is silence golden? earnings warnings and subsequent
TRANSCRIPT
Is Silence Golden? Earnings Warnings and
Subsequent Changes in Analyst Following
JENNIFER WU TUCKER*
Skinner (1994) proposes that managers have reputational incentives topreempt negative earnings news and speculates that firms that fail towarn are less likely to be followed by financial analysts. My study for-mally puts forward this argument and empirically tests the reputationalcost of withholding bad news in the form of a decrease in analyst follow-ing. I find that among firms with a similar level of existing disclosurereputation, those that fail to warn experience a significant decrease inanalyst following relative to those that warn. This finding is consistentwith managers’ concerns that withholding bad news would damage theirreputation for being transparent about forward-looking earnings news.
1. Introduction
Anticipating an earnings disappointment near or after the end of a fiscal
quarter, some managers warn while others do not. The popular explanation for
earnings warnings is the litigation argument: Managers preempt bad news to
avoid litigation or reduce legal costs (Skinner [1994, 1997]). A reputation-related
explanation, however, has received little attention, even though Skinner (1994)
acknowledges throughout his paper that his findings are consistent with both the
litigation argument and the reputation argument. Skinner (1994, 40) suggests,
‘‘Managers may also have reputational incentives to preempt negative earnings
news.’’ Based on anecdotal evidence, he speculates that firms that fail to warn
are less likely to be followed by financial analysts. My study formally puts
*Fisher School of AccountingWarrington College of BusinessUniversity of FloridaThis paper is developed from my dissertation completed at the New York University. I am
indebted to Jim Ohlson (Chair), Bill Greene, Carol Marquardt, Stephen Ryan, and Paul Zarowin fortheir guidance and constant support. I thank Rowland Atiase, Orie Barron, Joel Demski, Rajib Doo-gar, late Pat Hughes, Baruch Lev, Doug Skinner, Phil Stocken, Surjit Tinaikar, Greg Waymire, PeteWilson, an anonymous referee, and the participants of 2005 AAA Mid-Year FARS Conference.I thank Thomson Financial, especially Pamela Grant, for the First Call Company-Issued Guidelines(CIG) data and the Luciano Prida, Sr. Term Professorship Foundation for financial support.
Keywords: Earnings Warning, Disclosure Reputation, Analyst Following,
Voluntary Disclosure
431
forward the reputation argument and examines it by testing the change in analyst
following after a firm’s failure to warn.
‘‘Reputation’’ has been loosely used by researchers and by the press in vari-
ous circumstances. In information economics, ‘‘reputation’’ is a concept used only
in a multiperiod game. It is the probability assessment of a player’s ‘‘type’’ or
‘‘attribute.’’ This assessment is updated after each period of the game.1 ‘‘Reputa-
tion’’ is specific to the attribute at issue. For example, a person may have a repu-
tation for being punctual for work but a reputation for being tardy for parties.
As Wilson (1985) succinctly summarizes, whenever a player’s type is not
fully known to others, his strategy inevitably will be affected by his concern
about others’ assessment of his type—his reputation. As a result, he may adopt a
strategy for a multiperiod game that he would not use in a one-period game. For
example, a manager may not warn about bad news in a one-period game, but
warn in a multiperiod game because he is concerned about the effect of having a
reputation for being opaque on the behavior of financial analysts in the subse-
quent periods.
The role of reputation has been emphasized by survey studies (Gibbins,
Richardson, and Waterhouse [1990]; King, Pownall, and Waymire [1990]). In a
survey of chief financial officers (CFOs), Graham, Harvey, and Rajgopal (2005,
54) report that ‘‘92.1 percent of the survey respondents believe that developing a
reputation for transparent reporting is the key factor motivating voluntary disclo-
sures.’’ Their study further notes, ‘‘76.8 percent of the respondents agree or
strongly agree that disclosing bad news faster enhances the firm’s reputation for
transparent and accurate reporting’’ (Graham, Harvey, and Rajgopal, 2005, 65). De-
spite such importance being placed on reputation, the archival evidence on the role
of reputation is scant and Skinner’s reputation argument for bad-news disclosure
(hereinafter ‘‘the reputation argument’’) remains untested.2 My study fills this gap.
The underlying assumption of the reputation argument is that managers’ con-
cerns about their reputations for being transparent about forward-looking earnings
news (hereinafter ‘‘reputation’’) propel them to warn the market of an anticipated
earnings shortfall.3 For example, if a firm enjoys a stellar reputation, the man-
ager may disclose the bad news that he has to assure market participants that the
firm is indeed transparent. If a firm’s reputation is low, the manager may strive
to build a reputation by issuing a warning. In a world in which managers and
analysts (investors) both hold rational expectations (Muth [1961]), the following
must be true for reputation to be a valid concern to managers: among firms with
1. See Kreps and Wilson (1982) and Milgrom and Roberts (1982) for the seminal reputationmodels and Diamond (1989) and Stocken (2000) for the applications of reputation models in the capital-market setting.
2. Williams (1996) and Hutton and Stocken (2006) examine a firm’s reputation for forecast ac-curacy. For bad-news disclosures, forecast accuracy or credibility is typically not an issue (Mercer[2005]); instead, the issue is whether to disclose—that is, transparency.
3. Throughout the paper, I assume that a manager’s interest is aligned with the firm’s interestand use the manager’s reputation and the firm’s reputation interchangeably.
432 JOURNAL OF ACCOUNTING, AUDITING & FINANCE
a high reputation, those that fail to warn will be perceived as less transparent
than those that warn; among firms with a low reputation, those that warn will be
perceived as more transparent than those that do not warn.
As a researcher, I do not observe that analysts update their beliefs about a
firm’s transparency after a disclosure decision. I do observe, however, the change
in analysts’ interest in following a firm after the disclosure decision. Previous
studies suggest that analysts largely play the role of public-information interme-
diaries, that is, they mainly collect, digest, synthesize, and disseminate public in-
formation. For example, Francis, Schipper, and Vincent (2002) and Frankel,
Kothari, and Weber (2006) find that analysts’ reports complement firms’ public
disclosures. Lang and Lundholm (1996) find high analyst coverage for high dis-
closure rating firms and Healy, Hutton, and Palepu (1999) find a decrease in ana-
lyst coverage for firms whose disclosure ratings have declined. Acting mainly as
public-information intermediaries, analysts would be less interested in following
a firm that has become less forthcoming, such as withholding bad news.
I hypothesize that firms that fail to warn will experience a decrease in ana-
lyst following relative to those that have similar reputations before the disclosure
event and warn. Comparing warning and nonwarning firms across different pre-
event reputations does not test the reputation argument. For example, a warning
firm with an already high reputation cannot further improve the reputation.
A nonwarning firm with an already low reputation cannot further damage its rep-
utation. Comparing these two groups would lead to a wrong conclusion that the
disclosure decision does not matter when in fact it matters.
For the empirical tests, I collect a group of warnings and a group of non-
warning events in the post–Regulation Fair Disclosure (post–Reg FD) era. Both
groups have a disappointing quarter, that is, the earnings realization (assuming
managers have private signals about this realization) is lower than the market ex-
pectation near the end of the fiscal quarter. The warning group warns about this
earnings disappointment; the nonwarning group does not warn. I refer to the dis-
appointing quarter as the ‘‘event quarter,’’ the four preceding quarters as ‘‘pre-
event,’’ and the four subsequent quarters as ‘‘postevent.’’ I examine the change
in analyst following from the pre-event to the postevent period.4
I find in the multivariate test that relative to warning firms with a similar
pre-event reputation, nonwarning firms experience a significant decrease in ana-
lyst following. This result is obtained after I control for firms’ performance, the
change in other disclosure activities from pre-event to postevent, and the estima-
tion bias from firms’ self-selection into the warning and nonwarning groups. Here
the control for self-selection is crucial because a firm may be more likely to warn
for unobserved reasons that also discourage analysts from following the firm.
