A New Measure of Disclosure Quality - The Level of Disaggregation of Accounting Data in Annual Reports
A New Measure of Disclosure Quality - The Level of Disaggregation of Accounting Data in Annual Reports
A New Measure of Disclosure Quality - The Level of Disaggregation of Accounting Data in Annual Reports
12094
Journal of Accounting Research
Vol. 53 No. 5 December 2015
Printed in U.S.A.
ABSTRACT
1017
Copyright
C , University of Chicago on behalf of the Accounting Research Center, 2015
1018 S. CHEN, B. MIAO, AND T. SHEVLIN
1. Introduction
We construct a new measure of disclosure quality, disaggregation qual-
ity (DQ), based on the level of disaggregation of financial data items in
firms’ annual reports, and provide validation tests. We base DQ on the
theoretical premise that finer information is of higher quality (Blackwell
[1951]). Greater disaggregation leads to more, finer information available
to investors. More detailed disclosure reduces information asymmetry, ar-
guably increases the precision of the information in the financial state-
ments, and provides investors with more information for valuation and
mitigates mispricing (Fairfield, Sweeney, and Yohn [1996], Jegadeesh and
Livnat [2006]). Greater disaggregation also enhances the credibility of
firms’ financial report as it gives managers fewer degrees of freedom to
manage the reported numbers (Hirst, Koonce, and Venkataraman [2007],
D’Souza, Ramesh, and Shen [2010]), enhancing the contracting and stew-
ardship role of accounting information. Reasoning along this line, we argue
that a greater degree of disaggregation represents higher disclosure quality.
DQ is parsimonious and applicable to all Compustat industrial firms: we
count the number of nonmissing financial items reported in firms’ annual
reports, including items both in the financial statements and in the foot-
notes, as captured by Compustat. A higher count of nonmissing account-
ing data items represents higher disclosure quality. Despite a vast empirical
literature on disclosure in general and voluntary disclosure in particular,
there is surprisingly no overall measure of disclosure quality based on a
comprehensive set of accounting data as reported in financial reports.1
1 While the concept of disaggregation has been explored by researchers in a segment re-
porting setting (e.g., Berger and Hann [2003, 2007], Bens, Berger, and Monahan [2011]),
the concept of disaggregation of financial statement data items has received scant research
attention. We are aware of only one other paper (D’Souza, Ramesh, and Shen [2010]) that ex-
plicitly addresses the disaggregation of accounting data items in firms’ earnings press releases.
A NEW MEASURE OF DISCLOSURE QUALITY 1019
2 According to Compustat, the average tenure of its data collection staff in the United States
is 10.3 years, and some of them have been collecting data for over 20 years. Also, the staff is
highly educated and primarily has finance and business degrees, as well as Masters Degrees
and CFA certifications. Also, each of the staff goes through extensive training before updating
the data on a company, and staff bonuses are awarded for low errors and productivity.
3 Another reason we do not construct a statement of cash flow DQ is because multiple
reporting formats are allowed for the Statement of Cash Flows over our sampling period 1973–
2009: the formats for pre-1989 and post-1989 years are dramatically different, and pre-1989
firms could employ three different formats in presenting their Statement of Cash Flows. This
heterogeneity makes it difficult to construct a meaningful, parsimonious disclosure score for
all years. This exclusion is a caveat to our DQ measure.
1020 S. CHEN, B. MIAO, AND T. SHEVLIN
4 Both Beyer et al. [2010] and Berger [2011] highlight the difficulty in drawing causality
between disclosure and cost of capital. Neither theoretical nor empirical research agrees on
whether information quality should impact cost of capital. While many researchers demon-
strate that higher information quality should lead to lower cost of capital (Easley and O’Hara
[2004], Francis et al. [2005], Lambert, Leuz, and Verrecchia [2007], Kelly and Ljungqvist
[2012]), others argue that information asymmetry can be diversified away in large economies
and thus should have no impact on cost of capital (e.g., Hughes, Liu, and Liu [2007]). See
Shevlin [2013] for a summary of the literature.
5 Note our validation tests rely on the concurrent validity approach instead of the conver-
gent validity approach. In the latter approach a new measure is validated through high cor-
relation with existing measures of the same construct. Since DQ is capturing a fundamentally
different construct from existing disclosure measures, such as management forecasts and the
Fog Index, the convergent validity approach is not appropriate for our setting. The concurrent
validity approach examines how well one measure relates to other criterion that are presumed
A NEW MEASURE OF DISCLOSURE QUALITY 1021
to occur simultaneously. We employ the concurrent validity approach for our validation tests,
as disclosure quality should be positively correlated with concurrent variables prior literature
has shown to be associated with information quality. For a more detailed discussion of conver-
gent and concurrent validity approaches, please see Kline [2014].
1022 S. CHEN, B. MIAO, AND T. SHEVLIN
time generally yield significant though weaker results than the pooled cross-
sectional tests. These results suggest that DQ is likely better suited to cross-
sectional studies or event studies surrounding changes in firms’ operations.
For example, DQ can be used in settings where a firm changes its auditor
or top management or when a firm undergoes mergers and acquisitions.
Further, we caution that, since DQ captures a very unique aspect of dis-
closure quality, the level of details of accounting data in annual reports,
DQ differs from all existing measures and as such cannot be construed as
a simple replacement of existing measures without giving thought to the
underlying theoretical construct of interest.
The rest of our paper is organized as follows. In the next section we mo-
tivate our new measure of disclosure quality DQ. Section 3 offers detailed
discussions of the construction of DQ. Section 4 presents descriptive statis-
tics on DQ and section 5 presents results on the validation tests and addi-
tional tests separating DQ into components and on the temporal variation
in DQ. We conclude in section 6.
6 We acknowledge the possibility that finer information is not necessarily better information
when strategic concerns are involved. If some firms or industries produce more aggregated
information because of concerns with proprietary costs, DQ will be lower. Even though such
strategic nondisclosure might benefit managers, the lower DQ score still reflects lower quality
disclosure to users of annual reports.
A NEW MEASURE OF DISCLOSURE QUALITY 1023
of managerial discretion in financial reporting within the confines of GAAP. Other specific
examples of managerial discretion include the choices between using LIFO versus FIFO for
inventory costing and choices among different presentation formats for market risk disclo-
sures under FRR48.
1024 S. CHEN, B. MIAO, AND T. SHEVLIN
8 Lang and Lundholm [1993] report that a typical year’s Financial Analysts Federation re-
port provides ratings for about 27 industries, each with an average of 17 firms evaluated by an
average of 13 analysts in each industry. The firms covered by AIMR are the largest and most
heavily followed firms in each industry.
A NEW MEASURE OF DISCLOSURE QUALITY 1025
Fog Index can be applied to all firms, it is only applicable to MD&A, not to
the entirety of annual reports.
In sum, DQ is easy to construct and can be calculated for all Compus-
tat industrial firms, and is more objective relative to other self-constructed
measures of disclosure. These advantages of DQ make it possible for re-
searchers to replicate or test new hypotheses on disclosure quality on a
much wider set of firms.
