number of observations available to model disclosure choice while still maintaining a level of comparability across firms.
the number of quantitative disclosures, Tags_Notes, for all filings using Perl to count the numbers in the footnotes.18 Similar to previous disclosure measures such as the number of press releases or the average number of words in press releases, Tags_Notes captures the quantity of disclosure, which may or may not equate to the quality of disclosure. Just as more words can obfuscate the meaning of disclosure, more numbers can increase the noise that investors are required to sift through. Still, to the extent that each quantitative disclosure provides investors with another piece of information, Tags_Notes captures an aspect of the depth and quality of a firm’s disclosure in a way unique from prior disclosure literature.
Control Variables To model firms’ disclosure choice and ensure my results are not driven by other firm characteristics, I include several control variables, including the qualitative information content of the filing, firm performance, the presence of information intermediaries, and additional firm characteristics associated with the level of disclosure. I discuss below the motivation for including each control variable, and I provide detailed variable definitions in Appendix B.
Following Li (2008) and Miller (2011), I capture the log of the number of words in the footnotes (LnWords_Notes) and the footnotes’ fog score (Fog_Notes) as measures of the qualitative information and disclosure readability, respectively.19 In general, firms with longer reports are likely to have more information to provide and thus more quantitative disclosures as well. Since prior literature shows performance to be positively related to disclosure (Lang and Lundholm 1993, Miller 2002), I include the firm’s return on assets for the fiscal period (ROA) and the market-adjusted return over the twelve months ending in the filing’s fiscal period (PyAbnRet). I include the log of one plus the number of analysts covering the firm (LnAnalyst), the percent of shares outstanding held by See Appendix A for details of process of identifying the footnotes and counting the numbers.
I choose to use the log value for those control variables that are skewed or are likely to have a nonlinear relation with disclosure, based on prior literature. As robustness, I rerun my main results using quartile indicator variables for the number of words in the footnotes (instead of the log), and inferences remain unchanged.
institutions (InstHoldings), and the log of the number of shareholders (LnNumShareholders) to control for the effect of information intermediaries and differences in shareholders’ demands for disclosure quality on firms’ disclosure choices (Bushee, Matsumoto, and Miller 2003, Lehavy, Li, and Merkley 2011). Finally, I control for several firm characteristics that have historically been related to disclosure. I use the log of the firm’s market value (LnMV) and the log of the number of business segments (LnSegments) to control for firm size (Lang and Lundholm 1993, Li 2008), and I include the firm’s market-to-book equity ratio (Mtb) to control for the firm’s investment opportunities and growth potential. I control for the volatility of the firm’s operations using the standard deviation of the change in split-adjusted earnings per share over the previous five years (EarnVol) and the log of the standard deviation of the firm’s daily stock returns over the twelve months ending in the filing’s fiscal period (LnRetVol) (Waymire 1985, Bushee and Noe 2000). Also, I winsorize all variables at 1% and 99% to reduce the effect of outliers.
3.3 Descriptive Statistics Table 2 provides descriptive statistics for the 10-K filing sample. As shown in Panel A, the mean number of quantitative disclosures in the footnotes is 1,102, as compared to the 315 average items on the face of the financial statement. The footnotes contain 11,749 words on average and have a Fog score of 19.7, which is similar to Li (2008). The mean (median) firm has market float (per the 10-K disclosure) of $2.7 billion ($345 million), assets of $4.9 billion ($539 million), analyst following of seven (five), and institutional holdings of 57% (64%). Table 2 Panel B provides the Pearson and Spearman correlations between variables, with Spearman above the diagonal and Pearson below. The number of quantitative disclosures and the (log of the) number of words in the footnotes are highly correlated (0.71 Pearson), which is not surprising since they capture different aspects of disclosure within the footnotes. The number of quantitative disclosures is also positively correlated with firm size, number of analysts, institutional holdings, the number of shareholders, and firm performance.
Table 3 Panel A compares the XBRL and non-XBRL samples. Since the largest firms were required to adopt XBRL first, the XBRL sample is larger than the non-XBRL sample, with an average market float of $18.9 billion for XBRL versus $1 billion for non-XBRL. The average amount of footnote disclosure is larger for XBRL firms, with the mean XBRL firm having 1,825 numbers and 17,787 words in the footnotes and the mean non-XBRL firm having 1,024 numbers and 11,100 words. Since XBRL firms are larger on average than non-XBRL firms and size has historically been associated with more disclosure, this difference in disclosure amount is understandable. The complexity of the footnotes are similar, though, with a fog score for the footnotes of 19.6 and 19.7 for XBRL and non-XBRL, respectively.20 Table 3 Panel B (C) examines the mean (median) change in the number of numbers in the footnotes for XBRL and non-XBRL firms in the pre- and post-XBRL periods. As shown, XBRL firms significantly increase their quantitative footnote disclosures from the pre- to the post-XBRL period, while the increase for non-XBRL firms is not significant. When I compare the two groups, the increase for XBRL firms is significantly larger than for non-XBRL firms, implying that XBRL firms respond to the anticipated reduction in investor processing costs by increasing their quantitative disclosures. However, these comparisons are univariate. To ensure there are not other changes in firms’ information environments driving the disclosure choice, I turn next to multivariate tests.