A decrease in analyst following is a cost to firms. Analyst coverage makes a
firm better known to investors and generates investors’ interest in the stock.
4. I use four quarters in each period because analyst coverage varies within a reporting cycle offour quarters (O’Brien and Bhushan [1990]).
433IS SILENCE GOLDEN?
Financial theories on information risk argue that high analyst coverage likely
reduces information asymmetry, which in turn reduces the cost of capital (Easley
and O’Hara [2004], 1573). Empirical studies have found that analyst following
improves stock liquidity and firm value (Chung and Jo [1996]; Healy, Hutton,
and PalepuHutton [1999]; Roulstone [2003]). Thus, my finding of a decrease in
analyst following after a firm’s failure to warn suggests a reputational cost to
firms for withholding bad news.
The rest of this paper is organized as follows. Section 2 develops the
research hypothesis and Section 3 describes the data. Section 4 presents the em-
pirical model and Section 5 reports the test results. Section 6 examines the sub-
periods. Section 7 concludes.
2. Hypothesis
2.1 Direct Implications of the Reputation Argument
The reputation argument implies that a firm will be viewed as less transparent
when bad news is withheld than when it is disclosed. Suppose discrete measures
are used for reputation—high, medium, and low, which is often the case in prac-
tice. The reputation argument implies the following. First, if a firm with an already
high pre-event reputation warns, the reassessed probability of its being transparent
(hereinafter ‘‘perceived transparency’’) will remain high; if it fails to warn, the per-
ceived transparency will decrease. Second, if a firm with a medium pre-event rep-
utation warns, the perceived transparency will increase but will decrease if the
firm fails to warn. Third, if a firm with a low pre-event reputation warns, the per-
ceived transparency will increase but will remain low if the firm fails to warn.
2.2 Testable Implications of the Reputation Argument
Because analysts’ beliefs are unobservable, I use the change in analyst fol-
lowing after the warning decision to infer the change in a firm’s transparency as
perceived by analysts. Analysts’ interest in following a firm could be affected by
their assessment of the firm’s transparency in one of two ways. If analysts
mainly act as public-information intermediaries, a decrease in perceived transpar-
ency would discourage analysts from following the firm. If analysts are mainly
private information generators, a decrease in perceived transparency would en-
courage analysts to follow the firm.
The decision for an analyst to follow a firm depends on the costs and bene-
fits of following the firm. A major benefit of following a firm is to generate
commission fees for the brokerage house because, thanks to the analyst who fol-
lows the firm, investors trade stocks with which they are familiar. Another bene-
fit of following a firm is to know the firm well so that the brokerage house will
have the opportunity to underwrite the firm’s future equity issues and consult the
firm on mergers and acquisitions.
434 JOURNAL OF ACCOUNTING, AUDITING & FINANCE
Analysts incur costs in their activities. When they are public-information
intermediaries, they collect, digest, integrate, and disseminate market, industry,
and firm-specific public information. In doing so, analysts take advantage of their
financial expertise, the economy of scale in information gathering, and the syn-
ergy of covering multiple firms in one industry. These activities thrive on corpo-
rate disclosures and analysts’ reports complement such disclosures. A decrease in
a firm’s transparency would discourage analysts from following the firm. When
analysts search for private information, they visit stores and factories and inter-
view a firm’s major customers, suppliers, and alliances. A decrease in transpar-
ency would encourage analyst following because corporate disclosures likely
preempt analysts’ disclosure of private information.
Most archival evidence suggests that analysts mainly act as public-information
intermediaries. Ivkovic and Jegadeesh (2004) report that about 20 to 26 percent of
analyst recommendations and revisions of earnings estimate are issued at the earn-
ings announcement date and the following two days, even though these three days
account for only 5 percent of the sixty-day examination window. This observation
suggests that analysts’ reports complement corporate disclosure. Francis, Schipper,
and Vincent (2002) find that the magnitude of market reaction to analysts’ reports
and the magnitude of market reaction to firms’ earnings announcements are posi-
tively correlated. This result is confirmed by Frankel, Kothari, and Weber (2006),
suggesting the complementary nature of analysts’ activities and corporate disclo-
sures. Furthermore, Lang and Lundholm (1996) and Healy, Hutton, and Palepu
(1999) report that analyst following is higher for firms that have higher or
increased disclosure ratings. These studies strongly support the public-information-
intermediary role of financial analysts. Thus, I predict the following:
H1: Nonwarning firms will experience a decrease in analyst following relative
to warning firms that have similar pre-event reputations. In particular,H1a: Among firms with a high reputation, those that fail to warn will experi-
ence a decrease in analyst following relative to those that warn.H1b: Among firms with a medium reputation, those that fail to warn will ex-
perience a decrease in analyst following relative to those that warn.H1c: Among firms with a low reputation, those that warn will experience an
increase in analyst following relative to those that do not warn.
3. Data
I collect earnings warnings from the First Call Company-Issued Guidelines
(CIG) database (see Table 1). Before Reg FD, which was enacted on October 23,
2000, firms that did not warn publicly might have warned analysts privately. To
maintain a good reputation with analysts, managers may resort to either private or
public warnings. Given that private warnings are unobservable, testing the effects
of reputation on analyst following before Reg FD is not feasible. Therefore, I use
post–Reg FD disclosure data (including the four pre-event quarters), resulting in
435IS SILENCE GOLDEN?
TABLE 1
Sample Collection
Panel A: Sample collection of warning events
Procedures Change Events
Negative guidance about quarterly earnings for calendarized quarters
2001Q4–2005Q1 identified in the First Call CIG database.
5,892
Exclude negative guidance about quarterly earnings issued before the
beginning of the third fiscal month.
�3,819 2,073
Exclude warnings that do not have the unique identifying variables
for Compustat, CRSP, and I/B/E/S.
�156 1,917
Exclude duplicate warnings for the same fiscal quarter. �51 1,866
Exclude warnings issued since three days before the earnings
announcement date.
�25 1,841
Exclude warnings for which the recent analyst consensus forecast
before the warning is unavailable.
�5 1,836
Exclude warnings for which the split-adjusted stock price at the
beginning of the event quarter is less than $1.
�8 1,828
Exclude extremely small news (i.e., the price-deflated earnings
surprise is above �0.001).
�311 1,517
Warning sample 1,517
Panel B: Examples of earnings warnings
Example 1—EGL
‘‘Shares of EGL Inc (EAGL.O) fell 16 percent on Tuesday, a day after the company issued an earn-
ings warning that also stung the shares of other logistics providers. The Houston transport group said
late on Monday that costlier air cargo service from Asia and weak demand for air deliveries would
hurt third-quarter earnings. EGL said it expected share profits for the quarter of 12 cents, down from
its earlier forecast of 18 cents to 20 cents.’’
Source: Reuters News, ‘‘EGL Shares Fall 16 pct on Earnings Warning,’’ October 28, 2003.
Example 2—American Eagle
‘‘Shares of American Eagle Outfitters Inc. fell to their lowest levels in 10 months on Thursday after
the casual-wear retailer warned that quarterly earnings would be lower than many analysts expected.
After the market closed on Wednesday, the company said disappointing sales in May and June would
reduce second-quarter earnings to a range of 14 cents to 17 cents a share. Analysts’ average forecast
is 19 cents a share, according to Thomson First Call, with estimates ranging from 14 cents to 23
cents.’’