3. Construction of DQ
3.1 COMPUSTAT BALANCING MODEL AND COUNTING NONMISSING ITEMS
DQ rests on a count of nonmissing data items in firms’ annual reports as
reported by Compustat. There are two possible scenarios that can lead to a
Compustat missing line item: (1) the firm has the underlying item but does
not report it and Compustat reports it as missing, and (2) the firm does
not have the underlying item and Compustat codes it as missing.9 Though
we do not have reason to believe that missing items as a result of scenario
(2) will be systematically different across firms, and moreover systematically
related to firms’ disclosure behavior, we nevertheless aim to only capture
scenario (1) and purge from DQ items that firms do not have but appear
as missing fields on Compustat. We build in screening mechanisms based
on the nesting feature (i.e., sum of the components equals the total) of
the Balance Sheet and, to a lesser extent, the Income Statement accounts
to mitigate the impact of scenario (2) on DQ. We discuss each of them in
detail below.
In capturing line items reported by firms, Compustat has three “Balanc-
ing Models” for each of the three financial statements, which Compustat
uses as the basic templates in gathering financial statement data. These
models lay out the interrelations among standardized data items on the fi-
nancial statements. Abbreviated versions of the templates (the “Balancing
Models”) for the Balance Sheet and the Income Statement are attached in
appendix A.10 Since our coding of the Balance Sheet and Income State-
ment follows the same logic, we first illustrate our method to arrive at DQ
using the Balance Sheet as an example.
3.2 COUNTING NONMISSING ITEMS RELATED TO THE BALANCE SHEET
We refer to all line items under the column “Item Description” in the
Balancing Model (appendix A) as “subaccounts.” Our goal is to capture
9 A third possible scenario is that a firm reports an item but Compustat does not capture it,
resulting in a missing field. Our extensive communication with experienced Compustat data
experts reveals that these cases are rare, and are usually a result of a firm misplacing an item
in an irrelevant section.
10 The full-blown version of the Balancing Models is available from Compustat’s
11 We read the detailed definition of each of the subaccounts on the Compustat data manual
to link them to a specific parent account. The definitions are straightforward, and only in
rare cases do we need to exercise judgment. To minimize coding errors, we communicate
extensively with senior Compustat data representatives throughout the coding process to gain
an accurate understanding of Compustat’s balancing models.
12 Note that, while removing these two group accounts might render the assets not equal
to the liabilities and shareholders’ equity on the balance sheet, the impact of such removal is
minimal on our DQ measure. The correlation between DQ with and without these two group
accounts is 99.9%.
A NEW MEASURE OF DISCLOSURE QUALITY 1027
counting scheme Balance Sheet items that are irrelevant to firms’ opera-
tions even though they appear as missing fields on Compustat.13
We further check if the subaccounts, where applicable, add up to the
parent account in coding a missing subaccount. This step is our second
screening mechanism for the Balance Sheet DQ measure. For example, if
three of the four inventory components, INVRM, INVWIP, and INVFG add
up to inventory total INVT, then the fourth component INVO (inventory-
other), though a missing field on Compustat, does not count as missing
in DQ because the balance of this account should be zero, indicating that
the firm does not have INVO in that particular year.14 When the subac-
count total does not equal the parent account, we treat a single missing
subaccount as missing.15 If two out of the four inventory subaccounts have
missing Compustat fields, then we count both as missing as we cannot dis-
tinguish between which one of these two subaccounts is truly missing, or if
both are missing. Nevertheless, this second screening mechanism produces
a DQ measure that is strictly better than no screening because such a mech-
anism mitigates Type I error (coding an item as missing when in fact it is
not missing).
While this second screening mechanism for the balance sheet accounts
mitigates Type I error—it corrects over 56% of the potential misclassifica-
tion of subaccounts—it does come with a cost: when the subaccounts add
up to the parent accounts, DQ captures only aggregation by omission and
not aggregation by classification shifting. We note that, to the extent that
DQ misses classification shifting and wrongly codes a firm engaging in clas-
sification shifting as having a higher DQ, it will work to weaken the associa-
tion between DQ and the established information quality variables.
Not all subaccounts lend themselves neatly to a sum that equals the value
of the parent account. For example, while the six subaccounts DCLO (Debt
− Capitalized Lease Obligations), DCVT (Debt – Convertible), DD (Debt –
Debentures), DN (Debt – Notes), DS (Debt – Subordinated), DLTO (Other
Long-Term Debt) on the Linking Table in appendix B add up to the total
DLTT (Long-Term Debt), DD2DD5 (Debt Maturing 2nd5th years) do
13 As another example, a firm that does not have long-term debt will have DLTT (Long-
Term Debt – Total) appearing as zero on the Balancing Model and the associated subaccounts
as missing fields. We do not count these missing fields as missing in our coding.
14 Our coding of inventory items takes into account the fact that a nonmanufacturing firm
does not have INVRM (raw materials), INVWIP (work-in-progress), and INVFG (finished
goods) accounts. For nonmanufacturing firms, we exclude these subaccounts when we count
missing items.
15 We believe that, even though a single missing account’s magnitude can be determined
by subtracting the sum of N−1 nonmissing accounts from the total amount, it nevertheless
imposes more information processing costs on users, and as such is not as transparent as
if the firm itself disclosed it and constitutes lower disclosure quality than if all N accounts are
disclosed. Empirically, only 3,757 observations out of 14,300,972 subaccount level observations
are affected, and our results remain the same if we treat these accounts as nonmissing.
1028 S. CHEN, B. MIAO, AND T. SHEVLIN
16 Note that, even though the group account CITOTAL (Comprehensive Income – Total) is
not on the Income Statement Balancing Model, we classify the associated accounts as income
statement accounts rather than balance sheet items.
17 Note that these percentages are upper bound estimates, as we consider a missing Com-
pustat item as an error even if the firm only mentions the item (without a clear disaggregated
dollar value for the associated magnitudes) in its annual reports.
A NEW MEASURE OF DISCLOSURE QUALITY 1029
where k indexes group accounts. For the Balance Sheet, we are able
to create 11 groups and link these group accounts to 25 second level
accounts (parent accounts), which are in turn linked to 93 associated
subaccounts (see appendix B).18 For each of these 11 groups, we count the
number of nonmissing items in the subaccounts and divide this number
by the total number of subaccounts in that group. For example, the group
ACT (Current Assets - Total) is associated with seven parent accounts,
which are in turn linked to 20 subaccounts. Assuming only 2 out of 20
subaccounts under ACT are missing, then the ratio of nonmissing items in
this group is 18/20.
We then arrive at a value-weighted DQ score for each of the 11 group
accounts by multiplying the above ratio of nonmissing items with a weight,
defined as the asset value of that Balance Sheet group over total asset value.
This value-weighting scheme gives more weight to items that presumably
are more important to firms’ operations and thus to investors.19 Finally,
we sum the value-weighted nonmissing item ratios across the 11 Balance
Sheet groups, leading to a disclosure score with a theoretical minimum of
zero and a theoretical maximum of two (because we have both the asset
side and the liabilities/shareholders’ equity side of the Balance Sheet). We
further divide the score by 2 so the Balance Sheet DQ score, DQ BS, varies
between 0 and 1.