4. Research Design and Results
4.1 Investor Information Costs and Firm Disclosure Main Research Design To examine the effect of adoption of XBRL detailed tagging requirements on firm disclosure
choice, I first estimate the following OLS regression using XBRL firms only:
TagsNotesi,t = β0 + β1Postt + βi ∑ControlVariablesi,t + Firm FEi + ε (1) To ensure that any difference between the change in disclosure for XBRL firms versus non-XBRL firms is due to XBRL adoption rather than systematic differences in firm characteristics, I include firm fixed effects as well as numerous determinants of disclosure in my analyses. I also perform a robustness test using a subsample of more similarly sized XBRL and non-XBRL firms and find similar (although weaker) results. See section 5.2 for details.
TagsNotes and the control variables are as defined in section 3. Post is an indicator variable equal to one if the filing’s fiscal period is June 15, 2010 or later (fiscal year 2010). A positive coefficient on Post (β1) indicates an increase in XBRL firms’ quantitative footnote disclosures in the year of detailed XBRL tagging of quantitative footnote disclosures. In addition, I control for fixed idiosyncratic firm disclosure choices by including firm fixed effects, and I cluster standard errors by firm to control for transitory shocks that are correlated across time for a given firm. 21 In the second model, I utilize the staggered adoption of XBRL by examining filings of nonadopting firms for the same time periods and using these firms to control for any systematic changes other than XBRL that affected firms’ disclosures. Specifically, I estimate the following OLS
regression using all available firm-year observations:
TagsNotes and the control variables are as defined above. In addition to firm fixed effects and firmclustered standard errors, I control for time-related effects by including year fixed effects.
Post*XBRL is the interaction of Post (an indicator variable for the post-adoption year, 2010) and XBRL (an indicator variable for firms that adopt XBRL detailed tagging requirements). I do not include the main effects of Post and XBRL in model 2 because the year and firm fixed effects encompass the variation in Post and XBRL, respectively, preventing estimation of their coefficients.
However, I also report results from a variant of model 2 that does not include fixed effects and thus does include Post and XBRL. For either specification, the coefficient on Post*XBRL (β1) captures the difference between the change in XBRL firms’ disclosure and the change in non-XBRL firms’ disclosure before and after implementation of XBRL, or the difference-in-difference impact of detailed tagging adoption on the amount of disclosure, controlling for other firm and time effects.
Note that I do not include time fixed effects in this model because they would encompass the variation in the variable of interest (Post). Also, I do not cluster standard errors by time (to create two-way clustered standard errors) because there are five years in my sample, and to create consistent estimates of standard errors, at least 10-50 clusters (i.e. years, in this case) are recommended (Petersen 2009, Gow, Ormazabal, and Taylor 2010).
Main Results Table 4 provides the multivariate regression results for the effects of XBRL detailed tagging on the amount of quantitative disclosure in 10-K filings. Model 1 shows a positive coefficient for Post significant at the 1% level, confirming the main prediction of increased disclosure for XBRL firms upon adoption of detailed tagging. In addition, the magnitude of the effect appears to be economically significant; the coefficient for Post of 135 implies an average increase of 135 footnote numbers, or approximately 7% of the mean XBRL firm’s quantitative footnote disclosures. Model 2, which includes both XBRL and non-XBRL firms, shows a positive coefficient for Post*XBRL significant at the 1% level, confirming that the increase in XBRL firms’ disclosure is significantly greater than any change in non-XBRL firm disclosure. I also report a third model that removes the firm and year fixed effects from model 2, allowing estimation of coefficients for Post and XBRL. The coefficient for Post*XBRL remains positive and significant at the 1% level. The coefficient for XBRL is positive and significant, consistent with univariate statistics in Table 3. The coefficient for Post is negative and significant, rather than insignificant as shown in Table 3, but it does not seem to drive the main difference-in-difference results, given the much smaller magnitude (18.9 versus 125.1).
The coefficient estimates for the control variables show that the number of quantitative footnote disclosures increases with the number of words in the footnotes (LnWords_Notes), the market value (LnMV), the number of segments (LnSegments), the number of shareholders (LnNumShareholders), and the firm’s return on assets (ROA), consistent with prior findings that disclosure increases with firm size and performance (Lang and Lundholm 1993, Miller 2002). Surprisingly, disclosure is negatively related to the firm’s prior year stock performance (PyAbnRet) and number of analysts (LnAnalysts).22 Volatility of earnings and price is positively associated with quantitative disclosure, The negative coefficient on LnAnalysts (and insignificant coefficient on Inst_Hold) is surprising because prior literature shows that analysts and institutions are associated with more disclosure (e.g. Lang and Lundholm 1996, Ajinkya, Bhojraj, and Sengupta 2005). To provide additional comfort that analysts and institutions do help discipline rather than negative as might be expected given Waymire’s (1985) finding that firms with more volatile earnings are less likely to provide earnings forecasts. However, detailed footnote disclosure has a different context and purpose than summary earnings guidance (Merkley 2011). If volatile firms find earnings measures to be less helpful for investors, they would be likely to provide more detailed footnote disclosures and fewer earnings forecasts.
4.2 Variations in Information Costs To further determine whether the increase in disclosure is related to anticipated decreases in investor information processing costs, I examine two cross-sectional settings where the level of investor processing costs vary based on firm or investor characteristics.
Complicated Firms To test whether complicated firms have a greater increase in disclosure as investor processing costs decrease (P2), I estimate the following OLS regression, including year and firm fixed effects
and firm-clustered standard errors:
TagsNotesi,t = β0 + β1Post*XBRLi,t + β2Post*XBRL*ComplicatedFirmsi,t + β3ComplicatedFirmsi,t