Source: Reuters News, ‘‘American Eagle Shares Fall after Earnings Warning,’’ July 11, 2002.
2001Q4–2005Q1 (calendarized) being my sample period.5 During this period,
U.S. companies issued 5,892 negative guidelines about quarterly earnings.
Figure 1 shows that firms’ negative quarterly guidance arrives in two waves.
The first wave centers at the end of the first fiscal month (i.e., around the previ-
ous-quarter earnings announcement date) and I refer to such guidance as ‘‘fore-
casts.’’ The second wave is around the fiscal quarter-end and, consistent with
Wall Street, I refer to such guidance as ‘‘warnings.’’ The key difference between
these two types of disclosures is the amount of private information that managers
have at the time of disclosure. ‘‘Forecasts’’ require both a manager’s ability to
predict early in the quarter and his willingness to share the prediction with the
public. At the time of the disclosure decision, the fiscal quarter is almost over
and the manager should know how the firm has fared, and thus ‘‘warnings’’
require mainly a manager’s willingness to share.
I use warnings rather than forecasts to test the reputation argument for two
reasons. First, what underlies the reputation argument is ‘‘transparency,’’ that is,
the willingness to disclose. The absence of a warning indicates a manager’s
unwillingness to disclose, whereas the absence of a forecast indicates either a
FIGURE 1
Firms’ Voluntary Disclosure of Negative Earnings News
Source: The data source is the First Call Company-Issued Guidelines (CIG) database.
Note: First Call collects company disclosures from press releases and interviews, compares a disclo-
sure with existing market expectations, and codes it as D for negative news for the variable CIG-CODED. The figure shows the timing of negative earnings news issued in my sample period after I
exclude 178 observations for which the news is issued 180 days before the fiscal quarter-ends and 20
observations for which the news is issued 90 days after the fiscal quarter.
5. To better control for time trends in the tests, I calendarize the quarters: A firm’s fiscal quar-ter is labeled to the calendar quarter with which it overlaps most. For example, fiscal quarters thatend in May, June, and July belong to the second calendar quarter.
437IS SILENCE GOLDEN?
manager’s inability to predict early in the quarter or his unwillingness to dis-
close.6 Second, my examination of warnings is consistent with a series of recent
studies, such as Kasznik and Lev (1995); Soffer Thiagarajan, and Walther
(2000); and Atiase, Supattarakul, and Tse (2006).7 Therefore, I retain the 2,073
warnings from the initial sample.
The sample cleaning is as follows. I exclude 156 events whose unique iden-
tifiers in Compustat (‘‘gvkey’’), Center for Research in Security Prices (CRSP)
(‘‘permno’’), and Institutional Brokers’ Estimate System (I/B/E/S) (‘‘ticker’’) are
unavailable. Fifty-one firms issue a second warning for the same quarter, in
which case I keep the first warning because the first dose of warning is arguably
more alarming. I delete twenty-five warnings that are issued close to (i.e., within
three days) or after the event-quarter earnings announcement: the former have lit-
tle information value and the latter probably are caused by data errors. Among
the remaining warnings, five are excluded because no recent analyst consensus
from I/B/E/S is available and eight are dropped because their stock prices are
less than $1—small deflators may lead to outliers.
Finally, 311 warnings have the price-deflated earnings surprise between 0
and �0.001. The magnitude of such earning surprises is extremely small. These
earnings surprises are calculated by researchers using earnings realizations. At
the time of the disclosure decision, a manager may not be aware of such an
extremely small earnings shortfall unless he has perfect foresight. I exclude these
311 events and am left with 1,517 events as my final warning sample.
Nonwarning observations are identified as the firm-quarters for which the
forthcoming earnings are less than analysts’ most recent consensus before the
third fiscal month but for which the firms do not warn according to CIG. After
applying the same screening procedures as for the warning sample, I am left with
9,057 nonwarning events. Table 2 shows the industry distributions of warning
and nonwarning events. The business services, retail, and chips industries have
the highest number of warnings, while the percentage of warning firms is the
highest in the retail, consumer goods, and steel industries.
4. Empirical Model
4.1 Dependent Variable
I measure the dependent variable ChgFollow as the change in average ana-
lyst following from pre-event to postevent. Analyst following in each quarter is
6. If a firm issues an unfavorable earnings forecast early in the quarter, analysts would revisetheir earnings estimates downward and consequently the earnings surprise calculated using the expec-tation later in the quarter is probably non-negative. The firm is then neither in my warning nor non-warning group.
7. These studies examine disclosures issued in the sixty-day window before the quarterly earn-ings announcement date. This window is comparable to mine because earnings typically areannounced twenty-five to thirty days after the fiscal quarter end.
438 JOURNAL OF ACCOUNTING, AUDITING & FINANCE
the number of analysts whose earnings estimates are included in the most recent
I/B/E/S consensus compiled before the earnings announcement date. If a firm-
quarter is covered by Compustat but not by I/B/E/S, analyst following for this
firm-quarter is zero.
4.2 Explanatory Variables and Coefficient Predictions
My explanatory variables are the dummy variable Warn (one for a warning
observation and zero for a nonwarning observation), the proxies for pre-event
reputation, and the interaction terms between Warn and these proxies. A firm’s
pre-event reputation is unobservable; I proxy it by the disclosure frequency in
the pre-event period on the ground that pre-event reputation should be highly
positively correlated with the number of actual forward-looking earnings disclo-
sures issued by the company. To measure this frequency, I identify earnings
guidance for either forthcoming fiscal years or quarters in the CIG database. In
my sample, the median of this frequency is one, the seventy-fifth percentile is
four, and the ninetieth percentile is seven. Of the sample, 46.8 percent do not
issue any guidance in the pre-event period.
TABLE 2
Industry Distribution
Top Ten Industries with the Largest Number of Warnings, 2001Q4–2005Q1
Industry Warning Nonwarning Total Relative Frequency
Business Services 229 1,053 1,315 17.9%
Retail 215 384 599 35.9%
Chips 114 561 690 16.9%
Computers 76 370 460 17.0%
Banks 48 867 916 5.2%
Steel Works 44 145 194 23.3%
Machinery 41 274 318 13.0%
Wholesale 41 240 282 14.6%
Consumer Goods 40 88 129 31.3%
Medical Equipment 38 264 306 12.6%
Other 631 4,811 5,442 11.6%
Total 1,517 9,057 10,574 14.3%
Note:
1. Warning and nonwarning firms experience an earnings disappointment in the event quarter.
A warning (nonwarning) observation is a firm-quarter for which, according to First Call, the firm
publicly communicated (did not communicate) the disappointment to investors between the begin-
ning of the third fiscal month and three days before the earnings announcement.
2. I use the Fama and French (1997) industry classifications, which were then modified by Professor
French on his Web site.
439IS SILENCE GOLDEN?