18 Even though Compustat provides 212 items for the Balance Sheet and 131 items for the
Income Statement, many of these items are not relevant for our research setting and are thus
excluded. In counting missing items, we first exclude all items related to financial and utility
firms. We also exclude “formula” items; these “formula” items can be directly derived from
other items on the face of the financial statements. For example, the item Operating Income
before Depreciation (OIBDP) is by construction always equal to Sales (SALE) minus Operat-
ing Expense (XORP). Including such formula items would be double counting the underlying
items twice. Other items excluded include items computed by Compustat (an example would
be Compustat reference items that are not tied to annual reports) and moving average items.
Another example of items computed by Compustat is Invested Capital – Total (ICAPT). This
item is defined by Compustat as the sum of Long-Term Debt, Preferred Stock, Minority Inter-
est, and Common Equity. We also exclude per share items as including them in the count of
missing items would be counting the same underlying data item more than once.
19 If a firm’s total assets consist of 20% intangibles and 80% tangibles, investors would want
more information disclosed on the tangibles as such information presumably is more impor-
tant to their decision making.
1030 S. CHEN, B. MIAO, AND T. SHEVLIN
20 Untabulated results show that, when we use an equal-weighting scheme to arrive at the
Balance Sheet disclosure score, our results are qualitatively similar. However, we believe that
value-weighting for the Balance Sheet is conceptually superior to equal-weighting, hence we
present all our results using the value-weighted Balance Sheet DQ BS.
21 An eighth group account, SALE, has only one subaccount, REVT (Revenue – Total), for
which there is no variation as virtually all firms have nonmissing value. Therefore, we drop this
group account in constructing our DQ measure. Including this group account has no impact
on our results.
A NEW MEASURE OF DISCLOSURE QUALITY 1031
22 Since we exclude financial/utility firms from our sample, the regression is based on
the remaining 10 industries. This industry classification is available from Professor Ken-
neth French’s Web site, http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data
library.html.
23 Note each industry is compared to the base industry reflected in the intercept. The base
industry, Consumer Durables, has a DQ score approximately equal to the sample average and
is thus representative of the overall sample. That is, the base industry does not represent an
outlier that differs from all other industries.
1032 S. CHEN, B. MIAO, AND T. SHEVLIN
24 We stop in 2008 for this temporal model estimation because the new accounting codifi-
cation effective in 2009 makes it difficult to measure one of our variables, STDWORDS , as it is
difficult to unambiguously attribute the issuance of a new Accounting Standard Update as a
new FASB standard.
25 Consistent with this, the average first-order autocorrelation for DQ is 0.668, calculated as
the mean of firm-specific autocorrelations for firms with a minimum of five years data avail-
able. Other researchers document similar “stickiness” in annual reports and 10-K filings of
adjacent years. See, for example, Francis, Nanda, and Olsson [2008]. We note that our sub-
sequent validation test results are strongly robust to clustering standard errors by firm. We
further address temporal variation in DQ in section 5.5.
26 We obtain the Fog Index and the number of word counts in 10-K from Feng Li’s Web site,
http://webuser.bus.umich.edu/feng.
A NEW MEASURE OF DISCLOSURE QUALITY 1035
TABLE 2
Correlation Between DQ and Other Disclosure Measures
(Pearson Upper Triangle, Spearman Lower Triangle)
DQ DQ BS DQ IS MF FOG #WORDS AIMR
DQ – 0.876 0.884 0.292 0.036 −0.000# −0.021#
DQ BS 0.884 – 0.548 0.262 0.041 −0.026 −0.030#
DQ IS 0.867 0.563 – 0.252 0.024 0.026 0.006#
MF 0.233 0.211 0.206 – 0.038 0.084 –
FOG 0.135 0.125 0.117 0.015 – 0.264 –
#WORDS 0.078 0.050 0.106 0.142 0.327 – –
AIMR 0.017# −0.025# −0.022# – – – –
The sample consists of 41,692 firm-year observations from 1993 to 2011. The AIMR sample consists of
3,265 firm-year observations from 1981 to 1995. Financial, utility, and foreign companies are excluded from
this sample. All correlations are significant at 1%, except for those marked with # , which are insignificant at
conventional levels. Variable definitions: MF = # of management forecasts, year t. If a firm does not provide
a management forecast, MF is set to be zero; FOG = Fog Index on readability of MD&A, available from Feng
Li’s Web site, http://webuser.bus.umich.edu/feng; #WORDS = the total # of words in 10-K, year t, available
also from Feng Li’s Web site; AIMR = percentage rank of AIMR annual report disclosure score; DQ BS =
value-weighted disclosure quality score of balance sheet items, [0,1]; DQ IS = equally-weighted disclosure
quality score of income statement items, [0,1]; DQ = the simple average of DQ BS and DQ IS (DQ = 0.5 ∗
(DQ BS + DQ IS). T-statistics are in parentheses.
the First Call CIG (Company Issued Guidance) file and Feng Li’s Fog index
per his Web site. Though DQ (as well as the component scores) is positively
correlated with MF and the two readability measures FOG and #WORDS, the
magnitudes of the correlation are fairly small—between 0.050 and 0.233 for
the Spearman correlations, all significant at the 1% level.
Second, we also report the correlation of DQ with AIMR. The sample
for this analysis is even smaller due to the inherent small sample size with
available AIMR scores. All of the correlation coefficients are small and none
of the correlation coefficients are significant.27
These correlation results are expected as DQ conceptually captures a
different aspect of disclosure quality. As such, our validation tests cannot
employ the widely used convergent validity approach—correlation of DQ
with other existing disclosure measures. The convergent validity approach
is only appropriate if the new measure is capturing the same underlying
construct as existing measures. We instead rely on the concurrent validity
approach by demonstrating DQ’s correlation with variables prior literature
27 This lack of a significant correlation can be due to a number of factors: First, finan-
cial statement and footnote details, the concept underlying DQ, is just one of the many as-
pects that analysts rate in constructing AIMR. AIMR also includes analysts’ ratings of many
qualitative aspects of firms’ disclosure quality, such as the “amount of detail about the cor-
porate officers” and the “availability and timeliness of other written materials, such as press
releases, proxy statements, summary of annual meeting proceedings, and presentations to
analyst groups” (Lang and Lundholm [1993], p. 254). Second, AIMR includes analysts’ quan-
tification of “qualitative disclosure (e.g., management discussion and analysis) and disclosure
which may not have been reflected in published financial statements” (Lang and Lundholm
[1996]). Third, AIMR is a weighted average of all the myriad components. Thus, it is possible
that the variation in AIMR is dominated by all the other aspects of nonfinancial statement
quantitative and qualitative disclosure, yielding a low correlation between AIMR and DQ.