I place firms that issue no disclosures in the pre-event period into the ‘‘low’’
reputation group. The dummy variable DisLow is one for these firms and zero for
the others. I place firms that issue four or more disclosures in the pre-event period
into the ‘‘high’’ reputation group. The dummy variable DisHigh is one for these
firms and zero for the others. I place the remaining firms into the ‘‘medium’’ rep-
utation group.8 In my sample, 25.8 percent of the firms are in the high-reputation
group, 27.4 percent are in the medium-reputation group, and 46.8 percent are in
the low-reputation group. Hypotheses H1a, H1b, and H1c are then tested by eq. (1):
ChgFollowi¼b0þb1Warniþb2DisHighiþb3Warni �DisHighi
þb4DisLowiþb5Warni �DisLowiþcontrol variablesþlI ð1Þ
The following illustration shows that testing Hypotheses H1a, H1b, and H1cis
equivalent to testing three sets of coefficients. In Figure 2, the symbol ‘‘X’’ indi-
cates that the referred coefficient is a component of the change in analyst follow-
ing for a particular group. For example, the average change in analyst following
for nonwarning firms in the high pre-event reputation group is (b0þ b2), while
the average change in analyst following for warning firms in the same group is
(b0þ b1þ b2þ b3). To test Hypothesis H1a—that is, whether among the firms
with a high pre-event reputation, nonwarning firms experience a significant
decrease in analyst following relative to warning firms—I test whether (b1þb3) > 0. Similarly, the test for Hypothesis H1b is whether b1 > 0 and the test for
Hypothesis H1c is whether (b1þ b5) > 0.
I also compare the effects of withholding bad news across firms of differentpre-event reputations. The average change in analyst following for nonwarning
firms with a high pre-event reputation is (b0þ b2), whereas the average change
in analyst following for nonwarning firms with a medium pre-event reputation
is b0. Thus, a negative b2 would indicate a higher penalty for silence to firms
with a high pre-event reputation than for those with a medium pre-event reputa-
tion. Similarly, a positive b4 would indicate a higher penalty for silence to
firms with a medium pre-event reputation than to those with a low pre-event rep-
utation.
4.3 Control Variables
I control for five factors that affect analyst coverage: (1) performance, (2)
earnings variability, (3) the change in a firm’s other disclosure activities (rather
than the warning event), (4) the level of analyst following before the event quar-
ter, and (5) self-selection.
8. In robustness tests, I find no change in the conclusions if the cutoff of three or five is usedfor the medium and high reputation groups.
440 JOURNAL OF ACCOUNTING, AUDITING & FINANCE
4.3.1 Performance
Prior studies suggest that analysts tend to follow firms that are performing
well. Chung and Jo (1996, 496) argue that ‘‘more analysts follow high quality
firms than low quality firms because brokers find it easier to market stocks of high
quality firms.’’ Hayes (1998) demonstrates analytically and McNichols and
O’Brien (1997) find empirical evidence that analysts initiate coverage of firms that
have good earnings prospects and drop those with lackluster earnings performance,
because a trade on a poorly performing stock is more likely to be a sale and the
trading volume generated by a sale is limited to investors’ initial holdings when
short sale is restricted. Thus, analysts potentially can generate more trading com-
missions from well-performing stocks than from poorly performing stocks.
I control for performance by including (1) the event-quarter earnings disap-
pointment (Surprise), (2) the number of quarters in the pre-event period for
which the firm fails to meet or beat analyst expectations (PastMiss), (3) the
change in core earnings from the pre-event to the postevent period (ChgEPS),
and (4) the event-quarter stock performance (Return).
I use Surprise to control for the severity of the negative earnings news in
the event quarter and include its quadratic term to allow for curvature in the rela-
tion. For a warning observation, Surprise is the difference between the forthcom-
ing earnings per share (EPS) and the most recent analyst consensus before the
warning (both from I/B/E/S). For a nonwarning observation, the variable is the
difference between the forthcoming EPS and the last analyst consensus before
FIGURE 2
Empirical Predictions
Pre-event
Reputation
Disclosure
Decision b0 b1 b2 b3 b4 b5
High Warn X X X X
Not Warn X X
Medium Warn X X
Not Warn X
Low Warn X X X X
Not Warn X X
Predictions:
H1a: (b1þ b3) > 0
H1b: b1 > 0
H1c: (b1þ b5) > 0
Note: X indicates that the referred coefficient is a component of the change in analyst following for a partic-
ular group. For example, the average change in analyst following for nonwarning firms with a high pre-
event reputation is (b0þb2).
441IS SILENCE GOLDEN?
the third fiscal month of the event quarter. As in Kasznik and Lev (1995), Sur-prise is deflated by the split-adjusted beginning-of-event-quarter stock price.9
Note that Surprise is negative for all sample firms. I expect a positive coefficient
on Surprise (i.e., a larger decrease in analyst following after a more negative
value of Surprise) and a positive coefficient on the quadratic term (i.e., the curve
flattens out for a large magnitude of bad news).
Analysts may react to a firm’s poor performance with a delay, so I use Past-Miss to control for recent lackluster earnings performance. PastMiss is the num-
ber of earnings-disappointing quarters in the pre-event period. A quarter is
disappointing if a company’s realized earnings are lower than the most recent an-
alyst consensus before the third fiscal month (both from I/B/E/S).10 I use the
consensus before the third fiscal month instead of the consensus just before the
earnings announcement as the expectation benchmark. I do not use the latter
because analyst expectations could have been revised downward after a public
warning and the firm would appear to have met or beat consensus on the earn-
ings announcement date even though the quarter is truly disappointing. I expect
a negative coefficient on PastMiss.I control for the change in average core earnings from pre-event to postevent
(ChgEPS). Core earnings are the realized EPS recorded by I/B/E/S, which
removes special items. The earning change is deflated by the beginning-of-the-
event-quarter stock price. I predict a positive coefficient on ChgEPS.
Finally, I control for stock performance. Return is the buy-and-hold stock
return from the third fiscal month of the event quarter to five days after the event-
quarter earnings announcement, less the buy-and-hold return of an equal-weighted
market index over the same period. I expect a positive coefficient on Return.
4.3.2 Earnings Variability
I control for the change in earnings variability from pre-event to postevent
(ChgEarnVolt). Bhushan (1989) argues that analysts follow firms that have high
performance variability so that analysts can add value to the earnings-predicting
activity, presumably by exerting their superior information-processing abilities. I
measure earnings variability for the pre-event and postevent periods using the av-
erage absolute seasonal-differenced, split-adjusted diluted EPS (from Compustat)
9. In this study, all earnings-related variables are deflated by stock price to control for the scal-ing difference in EPS. To avoid outliers created by small deflators, I drop the observations whosedeflator is less than one.
10. If a quarter is not covered by analysts, the quarter is not marked as ‘‘disappointing.’’ Thus,my measure of PastMiss is lower for firms with lower analyst coverage in the pre-event period.Because I control for pre-event analyst following in the multivariate test, this measurement issueshould not be a concern.
442 JOURNAL OF ACCOUNTING, AUDITING & FINANCE
in that period.11 The variable is deflated by the split-adjusted beginning-of-event-
quarter stock price and I predict a positive coefficient.
4.3.3 Other Disclosure Activities
I control for the change in firms’ other forward-looking disclosure activities.
Warning is only one disclosure event, although a significant one. Analyst follow-
ing may change as a result of the change in a firm’s other disclosure activities
after the warning event. As discussed in Section 2, an increase in other corporate
disclosures should attract analysts. I include ChgDis, measured as the change in
the total number of company guidance events recorded in the CIG database from
pre-event to postevent and predict a positive coefficient.
4.3.4 Pre-Event Analyst Following
The change in analyst following is perhaps nonlinear to the pre-event level.
An increase in analyst following is more likely for firms with a low level to
begin with than for those with a high level. Similarly, a decrease in analyst fol-
lowing is more likely for firms with a high level to begin with than for those
with a low level. I control for PreFollow, measured as the average analyst fol-
lowing in the pre-event quarters. Analyst following in each quarter is the number
of analysts whose estimates are included in the most recent consensus before the
quarterly earnings announcement. I predict a negative coefficient on PreFollow.
4.3.5 Estimation Bias from Self-Selection
Some unobservable reasons may make firms more likely to warn. For exam-
ple, a new chief executive officer (CEO) may be more likely to warn to blame
the departing CEO for the shortfall. If these reasons also affect analysts’ deci-
sions to continue to cover a firm (e.g., the analyst’s contact is no longer with the
firm), a portion of the observed change in analyst coverage is attributable to the
effect of unobservable reasons. I control for this estimation bias from self-selec-
tion by using a modified treatment-effect model. A frequently used method in
the accounting literature to control for self-selection bias is the treatment-effect
model described by Greene (2003). Under this method, researchers estimate a
first-stage choice model, calculate the inverse Mills ratio for the event and none-
vent firms (the calculations are slightly different for the two groups), and then
add this variable to the regression of interest in the second stage. Tucker (2007)
modifies the standard treatment-effect model by allowing the event and nonevent
11. I do not use the realized earnings in I/B/E/S to construct earnings variability because sea-sonal-differencing cannot be performed when a firm’s quarter or its seasonal quarter in the previousyear is not covered by analysts.
443IS SILENCE GOLDEN?
groups to each have its own coefficient on its respective inverse Mills ratios.
She also provides concrete interpretation for the inverse Mills ratios and their
coefficients.
I estimate a probit model of warning versus nonwarning decisions using the
specifications in Tucker (2007). For brevity, the variable definitions and estima-
tion results are provided in Appendix A. Let Z denote the row vector of the
observable variables capturing managers’ considerations in the warning decision
and g be the column vector of the estimated coefficients. The inverse Mills
ratio (Mill) for a warning observation is calculated as /ðZicÞ=UðZicÞ, and that
for a nonwarning observation is calculated as ½�/ðZicÞ=ð1� UðZicÞÞ�. Here, f(.)
and F(.) are the p.d.f. and c.d.f. of the standard normal distribution, respectively.
After adding the quarterly dummies to control for time trends in analyst fol-
lowing, I sum up the regression in the following equation:
ChgFollowi¼b0þb1Warniþb2DisHighiþb3Warni�DisHighiþb4DisLowi
þb5Warni�DisLowiþb6Surpriseiþb7Surprise2iþb8PastMissi
þb9ChgEPSiþb10Returniþb11ChgEarnVoltiþb12ChgDisi
þb13 PreFollowiþb14Milli�Warniþb15Milli�ð1�WarniÞþquarterly dummiesþli ð2Þ
5. Empirical Results
After applying the data requirement for the first-stage warning choice model
(Appendix A), 1,481 warning and 8,504 nonwarning observations remain. In this
section, I first present the descriptive statistics associated with the test of analyst
following changes. I then report the primary test results. Finally, I briefly discuss
alternative model specifications.
5.1 Descriptive Statistics
Panel A of Table 3 presents the number and percentage of warning and non-
warning firms in each pre-event reputation group. The high-reputation group has
807 warning firms and 1,774 nonwarning firms. The medium-reputation group
has 486 warning firms and 2,249 non-warning firms. The low-reputation group
has 188 warning firms and 4,481 nonwarning firms. Overall, the majority of
warning firms (54.5%) belong to the high-reputation group, whereas only 20.9
percent of nonwarning firms are from the high-reputation group. In contrast, the
majority of nonwarning firms issue no forward-looking earnings disclosures in the
pre-event period (i.e., low-reputation group), whereas only 12.7 percent of warn-
ing firms are silent (or opaque) in the pre-event period. Thus, most warning firms
have a high reputation to begin with and most nonwarning firms are opaque
444 JOURNAL OF ACCOUNTING, AUDITING & FINANCE
before the event quarter. It appears that in the face of bad news managers issue a
warning to maintain the existing high reputation.
Panel B of Table 3 reports the descriptive statistics of the dependent variable,
ChgFollow, and the control variables for this sample. The change in analyst fol-
lowing for nonwarning firms has a mean of �0.04 and the change for warning
firms has a mean of �0.23. The change for nonwarning firms is more favorable
TABLE 3
Descriptive Statistics
Panel A: Warning and nonwarning firms by pre-event reputation
Pre-event Reputation
Warning Nonwarning Subtotal
Obs. % Obs. % Obs. %
High 807 54.5% 1,774 20.9% 2,581 25.8%
Medium 486 32.8% 2,249 26.4% 2,735 27.4%
Low 188 12.7% 4,481 52.7% 4,669 46.8%
Total 1,481 100% 8,504 100% 9,985 100%
Panel B: Sample means and medians
Warning Nonwarning
Obs. Mean Median Obs. Mean Median
ChgFollow 1,481 �0.23 �0.25 8,504 �0.04 0
Surprise 1,481 �0.011 �0.005 8,504 �0.017 �0.005
PastMiss 1,481 0.77 0 8,504 0.33 0
ChgEPS 1,445 �0.006 �0.002 8,036 0.000 0.000
Return 1,481 �0.169 �0.154 8,500 �0.083 �0.079
ChgEarnVolt 1,469 �0.003 0.000 8,416 �0.118 �0.000
ChgDis 1,481 �0.80 0 8,504 �0.27 0
PreFollow 1,481 7.27 6 8,504 4.57 3
PreDisclosure 1,481 5.24 4 8,504 1.89 0
Note: Table 3 uses 1,481 warning and 8,504 nonwarning firms that have available data for the warning
choice model in Appendix A. Firms with four or more earnings guidelines in the pre-event period are classi-
fied into the ‘‘high’’ reputation group. Firms with one to three earnings guidelines in the pre-event period
are classified into the ‘‘medium’’ reputation group. Firms with no earnings guidelines in the pre-event period
are classified into the ‘‘low’’ reputation group.
Variable Definitions:
ChgFollow is the change in average analyst following from the four pre-event to the four postevent
quarters.
Surprise: For a warning observation, the variable is the difference between the forthcoming EPS and
the most recent analyst consensus before the warning (both from I/B/E/S). For a nonwarning observation,
the variable is the difference between the forthcoming EPS and the last analyst consensus before the third
fiscal month of the event quarter. The variable is deflated by the split-adjusted beginning-of-event-quarter
stock price.
(continued)
445IS SILENCE GOLDEN?
than the change for warning firms (between-group t-test statistic ¼ 2.50; Wil-
coxon z-statistic ¼ 2.54, unreported). This difference, however, could be due to
warning firms experiencing a more permanent decline in future earnings
(ChgEPS, between-group t-test statistic ¼ �5.00; Wilcoxon z-statistic ¼ �9.62,
unreported) or more nonearnings bad news (Return, between-group t-test statistic ¼�15.53; Wilcoxon z-statistic ¼ �15.80, unreported).12 It would be rather hasty to
conclude from the univariate comparison that withholding bad news helps firms
keep analysts.
5.2 Primary Results
I report the multivariate test of the full model in the last column of Table 4
with the coefficients on the quarterly dummies suppressed. The estimation is ro-
bust to possible violations of the normality assumption of the error term. More
important, it is robust to outliers in both the dependent and independent variables
by automatically, in each iteration, setting aside influential observations and
down-weighting observations with large residuals.
The coefficient on Warn, b1, is 1.398, both statistically (t ¼ 6.05) and eco-
nomically significant, indicating that among firms with a medium pre-event repu-
tation, nonwarning firms suffer from a loss of 1.398 analysts relative to warning
firms. This finding supports Hypothesis H1b. The sum of coefficients for Warnand Warn*DisHigh, (b1þb3), is 1.433 with the Wald-test F statistic being
PastMiss is the number of earnings-disappointing quarters in the four pre-event quarters by comparing
realized earnings with the most recent analyst consensus before the third fiscal month.