1036 S. CHEN, B. MIAO, AND T. SHEVLIN
5. Validation Tests
We perform three sets of validation tests: We examine the association be-
tween DQ and variables prior literature has shown to be associated with
information quality/asymmetry, namely, analyst forecast dispersion and
A NEW MEASURE OF DISCLOSURE QUALITY 1037
TABLE 3
DQ and Firm Fundamentals
Panel A: Correlation matrix between DQ and firm fundamentals (Pearson upper triangle,
Spearman lower triangle)
DQ Restructure M&A SI σ (RET) log(AT) log(NSEG)
DQ – 0.398 0.128 0.091 0.018 0.192 −0.024
Restructure 0.361 – 0.112 0.157 0.020 0.244 0.068
M&A 0.121 0.112 – 0.029 −0.081 0.298 0.075
SI 0.203 0.320 0.118 – 0.173 −0.046 −0.020
σ (RET) 0.032 0.007 −0.102 0.124 – −0.324 −0.125
log(AT) 0.174 0.235 0.292 0.141 −0.399 – 0.289
log(NSEG) −0.054 0.065 0.071 0.042 −0.180 0.280 –
accuracy and bid-ask spreads. We then relate DQ to the cost of equity. In all
our validation tests, we include control variables for firm fundamentals so as
to isolate the association between the discretionary component of DQ and
established measures of information quality. We include industry, two-digit
SIC codes, and year FEs to control for unobserved industry and year ef-
fects. Additionally, we cluster standard errors by industry and year to guard
against the effects of nonfixed (temporary) correlations between variables
within industries and years.28
28 We also estimate all models using industry times year fixed effects, which allows for more
flexible industry-specific trends and shocks, and find our results on DQ are robust to these
controls.
1038 S. CHEN, B. MIAO, AND T. SHEVLIN
29 For example, Duru and Reeb [2002] document that analyst forecasts are less accurate
for firms that are internationally diversified. Haw, Jung, and Ruland [1994] find that analyst
forecast accuracy decreases sharply after mergers. Lehavy, Li, and Merkley [2011] show that
communication complexity (or the inverse of overall readability of corporate 10-K filings)
reduces analyst forecast accuracy and increases forecast dispersion.
A NEW MEASURE OF DISCLOSURE QUALITY 1039
TABLE 4
DQ and Analyst Forecast Properties
Predicted Signs Dependent Variable = DISP Dependent Variable = |FE|
Intercept ? 7.725 4.255 27.564∗ 16.689
(1.52) (0.80) (1.92) (1.06)
DQ − −18.934∗∗∗ −17.305∗∗∗ −42.703∗∗ −39.13∗∗
(−3.05) (−2.82) (−2.47) (−2.26)
σ (EPS) + 1.099∗∗∗ 1.067∗∗∗ 2.563∗∗∗ 2.375∗∗∗
(7.11) (6.97) (6.60) (6.35)
GROWTH + 2.902∗∗∗ 2.072∗∗∗ 8.675∗∗∗ 6.085∗∗
(3.86) (2.85) (4.02) (3.53)
ROA − −38.568∗∗∗ −40.619∗∗∗ −94.993∗∗∗ −95.19∗∗∗
(−6.77) (−7.14) (−6.54) (−6.04)
log(AF) − −4.655∗∗∗ −4.800∗∗∗ −19.215∗∗∗ −19.516∗∗∗
(−5.83) (−5.94) (−7.11) (−7.20)
Log(AT) + 2.953∗∗∗ 3.389∗∗∗ 8.391∗∗∗ 9.458∗∗∗
(6.44) (6.88) (7.16) (7.30)
Control for firm NO YES NO YES
fundamentals
Ind and Year Included Included Included Included
Fixed effects
NOBS 31,202 31,202 31,202 31,202
Adjusted R 2 0.179 0.186 0.084 0.087
DISP i,t+1 {|F E |i,t+1 } = γ1 + γ2 DQ i,t + γ3 σ (EPS)i,t + γ4 GROWTH i,t + γ5 ROAi,t + γ6 log (AF )i,t
+ γ7 log (AT )i,t + fundamenals i,t + Ind FE + Yr FE + εi,t (1)
The sample consists of 31,202 firm-years with at least three analyst forecasts of annual earnings from
1976 to 2011. DISP, |FE|, ROA, and GROWTH are winsorized at the extreme 1% and 99%. Standard errors
are two-way clustered by year and industry. Clustered T-statistics are in parentheses.
All coefficients are multiplied by 100 for exposition convenience. Variable definitions (all other variables
are defined as in the notes to table 3): DISPi, t +1 = forecast dispersion, measured as the average of standard
deviation of analyst forecast of year t+1 earnings sampled at each month over year t; |FE| i,t+1 = forecast
accuracy, measured as the average of the mean absolute forecast error of year t+1 earnings samples at each
month of year t; σ (EPS) = decile ranks of earnings volatility, measured as standard deviation of EPS over
yeat t − 4 to year t, deflated by share price at the end of year t; GROWTH = average percentage growth
in sales over year t − 4 to year t; ROA = income before extraordinary items divided by total assets; AF =
number of analysts issuing EPS forecasts for the current year.
∗ ∗∗ ∗∗∗
, , indicate significance levels at 10%, 5%, and 1% (two-tailed).
30 Daily quoted bid-ask spread is calculated as the average of all midquote-deflated bid-
ask spreads, 0.5∗(Ask-Bid)/(Ask + Bid), quoted during regular trading hours (9:30–16:00).
Effective bid-ask spread measures the difference between the actual execution price and the
midpoint of the prevailing quote, and is calculated as (Price − MidQuote)/MidQuote, where
MidQuote = (Bid + Ask)/2. Each trade is matched with the quote at the previous second, and
the effective bid-ask spreads of all trades during regular trading hour are averaged to obtain
the daily estimates.
A NEW MEASURE OF DISCLOSURE QUALITY 1041
TABLE 5
DQ and Bid-Ask Spread
Predicted Signs Dependent Variable = QBAS Dependent Variable = EBAS
Intercept ? 8.786∗∗∗ 8.418∗∗∗ 6.673∗∗∗ 6.414∗∗∗
(11.73) (11.69) (14.15) (14.10)
DQ ∓ −1.287∗∗ −1.260∗∗ −1.188∗∗∗ −1.173∗∗∗
(−2.32) (−2.35) (−2.94) (−3.01)
Log(VOL) – −0.518∗∗∗ −0.565∗∗∗ −0.232∗∗∗ −0.267∗∗∗
(−14.70) (−17.36) (−9.89) (−11.65)
Log(PRICE) – −1.369∗∗∗ −1.319∗∗∗ −0.875∗∗∗ −0.84∗∗∗
(−10.52) (−10.63) (−11.59) (−11.65)
BTM + 0.181∗∗∗ 0.176∗∗∗ 0.172∗∗∗ 0.168∗∗∗
(3.00) (2.84) (3.86) (3.65)
Log(AT) ? −0.104∗ −0.075 −0.202∗∗∗ −0.181∗∗∗
(−1.96) (−1.52) (−4.39) (−4.19)
Control for firm NO YES NO YES
fundamentals
Ind and Year Included Included Included Included
Fixed effects
NOBS 63,462 63,462 63,948 63,948
Adjusted R 2 0.597 0.599 0.557 0.560
QBAS/EBAS i,t+1 = β1 + β2 DQ i,t + β3 log (VOL)i,t + β4 log (PRICE)i,t + β5 BTM i,t + β6 log (AT )i,t
+ Fundamentals i,t + Ind FE + Yr FE + ui,t (2)
The sample consists of 63,462 firm-years from 1991 to 2011 for QBAS regressions and 63,948 firm-years
for EBAS regressions. BTM is winsorized at the extreme 1% and 99%. All coefficients are multiplied by 100
for exposition convenience. Standard errors are two-way clustered by year and industry. Clustered T-statistics
are in parentheses.