ChgEPS is the change in average realized EPS (recorded by I/B/E/S) from the four pre-event to the
four postevent quarters, deflated by the split-adjusted beginning-of-event-quarter stock price.
Return is the buy-and-hold stock return from the third fiscal month of the event quarter to five days af-
ter the event-quarter earnings announcement, less the buy-and-hold market return.
ChgEarnVolt is the change in average absolute seasonal-differenced diluted EPS from the four pre-
event quarters to the four postevent quarters, deflated by the split-adjusted beginning-of-event-quarter stock
price.
ChgDis is the change in the total number of company guidance events, recorded in the First Call CIG
database, from the pre-event to the postevent period.
PreFollow is the average analyst following in the four pre-event quarters. Analyst following in each
quarter is the number of analysts whose forecasts are included in the most recent consensus before the quar-
terly earnings announcement.
PreDisclosure is the total number of company guidance events (about both fiscal year and fiscal quar-
ters) in the pre-event period, according to the CIG database.
TABLE 3 (Continued)
12. On average, the magnitude of the event-quarter earnings shortfall is smaller for nonwarningfirms than for warning firms, although the medians are similar.
446 JOURNAL OF ACCOUNTING, AUDITING & FINANCE
TABLE 4
Multivariate Test
ChgFollowi¼b0þb1Warniþb2DisHighiþb3Warni �DisHighiþb4DisLowiþb5Warni �DisLowi
þb6Surpriseiþb7Surprise2iþb8PastMissiþb9ChgEPSiþb10Returni
þb11ChgEarnVoltiþb12ChgDisiþb13 PreFollowi
þb14Milli �Warniþb15Milli �ð1�WarniÞþquarterlydummiesþli
Model 1 Model 2 Model 3 Full Model
Intercept 0.568*** 0.501*** 0.470*** 0.426***
(8.35) (7.14) (6.69) (5.41)
Warn 0.188*** 1.143*** 1.078*** 1.398***
(3.60) (7.10) (4.12) (6.05)
DisHigh �0.129**
(�2.05)
Warn*DisHigh 0.035
(0.27)
DisLow 0.090*
(1.85)
Warn*DisLow 0.187
(1.15)
PreDisclosure �0.041***
(�4.14)
Warn*PreDisclosure 0.045***
(2.93)
Surprise 9.813*** 10.264*** 10.842*** 10.599***
(12.54) (13.10) (13.64) (13.46)
Surprise2 10.234*** 10.584*** 11.092*** 10.900***
(9.85) (10.19) (10.61) (10.47)
PastMiss �0.157*** �0.171*** �0.158*** �0.148***
(�7.76) (�8.38) (�7.63) (�6.96)
ChgEPS 1.425*** 1.441*** 1.454*** 1.423***
(4.76) (4.82) (4.87) (4.76)
Return 0.685*** 0.711*** 0.735*** 0.720***
(7.79) (8.08) (8.33) (8.18)
ChgEarnVolt �0.003 �0.004 �0.004 �0.004
(�0.72) (�0.86) (�0.93) (�0.89)
ChgDis 0.051*** 0.064*** 0.057*** 0.060***
(7.91) (9.53) (7.97) (8.75)
PreFollow �0.069*** �0.084*** �0.085*** �0.084***
(�18.86) (�20.19) (�20.19) (�20.25)
Mill*Warn �0.560*** �0.507*** �0.692***
(�5.24) (�3.39) (�5.18)
Mill*(1-Warn) �0.660*** �1.178*** �0.878***
(�5.35) (�6.93) (�6.30)
Adj. R2 9.7% 10.2% 10.2% 10.3%
(continued)
significant at 64.00, indicating that when the pre-event reputation is high, a fail-
ure to warn is associated with a decrease of 1.433 analysts. This finding supports
Hypothesis H1a. The sum of coefficients for Warn and Warn*DisLow, (b1þ b5),
is 1.585 with the Wald-test F statistic being significant at 32.37, suggesting that
among firms with a low pre-event reputation, those that warn experience an
increase of 1.585 analysts relative to those that do not warn. This finding sup-
ports Hypothesis H1c. The joint test for all the three pre-event reputation levels
yields an F statistic of 21.59, statistically significant, supporting my overall hy-
pothesis (Hypothesis H1). The documented decrease in analyst following repre-
sents a reputational cost to firms for withholding bad news.
I compare the reputational cost for withholding bad news to firms of differ-ent pre-event reputations. The coefficient on DisHigh, b2, is significantly nega-
tive at �0.129 with a t-statistic of �2.05, suggesting that among nonwarningfirms those with a high reputation experience a larger loss of analysts than those
with a medium reputation. The coefficient on DisLow, b4, is weakly significantly
positive at 0.090 with a t-statistic of 1.85, suggesting that among nonwarningfirms, those with a medium reputation experience a (weakly) larger loss of ana-
lysts than those with a low reputation. These results imply that withholding bad
news is more costly to firms with a higher pre-event reputation than for those
with a lower reputation.
Most of the coefficients on the control variables have the expected sign. As
predicted, analyst following decreases for firms with poor past performance
Note:
1. The estimation uses 1,445 warning (Warn¼ 1) and 8,033 nonwarning (Warn¼ 0) observations that
have available data. The estimation requires two steps to control for self-selection bias. In the first
step, I estimate the warning-choice probit model (Appendix A) and calculate the inverse Mills ratio
(Mill). In the second step, I add Mill to the analyst coverage regression, allowing the warning and
nonwarning groups to each have its own coefficient.
2. I report the robust-regression estimations with the coefficients on the quarter dummies suppressed.
Such estimations are robust to outliers in the dependent and independent variables by automatically,
in each iteration, setting aside influential observations and downweighting observations with large
residuals. This method does not assume normality and theoretically possesses about 95 percent of
the efficiency of ordinary least squares (OLS).
3. ***, **, and * indicate statistical significance at 1 percent, 5 percent, and 10 percent in a two-tailed
test, respectively. The t-statistics are in the parentheses.
Variable Definitions
ChgFollow is the change in average analyst following from the pre-event to the postevent period.
DisHigh is one if the total number of earnings guidelines in the pre-event period (i.e., Predisclosure)
is four or higher and zero otherwise.
DisLow is one if the firm issues no earnings guidelines in pre-event period (i.e. Predisclosure) and
zero otherwise.
See Table 3 for other variable definitions.
TABLE 4 (Continued)
448 JOURNAL OF ACCOUNTING, AUDITING & FINANCE
(Surprise, PastMiss, and Return) and a deterioration in future performance
(ChgEPS). Analyst following increases when firms increase other disclosures
(ChgDis). Contrary to my prediction, the change in earnings variability is not
associated with changes in analyst following.
The coefficient on the inverse Mills ratio is significantly negative for both
the warning and nonwarning groups. The negative coefficient for the warning
group suggests that, within this group, some unobservable reasons make a firm
more likely to warn and at the same time discourage analysts from following the
firm (see Tucker [2007]). The negative coefficient for the nonwarning group sug-
gests that within this group some unobserved reasons make a firm less likely to
warn and at the same time encourage analysts to follow the firm. The significance
of both coefficients indicates the importance of controlling for self-selection.13
5.3 Alternative Model Specifications
The first three columns of Table 4 report the test results of alternative model
specifications (quarterly dummies are included in the estimations but suppressed
in the table). In the first column, I leave out the pre-event reputation proxies and
the terms for correcting self-selection estimation bias. The coefficient on Warn is
significantly positive at 0.188 with a t-statistic of 3.60, indicating that, on aver-
age, analyst following decreases for nonwarning firms relative to warning firms.
Recall that in Panel B of Table 3 the univariate comparison of ChgFollow yields
the opposite result. The different conclusion in the multivariate test shows the
importance of controlling for performance, other disclosure activities, and the
mean-reverting property of analyst following.
The second column adds the inverse Mills ratios to the regression. Mill is
significantly negative for both the warning and nonwarning groups. After addi-
tionally controlling for self-selection bias, the coefficient on Warn is significantly
positive at 1.143 with a t-statistic of 7.10.14 These results suggest that the control
for self-selection substantially improves the test results, although it does not
change the conclusion.
The third column uses a simple count of pre-event-period disclosures, Pre-Disclosure, rather than the trichotomous variables DisHigh and DisLow. The
coefficient on Warn is significantly positive at 1.078 with a t-statistic of 4.12,
13. The results are similar if the standard treatment-effect model is used. I also estimate themodel without Mill. The sum of coefficients on Warn and Warn*DisHigh, (b1þb3), is 0.278 with theWald-test F statistic being significant at 13.65, indicating that when the pre-event reputation is high,a failure to warn is associated with a decrease of about 0.3 analysts. The coefficient on Warn and thesum of Warn and Warn*DisLow coefficients are both positive but statistically insignificant, suggest-ing no difference in analyst following between the warning and nonwarning groups when firms havemedium or low reputations. The weaker evidence confirms the importance of controlling for self-selection estimation bias.
14. If a standard treatment-effect model were used, the coefficient on Warn would be 1.188with a t-statistic of 7.90, and the coefficient on Mill would be �0.603 with a t-statistic of �6.93.
449IS SILENCE GOLDEN?
suggesting that a firm with no disclosures in the pre-event period would have
about one more analyst after it warns than a similar firm that does not warn. The
coefficient on PreDisclosure is significantly negative at �0.041 with a t-statistic
of �4.14, suggesting that the more transparent a firm is before the event, the
larger the decrease in analyst following in the absence of a warning. If a firm
warns, however, the frequency of pre-event disclosures does not have an effect
on analyst following subsequent to the disclosure event, because the coefficient
on Warn*PreDisclosure of 0.045 (t-statistic ¼ 2.93) offsets the coefficient on
PreDisclosure of �0.041. The result that PredDisclosure does not matter to
warning firms likely is due to my constraining the relation between ChgFollowand PreDisclosure to be linear.
5.4 Disclosure Reputation and Stickiness of Corporate Guidance Practices
My paper focuses on firms’ disclosure reputation for being transparent—a
concept in a multiperiod game. Prior research has noted that corporate earnings
guidance practices are sticky (Anilowski, Feng, and Skinner [2007]). An interest-
ing question is to what extent a firm warns as a result of following its sticky
guidance practice. If a firm typically provides guidance at the time of the warn-
ing, the warning in the event quarter would be unsurprising. I collect earnings
guidance for the same quarter (as the warning-event quarter) in the prior year
issued during the window in which warnings typically are collected (i.e., from
the beginning of the third fiscal month to three days before the earnings
announcement date). I find that 4.6 percent of the sample firms issued such guid-
ance and warn in the event quarter, 8.8 percent issued such guidance but do not
warn in the event quarter, 9.8 percent did not issue such guidance but warn in
the event quarter, and 76.8 percent did not issue such guidance and do not warn
in the event quarter. These patterns suggest that most warnings are issued notbecause of the stickiness of a firm’s guidance practice.15
6. SubPeriods
I divide the sample into an early period and a recent period using the cutoff
of 2003Q1 for three reasons. First, although the Sarbanes-Oxley Act—the most
sweeping security regulation after the security laws of 1933 and 1934—was
15. Among the 1,517 warning firms, 484 (31.9%) issued guidance in the same window of thesame quarter in the prior year and 1,033 (68.1%) did not. For 302 (62.4%) warning firms that issuedsuch guidance, the warning in the event quarter is more than five days earlier or later than the timingof this prior guidance (When the timings of the two events are random, the chance of the warningto be within five days of prior guidance timing in the sixty-day warning-collection window is about10/60¼16.7%.) This observation again confirms that most warnings are issued not out of sticky guid-ance practices.
450 JOURNAL OF ACCOUNTING, AUDITING & FINANCE
passed in July 2002, most of the recommended changes took effect near or after
the first quarter of 2003. Examples of such changes were requiring at least one
audit committee member with financial expertise, transparent communications
about board operations, audit of internal control over financial reporting, and dis-
closures of off-balance-sheet arrangements and non–Generally Accepted
Accounting Principles financial measures. Such stringent regulations and
enhanced corporate governance have provided investors with additional monitor-
ing means. As a market-based monitoring tool, reputation is expected to play a
smaller role in the recent period when other monitoring tools are strong.
My second reason for dividing the sample is that some of the market partici-
pants might have adopted a different view of firms’ disclosure practices after
2003. On December 13, 2002, Coca-Cola publicly announced that the company
would no longer provide quarterly earnings guidance. The announcement had a
ripple effect: Several high-profile companies followed suit in early 2003, claim-
ing that the guidance cessation was intended to direct investors’ focus from
short-term results to long-term goals. As a result, in the mind-set of some, nondi-
sclosure is a virtue, not a sin (CFA Institute [2006]; Hsieh, Koller, and Rajan
[2006]; U.S. Chamber of Commerce [2007]). This change in attitudes may affect
analysts’ responses to warning versus nonwarning in the recent period.
My third reason for dividing the sample is that the early period is a bear
market and the recent period is a bull market. Veronesi (1999) argues that the
market handles bad news differently in a bear market than in a bull market. He
predicts that market participants are more disappointed with bad news in good
times than in bad times because bad news in good times likely signals a regime
shift: The good state is over. Conrad, Cornell, and Landsman (2002) find sup-
porting evidence for this prediction. Therefore, being informed of bad news in
good times could be more valuable than in bad times. Consequently, failures to
warn in a bull market may lead to more reputational damages than such failures
in a bear market.
Panels A and B of Table 5 provide the descriptive statistics for the early and
recent periods separately. The two periods differ in several variables, confirming
that the sample partition is justified. In the early period, analyst following
decreases in a similar degree for both warning and nonwarning firms, whereas in
the recent period the change for both groups is less negative and the change for
nonwarning firms is in fact positive. In addition, the stock returns for warning
firms (Return) are more negative in the early period than in the recent period,
even though the magnitude of earnings surprise is similar. This comparison sug-
gests that warning firms in the early period might have accumulated a greater
amount of nonearnings bad news and the release of bad news is less timely than
in the recent period.
Panel C of Table 5 reports the subperiod estimations. The test results are
similar to the full-sample results with one notable exception. The decreases in
analyst following for nonwarning firms relative to warning firms with a high,
medium, and low pre-event reputation are 2.396, 2.085, and 2.138 in the early
451IS SILENCE GOLDEN?