Variable definitions: QBAS = average daily quoted bid-ask spread over the 12-month period beginning
with 4 months after the end of current fiscal year. Daily quoted bid-ask spread is calculated as the average of
all bid-ask spreads, 0.5∗ (Ask−Bid)/(Ask+Bid), quoted during regular trading hour (9:30–16:00). Intraday
quotes data are from TAQ; EBAS = average daily effective bid-ask spread over the 12-month period begin-
ning with 4 months after the end of current fiscal year. Daily effective bid-ask spread is calculated using all
trades during regular trading hours (9:30–16:00). Trades are matched with prevailing quotes at the previous
second to estimate effective bid-ask spread using the equation (Price−MidQuote)/MidQuote, where MidQuote
= (Bid+Ask)/2. Intraday trades and quotes data are from TAQ; VOL = average daily trading volume over
year t; PRICE = average daily closing price over year t; BTM = book value of common equity divided by
market value of common equity. All other variables are defined as in notes to tables 3 and 4.
∗ ∗∗ ∗∗∗
, , indicate significance levels at 10%, 5%, and 1% (two-tailed).
and Verrecchia [2007], Kelly and Ljunqvist [2012]), and more specifically,
accounting information can affect cost of capital. Empirical research in ac-
counting has linked reporting quality, broadly defined, to lower cost of eq-
uity (Leuz and Verrecchia [2000], Lang and Lundhom [2000], Francis et al.
[2004], Hail and Leuz [2006], Daske et al. [2008], Francis, Nanda, and Ols-
son [2008], Ashbaugh-Skaiffe, Collins, and Kinney [2009], Baginski and
Rakow [2011], Daske et al. [2013]). Thus, we expect a negative association
between DQ and cost of equity.31 We estimate the following cost of equity
regression:
CofE i,t+1 = α1 + α2 DQ i,t + α3 Beta i,t + α4 BTM i,t + α5 log (MV )i,t
+ Fundamenals i,t + I nd F E + Y r F E + εi,t , (3)
31 We acknowledge that the literature on the causality between information quality and cost
of capital is controversial. Neither empirical nor theoretical research agrees on whether there
is a causal link (see Berger [2011], Shevlin [2013], section 4). We note that our research
purpose only calls for a demonstration of correlation, not causality.
1042 S. CHEN, B. MIAO, AND T. SHEVLIN
where cost of equity, CofE, is estimated using the average of three implied
cost of equity capital measures developed in prior literature (MPEG, GM,
and Claus and Thomas [2001]) evaluated in Botosan and Plumlee [2005]
and Easton and Monahan [2005]. The forecasts of future earnings in all
three methods are based on the approach proposed in Li and Mohanram
[2014] to address concerns that optimistic analysts’ forecasts lead to bi-
ased estimates of implied cost of capital (see Easton and Monahan [2005],
Kothari, Li, and Short [2009]). Following prior research (Francis et al.
[2004], Francis, Nanda, and Olsson [2008]), we include beta (Beta) esti-
mated using daily returns and the Scholes–Williams (Scholes and Williams
[1977]) adjustment method and book-to-market ratio (BTM).32 Under the
maintained assumption that higher disclosure quality should be associated
with lower cost of equity, we expect α 2 to be negative. As in Francis, Nanda,
and Olsson [2008], we expect positive coefficients on Beta and BTM.
The sample for this test consists of 35,474 firm-year observations from
1976 to 2011.33 Table 6 presents the cost of equity regression results. The
coefficient on the summary measure DQ is significantly negative at better
than 1% in both specifications, before and after we include controls for
firm fundamentals. The coefficient estimate of −0.057 suggests that a one
standard deviation increase in unconditional DQ of 0.113 is associated with
a decrease in CofE 0.6%. A one standard deviation increase in orthogonal-
ized DQ of 0.054 is associated with a decrease in CofE of 0.3%, which is still
nontrivial. Thus, higher DQ is associated with lower cost of equity capital.
Taken together, our three sets of validation tests offer strong evidence con-
sistent with DQ capturing disclosure quality.
5.4 ADDITIONAL ANALYSES: DISAGGREGATING DQ INTO COMPONENTS
Our research goal is to advance one summary measure of firms’ disclo-
sure quality, DQ, based on the level of disaggregation of items in firms’
annual reports covering both the body of the two financial statements and
associated footnotes. In this section, we report results disaggregating DQ
along two dimension: DQ arising from operating and from financing ac-
tivities (DQ OP and DQ FIN), and DQ disaggregated into Balance Sheet
and Income Statement components (DQ BS and DQ IS), respectively. Dis-
aggregation along these two dimensions serves as a starting point for future
researchers wishing to examine the components of DQ.
It is possible that for a typical firm there is more information asymmetry
about its operating performance than its financing decisions. If this is the
32 We do not include log(AT) in the cost of capital validation tests as we already include
log(MV) in all our models. The correlation coefficient between log(MV) and log(AT) exceeds
0.90 and would lead to severe multicollinearity if we include both variables in our models.
33 Untabulated descriptive statistics show that our cost of equity estimates yield statistics
comparable to prior literature (Easton and Monahan [2005]). The mean cost of equity, CofE,
is 0.132 with a median of 0.111. Though the sample in this test is much smaller than the sample
in table 1 due to data restrictions, the distributions of DQ and the component scores DQ BS
and DQ IS (untabulated) are very similar to that presented in panel A of table 1.
A NEW MEASURE OF DISCLOSURE QUALITY 1043
TABLE 6
DQ and Cost of Equity
Predicted Signs (1) (2)
∗∗∗
Intercept ? 19.060 15.355∗∗∗
(10.98) (9.27)
DQ – −5.591∗∗∗ −5.721∗∗∗
(−4.05) (−4.42)
Beta + 0.284∗ −0.166
(1.68) (−1.15)
BTM + 3.501∗∗∗ 3.737∗∗∗
(12.75) (12.75)
Log(MV) – −1.997∗∗∗ −1.821∗∗∗
(−7.76) (−7.43)
Control for firm fundamentals NO YES
Industry and Year Fixed Effects Included Included
NOBS 35,474 35,474
Adjusted R 2 0.414 0.432
CofE i,t+1 = α1 + α2 DQ i,t + α3 Beta i,t + α4 BTM i,t + α5 log (M V )i,t + Fundamenalsi,t
+ Ind FE + Yr FE + εi,t (3)
The sample consists of 35,474 firm-year observations from 1976 to 2011. All coefficients are multiplied
by 100 for exposition convenience. Standard errors are two-way clustered by year and industry. Clustered
T-statistics are in parentheses.
Variable definitions: CofE = average of implied cost of equity estimated using the MPEG, GM, and CT
methods. The forecasts of future earnings used in all three estimation methods are based on the approach
proposed in Li and Mohanram [2014]; Beta = CAPM beta estimated using the Scholes–Williams method
over the most recent calendar year ending before current fiscal year end; MV = market value of equity at
the current fiscal year end. All other variables are defined as in the notes to tables 3 and 4.
∗ ∗∗ ∗∗∗
, , indicate significance levels at 10%, 5%, and 1% (two-tailed).
case, then we would expect the link between established proxies for infor-
mation asymmetry and DQ to be primarily driven by the disaggregation of
operating items.34
To explore this possibility, we classify items representing operating and
financing activities into DQ OP (DQ representing operating activities)
and DQ FIN (DQ representing financing activities) using the classification
scheme advanced in Nissim and Penman [2001]. We equal weight each line
item and reestimate all three sets of validation tests on both components.
We present the results in panel A of table 7. For parsimony, we omit the
tabulation of all control variables. Because the samples used in each esti-
mation are identical to those for the main tests and the adjusted R2 are
nearly identical, we omit the tabulation of these two items as well.
Panel A results show that DQ OP drives the link with analyst forecast dis-
persion and accuracy, whereas DQ FIN drives the link with bid-ask spreads.
Both DQ OP and DQ FIN are significantly negatively associated with the cost
of equity. Panel B of table 7 presents the results separating DQ into DQ BS
and DQ IS. The results show that both components drive the link with bid-
ask spreads and cost of equity capital, but the link with analyst earnings
forecast properties is driven by DQ IS.
TABLE 7
Separating DQ into Operating (DQ OP) vs. Financing (DQ FIN) Components and Balance Sheet
(DQ BS) vs. Income Statement (DQ IS) Components
Panel A: Separating DQ into DQ OP and DQ FIN
Panel A1: Analyst forecasting properties (reestimating equation (1))
Dependent Dependent
Variable = DISP Variable = |FE|
Predicted Signs (1) (2) (1) (2)
DQ OP – −22.995∗∗∗ −19.409∗∗∗ −52.515∗∗ −45.033∗∗
(−3.05) (−2.57) (−2.47) (−2.02)
DQ FIN – −3.301 −4.107 −18.734 −20.674
(−0.89) (−1.11) (−1.43) (−1.57)
35 Untabulated analysis shows that 40% (30%) of the firms remains in the same DQ decile
classification in year t to year t+1 (t+2); in other words, 60% (70%) of the firms fall into a
different DQ decile in the following year.
TABLE 8
1046
(1) (2) (1) (2) (1) (2) (1) (2) (1) (2)
∗∗∗ ∗∗∗ ∗∗∗ ∗∗∗ ∗ ∗∗∗
−2.32 −1.90 −10.52 −9.53 −0.12 −0.062 −0.053 −0.014 −1.27 −1.26∗∗∗
(−3.13) (−2.52) (−3.31) (−2.98) (−1.47) (−0.78) (−1.01) (−0.27) (−3.97) (−4.00)
N = 7,592 N = 7,592 N = 7,592 N = 7,592 N = 14,157 N = 14,157 N = 14,081 N = 14,081 N = 9,450 N = 9,450
0.073 0.094 0.041 0.049 0.549 0.555 0.532 0.538 0.252 0.291
Panel A presents the coefficients (t-statistics) on DQ after adding firm fixed effects and removing industry fixed effects to equations (1)–(3). Column (1) models exclude controls
for firm fundamentals, whereas column (2) models include controls for firm fundamentals. The last two rows present the number of observations used in each regression and
adjusted R 2 . Panel B presents coefficients β 2 (firm clustered t-statistics) on the interaction variable DQINC×POST from the model: Dep. Var. = α + β 0 ∗DQINC +β 1 ∗ POST + β 2
∗(DQINC×POST) + Control Variables + Control for Fundamentals + error term. The dependent variables are DISP, |FE|, QBAS, EBAS, and COE, respectively. DQINC is an
indicator variable identifying changes in average DQ greater than 30% from years t − 2, t − 1 to years t, t+1, and t+2: if DQ increases by more than 30% over the five-year window,
DQINC = 1; if DQ decreases by more than 30% over the five-year window, DQINC = 0. POST is an indicator variable coded as 1 for years t, t+1, and t+2 and 0 for years t−2 and t−1.
The “Control Variables” are regression-specific controls employed in equations (1) through (3), respectively. Column (1) models exclude controls for firm fundamentals, whereas
column (2) models include controls for firm fundamentals. The last two rows present the number of observations used in each regression and adjusted R 2 .
∗ ∗∗ ∗∗∗
, , indicate significance levels of 10%, 5%, and 1% (one-sided).
A NEW MEASURE OF DISCLOSURE QUALITY 1047
DQ BS, and DQ IS by size quartile based on total assets. One striking pattern
emerges: all three DQ scores exhibit similar standard deviation across the
four size quartiles. Specifically, the standard deviation for the summary DQ
ranges from 0.123 to 0.125 across the size quartiles, and the range is 0.147–
0.152 for DQ BS and 0.118–0.127 for DQ IS. Note that these numbers are
very similar to those reported for the overall sample distribution in table 1.
Thus, the data does not support the assumption that larger firms dominate
the variation in DQ.
Furthermore, we note that our random sample check (detailed in section
3.1) shows that the Compustat coding error rate for the smallest size quin-
tile is smaller than that for the largest size quintile. This result again is in-
consistent with Compustat ignoring the collection of data on smaller firms
and spending more effort on the collection of data on the largest firms.
5.6.3. Do Existing Disclosure Measures Subsume DQ? Another concern re-
garding our disclosure score DQ is that it is subsumed by other existing
measures of disclosure, such as management earnings forecasts (MF) or
AIMR scores. We have argued that DQ is conceptually very different from
existing measures and demonstrated that DQ has low correlations with
these existing measures in table 2. As such, we do not expect that existing
measures can subsume DQ. Nevertheless, to alleviate the above concern, we
include the existing measures of disclosure as tabulated in table 2 in all our
validation regressions from tables 4 to 6. The sample sizes after the inclu-
sion of these existing disclosure measures are much smaller. In untabulated
analyses, we find that the results on DQ after the inclusion of these variables
are robust, while the correlations of these existing disclosure measures with
the information quality variables are either insignificant or inconsistent in
the presence of DQ.
5.6.4. Is DQ Simply the Inverse of Immaterial Items that Are Aggregated into
Other Items? Another potential concern with DQ is that, if managers sim-
ply aggregate immaterial items into other items, DQ will be capturing the
inverse of immaterial items. Following this logic, managers aggregate the
missing items into other items because the missing items are deemed im-
material and therefore unimportant by managers. If this is true, the quality
of financial reporting should not be adversely affected by the exclusion of
such immaterial items, and we should not expect higher DQ to represent
higher disclosure quality.
Our collective evidence from the three sets of validation tests does not
support the above argument. In addition, the mean magnitude of the miss-
ing items backed out in section 5.6.1 represents 17% of total assets. This
nontrivial amount is inconsistent with the concern that missing items are
immaterial items.
6. Conclusion
We develop a new measure of disclosure quality, DQ, which captures the
level of disaggregation of accounting line items in firms’ annual reports,
A NEW MEASURE OF DISCLOSURE QUALITY 1049
APPENDIX A
Compustat Template
(“Balancing Model”) for the Balance Sheet and Income Statement
Item Description Balancing Mnemonic
ASSETS
Current Assets
Current Assets - Total ACT
Non-Current Assets
Property Plant and Equipment - Total PPENT
(Net)
Investment and Advances - Equity IVAEQ
Investment and Advances - Other IVAO
Intangible Assets - Total INTAN
Assets - Other - Total AO
Assets - Total ACT + PPENT + IVAEQ + AT
IVAO + INTAN + AO
LIABILITIES & SHAREHOLDES’ EQUITY
Current Liabilities
Current Liabilities - Total LCT
Long-Term Liabilities
Long-Term Debt - Total DLTT
Deferred Taxes and Investment Tax Credit TXDITC
Liabilities - Other LO
Liabilities - Total LCT + DLTT + TXDITC + LO LT
Noncontrolling Interest - Redeemable - MIB
Balance Sheet
Shareholders’ Equity
Preferred/Preference Stock (Capital) - PSTK
Total
Common/Ordinary Equity - Total CEQ
Stockholders Equity - Parent - Total PSTK + CEQ SEQ
Noncontrolling Interest - Nonredeemable MIBN
- Balance Sheet
Stockholders Equity - Total SEQ + MIBN TEQ
Liabilities and Stockholders Equity - Total LT + MIB + TEQ LSE
(Continued)
A NEW MEASURE OF DISCLOSURE QUALITY 1051
APPENDIX A—Continued
Item Description Balancing Mnemonic
Sales/Turnover (Net if Excise Tax TXW) SALE
Operating Expenses - Total COGS + XSGA XOPR
Cost of Goods Sold COGS
Selling, General and Administrative XSGA
Expenses
Depreciation and Amortization - Total DP
Interest and Related Expense XINT
Nonoperating Income (Expense) - Total IDIT + NOPIO NOPI
Nonoperating Income (Expense) - NOPIO
Excluding Interest Income
Interest Income - Total IDIT
Special Items SPI
Pretax Income OIADP − XINT + NOPI + SPI PI
Income Taxes − Total TXT
Income Taxes - Current TXFED + TXS + TXFO + TXO TXC
Income Taxes - Deferred TXDFED + TXDS + TXDFO TXDI
Noncontrolling Interest - Income Account MII
Income Before Extraordinary Items IB
Dividends - Preferred/Preference DVP
Income Before Extraordinary Items - IB − DVP IBCOM
Available for Common
Extraordinary Items and Discontinued XI + DO XIDO
Operations
Extraordinary Items (including XI
Accounting Changes CCHG)
Discontinued Operations DO
Net Income (Loss) IBADJ + XIDO NIADJ
∗
Note that even though the group account CITOTAL (Comprehensive Income – Total) is not on Com-
pustat’s Income Statement Balancing Model, we classify the associated accounts as income statement ac-
counts rather than balance sheet accounts.
1052 S. CHEN, B. MIAO, AND T. SHEVLIN
APPENDIX B
The Three Level Structure to Link Subaccounts to Group Accounts
SUBACCOUNTS PARENT ACCOUNTS GROUP ACCOUNTS
[93] [25] [11]
Item 1
Item 2
Item 3 Parent account A
Item 4 Group Account #1
Item 5
Item 6 Parent account B
Item 7
…… …… ……
…… …… Group Account #11
TOTAL ASSETS
EXAMPLE
REFERENCES
AMIRAM, D.; E. OWENS; AND O. ROZENBAUM. Do Information Releases Increase or Decrease
Information Asymmetry? New Evidence from Analyst Forecast Announcements. Working
paper, Columbia University and University of Rochester, 2013.
ASHBAUGH-SKAIFFE, H.; D. COLLINS; W. KINNEY, JR.; AND R. LAFOND. “The Effect of SOX Inter-
nal Control Deficiencies on Firm Risk and Cost of Equity.” Journal of Accounting Research 47
(2009): 1–43.
BAGINSKI, S., AND K. RAKOW. “Management Earnings Forecast Disclosure Policy and the Cost
of Equity Capital.” Review of Accounting Studies 17 (2011): 279–321.
BENS, D.; P. BERGER; AND S. MONOHAN. “Discretionary Disclosure in Financial Reporting: An
Examination Comparing Internal Firm Data to Externally Reported Segment Data.” The
Accounting Review 86 (2011): 417–49.
BERGER, P. “Challenges and Opportunities in Disclosure Research—A Discussion of ‘The Fi-
nancial Reporting Environment: Review of the Recent Literature.’” Journal of Accounting &
Economics 51 (2011): 204–18.
BERGER, P., AND R. HANN. “The Impact of SFAS No. 131 on Information and Monitoring.”
Journal of Accounting Research 41 (2003): 163–223.
BERGER, P., AND R. HANN. “Segment Profitability and the Proprietary and Agency Costs of
Disclosure.” The Accounting Review 82 (2007): 869–906.
BEYER, A.; D. COHEN; T. LYS; AND B. WALTHER. “The Financial Reporting Environment: Review
of the Recent Literature.” Journal of Accounting & Economics 50 (2010): 296–343.
BLACKWELL, D. “Comparisons of Experiments.” in Proceedings of the Second Berkeley Symposium in
Mathematical Statistics and Probability, edited by J. Neyman. Berkeley and Los Angeles: Uni-
versity of California Press, 1951.
A NEW MEASURE OF DISCLOSURE QUALITY 1053
BOTOSAN, C. “Disclosure Level and the Cost of Capital.” The Accounting Review 72 (1997): 323–
49.
BOTOSAN, C., AND M. PLUMLEE. “Assessing Alternative Proxies for the Expected Risk Pre-
mium.” The Accounting Review 80 (2005): 21–53.
CLAUS, J., AND J. THOMAS. “Equity Premia as Low as Three Percent? Evidence from Analysts’
Earnings Forecasts for Domestic and International Stock Market.” The Journal of Finance 56
(2001):1629–66.
COLLER, M., AND T. YOHN. “Management Forecasts and Information Asymmetry: An Examina-
tion of Bid-Ask Spreads.” Journal of Accounting Research 35 (1997): 181–92.
COLLINS, D.; E. MAYDEW; AND I. WEISS. “Changes in the Value-Relevance of Earnings and Book
Values Over the Past Forty Years.” Journal of Accounting & Economics 24 (1997): 39–67.
DASKE, H.; L. HAIL; C. LEUZ; AND R. VERDI. “Mandatory IFRS Reporting Around the World:
Early Evidence on the Economic Consequences.” Journal of Accounting Research 46 (2008):
1085–142.
DASKE, H.; L. HAIL; C. LEUZ; AND R. VERDI. “Adopting a Label: Heterogeneity in the Economic
Consequences Around IAS/IFRS Adoptions.” Journal of Accounting Research 51 (2013): 495–
548.
DEMSETZ, H. “The Cost of Transacting.” Quarterly Journal of Economics 82 (1968): 33–53.
DHALIWAL, D.; S. RADHAKRISHNAN; A. TSANG; AND Y. YANG. “Nonfinancial Disclosure and An-
alyst Forecast Accuracy: International Evidence on Corporate Social Responsibility Disclo-
sure.” The Accounting Review 87 (2012): 723–59.
D’SOUZA, J.; K. RAMESH; AND M. SHEN. “Disclosure of GAAP Line Items in Earnings Announce-
ments.” Review of Accounting Studies 15 (2010): 179–219.
DURU, A., AND D. REEB. “International Diversification and Analysts’ Forecast Accuracy and
Bias.” The Accounting Review 77 (2002): 415–33.
EASLEY, D., AND M. O’HARA. “Information and the Cost of Capital.” The Journal of Finance 59
(2004): 1553–83.
EASTON, P., AND S. MONAHAN. “An Evaluation of Accounting-Based Measures of Expected Re-
turns.” The Accounting Review 80 (2005): 501–38.
ERTIMUR, Y.; J. LIVNAT; AND M. MARTIKAINEN. “Differential Market Reactions to Revenue and
Expense Surprises.” Review of Accounting Studies 8 (2003): 185–211.
FAIRFIELD, P.; R. SWEENEY; AND T. YOHN. “Accounting Classification and the Predicative Con-
tent of Earnings.” The Accounting Review 71 (1996): 337–55.
FRANCIS, J.; R. LAFOND; P. OLSSON; AND K. SCHIPPER. “Cost of Equity and Earnings Attributes.”
The Accounting Review 79 (2004): 967–1010.
FRANCIS, J.; R. LAFOND; P. OLSSON; AND K. SCHIPPER. “The Market Pricing of Accruals Quality.”
Journal of Accounting & Economics 39 (2005): 295–327.
FRANCIS, J.; D. NANDA; AND P. OLSSON. “Voluntary Disclosure, Earnings Quality, and Cost of
Capital.” Journal of Accounting Research 46 (2008): 53–99.
GJESDAL, F. “Accounting for Stewardship.” The Accounting Review 19 (1981): 208–31.
HAIL, L., AND C. LEUZ. “International Differences in the Cost of Equity Capital: Do Legal
Institutions and Securities Regulation Matter?” Journal of Accounting Research 44 (2006): 485–
531.
HAW, I.; K. JUNG; AND W. RULAND. “The Accuracy of Financial Analysts’ Forecasts After Merg-
ers.” The Journal of Accounting, Auditing, and Finance 9 (1994): 465–83.
HEALY, P.; A. HUTTON; AND K. PALEPU. “Stock Performance and Intermediation Changes Sur-
rounding Sustained Increases in Disclosure.” Contemporary Accounting Research 16 (1999):
485–520.
HEWITT, M. “Improving Investors’ Forecast Accuracy When Operating Cash Flows and Accru-
als Are Differentially Persistent.” The Accounting Review 84 (2009): 1913–31.
HIRST, E.; L. KOONCE; AND S. VENKATARAMAN. “How Disaggregation Enhances the Credibility
of Management Earnings Forecasts.” Journal of Accounting Research 45 (2007): 811–37.
HOPE, O. “Disclosure Practices, Enforcement of Accounting Standards, and Analysts’ Forecast
Accuracy: An International Study.” Journal of Accounting Research 41 (2003): 235–72.
1054 S. CHEN, B. MIAO, AND T. SHEVLIN
HUGHES, J.; J. LIU; AND J. LIU. “Information Asymmetry, Diversification, and Cost of Capital.”
The Accounting Review 82 (2007): 705–29.
JEGADEESH, N., AND J. LIVNAT. “Revenue Surprises and Stock Returns.” Journal of Accounting &
Economics 41 (2006): 147–71.
KELLY, M., AND A. LJUNGQVIST. “Testing Asymmetric-Information Asset Pricing Models.” Review
of Financial Studies 25 (2012): 1366–413.
KLINE, R. Principles and Practice of Structural Equation Modeling, Third edition (Methodology in
the Social Sciences series). New York, NY: The Gulf Press, 2014.
KOH, P., AND D. REEB. “Missing R&D.” Journal of Accounting & Economics 60 (2015): 73–94.
KOTHARI, S. P.; X. LI; AND J. SHORT. “The Effect of Disclosures by Management, Analysts, and
Financial Press on Cost of Capital, Return Volatility, and Analyst Forecasts: A Study Using
Content Analysis.” The Accounting Review 82 (2009): 1255–97.
LAMBERT, R.; C. LEUZ; and R. VERRECCHIA. “Accounting Information, Disclosure, and the Cost
of Capital.” Journal of Accounting Research 45 (2007): 385–420.
LANG, M., AND R. LUNDHOLM. “Cross-Sectional Determinants of Analyst Ratings of Corporate
Disclosures.” Journal of Accounting Research 31 (1993): 246–71.
LANG, M., AND R. LUNDHOLM. “Corporate Disclosure Policy and Analysts Behavior.” The Ac-
counting Review 71 (1996): 467–92.
LANG, M., AND R. LUNDHOLM. “Voluntary Disclosure and Equity Offerings: Reducing Informa-
tion Asymmetry or Hyping the Stock?” Contemporary Accounting Research 17 (2000): 632–62.
LEHAVY, R.; F. LI; AND K. MERKELEY. “The Effect of Annual Report Readability and Analyst
Following and the Properties of Their Earnings Forecasts.” The Accounting Review 86 (2011):
1087–115.
LEUZ, C., AND R. VERRECCHIA. “The Economic Consequence of Increased Disclosure.” Journal
of Accounting Research 38 (2000): 91–124.
LI, F. “Annual Report Readability, Current Earnings, and Earnings Persistence.” Journal Ac-
counting Research 45 (2008): 221–47.
LI, K., AND P. MOHANRAM. “Evaluating Cross-Sectional Forecasting Models for Implied Cost of
Capital.” Review of Accounting Studies 19 (2014): 1152–85.
MCVAY, S. “Earnings Management Using Classification Shifting: An Examination of Core Earn-
ings and Special Items.” The Accounting Review 81 (2006): 501–31.
NISSIM, D., AND S. PENMAN. “Ratio Analysis and Equity Valuation: From Research to Practice.”
Review of Accounting Studies 6 (2001): 109–54.
SCHOLES, M., AND J. WILLIAMS. “Estimating Betas from Nonsynchronous Data.” Journal of Fi-
nancial Economics 5 (1977): 309–27.
SHEVLIN, T. “Some Personal Observations on the Debate on the Link Between Financial Re-
porting Quality and the Cost of Equity Capital.” Australian Journal of Management 38 (2013):
447–73.
STOLL, H. R. “The Supply of Dealer Services in Security Markets.” Journal of Finance 33 (1978):
1133–51.
Supporting Information
Additional Supporting Information may be found in the online version of
this article at the publisher’s website:
Internet Appendix A. Linking Table for the Balance Sheet
Internet Appendix B. Linking Table for Income Statement
Internet Appendix C. An Example for Constructing Value-Weighted DQ BS