TABLE 5
Subperiod Tests
Panel A: Early period (2001Q4–2003Q1)
Warning Nonwarning
Obs. Mean Median Obs. Mean Median
ChgFollow 722 �0.30 �0.25 3,618 �0.37 �0.25
Surprise 772 �0.012 �0.006 3,618 �0.022 �0.005
PastMiss 772 0.85 0 3,618 0.44 0
ChgEPS 754 �0.006 �0.003 3,409 0.001 �0.001
Return 772 �0.189 �0.167 3,618 �0.098 �0.087
ChgEarnVolt 767 �0.006 0.000 3,592 �0.154 �0.001
ChgDis 772 �0.61 0 3,618 �0.26 0
PreFollow 772 7.05 5.75 3,618 4.87 3
PreDisclosure 772 4.54 3 3,618 1.90 1
Panel B: Recent period (2003Q2–2005Q1)
Warning Nonwarning
Obs. Mean Median Obs. Mean Median
ChgFollow 709 �0.16 0 4,886 0.20 0.17
Surprise 709 �0.009 �0.005 4,886 �0.014 �0.004
PastMiss 709 0.67 0 4,886 0.24 0
ChgEPS 691 �0.005 �0.001 4,627 �0.001 0.000
Return 709 �0.148 �0.142 4,882 �0.072 �0.076
ChgEarnVolt 702 0.001 0.001 4,824 �0.090 0.000
ChgDis 709 �1.01 0 4,886 �0.28 0
PreFollow 709 7.50 6 4,886 4.34 2.75
PreDisclosure 709 6.00 5 4,886 1.88 0
Panel C: Multivariate test
Early Period Recent Period
Intercept 0.427*** 0.391***
(4.32) (4.40)
Warn 2.085*** 1.238***
(6.05) (3.95)
DisHigh �0.270*** �0.097
(�2.63) (�1.21)
Warn*DisHigh 0.311 �0.280
(1.59) (�1.49)
DisLow 0.053 0.134**
(0.69) (2.15)
Warn*DisLow 0.293 �0.013
(1.26) (�0.06)
Surprise 11.320*** 17.184***
(10.50) (10.73)
(continued)
period, respectively. The corresponding numbers for the recent period are 1.238,
0.958, and 1.372. After stacking the subperiods with a dummy variable to distin-
guish the two sets of independent variables, I find that the differences above also
are statistically significant (F¼16.11, t¼1.89, and F¼4.42 for the high, medium,
and low reputation groups, respectively). These comparisons indicate that the
reputational cost for withholding bad news is higher in the early period than in
the recent period. This evidence is contrary to the regime-shifting argument, but
it is consistent with either tightened regulations or shifting attitudes against quar-
terly earnings guidance in the recent period.
7. Conclusion
My study fills the void in the literature by empirically testing whether firm
managers have reputation-related incentives to preempt negative earnings news.
I find that analyst following decreases for firms that fail to warn relative to those
that have similar pre-event disclosure reputations and warn. The loss of analyst
coverage represents a reputational cost to firms for withholding bad news. My
findings confirm managers’ concerns that withholding bad news would damage
their reputation with financial analysts. The evidence elevates the discussion of
reputation in both academia and practice.
Surprise2 10.936*** 41.849***
(8.44) (9.40)
PastMiss �0.141*** �0.157***
(�4.68) (�5.14)
ChgEPS 0.996** 3.880***
(2.52) (7.97)
Return 0.942*** 0.416***
(7.61) (3.29)
ChgEarnVolt �0.001 �0.010
(�0.19) (�1.46)
ChgDis 0.060*** 0.060***
(5.60) (6.74)
PreFollow �0.129*** �0.056***
(�19.83) (10.31)
Mill*Warn �1.015*** �0.527***
(�5.03) (�2.97)
Mill*(1-Warn) �1.976*** �0.516***
(�8.16) (�3.07)
Adj. R2 13.6% 8.1%
No. of Obs:
warn/nonwarn
754/3,409 691/4,624
Note: Please see the notes and variable definitions in Tables 3 and 4.
TABLE 5 (Continued)
453IS SILENCE GOLDEN?
APPENDIX A
Warning-Choice Probit Model
PrðWarni ¼ 1Þ ¼ Uða0 þ a1LitigRiski þ a2LogMVEi þ a3LogSurprisei
þ a4Forecasti þ a5QtrGuidancei þ a6Analysti þ a7M=Bi
þ a8EarnVolti þ eiÞ
REFERENCES
Anilowski, C., M. Feng, and D. J. Skinner. 2007. ‘‘Does Earnings Guidance Affect Market Returns?The Nature and Information Content of Aggregate Earnings Guidance.’’ Journal of Accountingand Economics 44 (1–2): 36–63.
Atiase R. K., S. Supattarakul, and S. Tse. 2006. ‘‘Market Reaction to Earnings Surprise Warnings:The Incremental Role of Shareholder Litigation Risk on the Warning Effect.’’ Journal ofAccounting, Auditing and Finance 21 (Spring): 191–222.
Coefficient z-statistic
Intercept �1.208 �11.83***
LitigRisk 0.241 3.18***
LogMVE 0.031 1.92***
LogSurprise 0.154 9.27***
Forecast 0.409 8.66***
QtrGuidance 0.132 17.01***
Analyst 0.024 5.49***
M/B 0.464 7.42***
EarnVolt �1.250 �6.52***
McFadden Pseudo R2 15.36%
Note: The model specification follows the choice model used in Tucker (2007).
*** indicates statistical significance at the 1 percent level in a two-tailed test. The numbers of warning
and nonwarning observations that have sufficient data are 1,481 and 8,504, respectively.
Variable Definitions:Warn is one for a warning observation and zero for a nonwarning observation. LitigRisk is the frac-
tional ranking (zero for the lowest and one for the highest) of the likelihood of being sued, calculated out of
sample using the estimated coefficients of the litigation-risk model in Appendix B of Houston, Lev, and
Tucker (2008) with the input variables measured in the pre-event period. LogMVE is the natural logarithm
of market value of equity at the beginning of the event quarter (in millions). LogSurprise is the natural loga-
rithm of the absolute value of event-quarter earnings surprise. For a warning observation, earnings surprise
is the difference between the forthcoming EPS and the most recent analyst consensus before the warning
(both from I/B/E/S). For a nonwarning observation, it is the difference between the forthcoming EPS and
the last analyst consensus before the third fiscal month of the event quarter. Surprise is deflated by the split-
adjusted beginning-of-event-quarter stock price. Forecast is one if a firm has issued an earnings forecast
about the event quarter before the third fiscal month for warning firms and before the consensus date in com-
puting Surprise for nonwarning firms and zero otherwise. QtrGuidance is the number of guidelines about
quarterly earnings issued by a firm in the 360 days before the event quarter. Analyst is the average number
of analysts whose earnings forecasts are included in the most recent consensus before quarterly earnings
announcements in the pre-event period. M/B is the fractional ranking of the market-to-book ratio at the be-
ginning of the event quarter (zero for the lowest and one for the highest). EarnVolt is the absolute change in
quarterly earnings from the same quarter in the previous year, averaged in the pre-event period.
454 JOURNAL OF ACCOUNTING, AUDITING & FINANCE
Bhushan, R. 1989. ‘‘Firm Characteristics and Analyst Following.’’ Journal of Accounting and Eco-nomics 11: 255–274.
CFA Institute. 2006. Breaking the Short-Term Cycle. CFA Institute Publications. Codes, Standards,and Position Papers.
Chung, K. H., and H. Jo. 1996. ‘‘The Impact of Security Analysts’ Monitoring and Marketing Func-tions on the Market Value of the Firms.’’ Journal of Financial and Quantitative Analysis 31,no. 4 (December): 493–512.
Conrad, J., B. Cornell, and W. R. Landsman. 2002. ‘‘When is Bad News Really Bad News?’’ Journalof Finance 57 (6): 2507–2532.
Diamond, D. W. 1989. ‘‘Reputation Acquisition in Debt Markets.’’ Journal of Political Economy 97(4): 828–862.
Easley, D., and M. O’Hara. 2004. ‘‘Information and the Cost of Capital.’’ Journal of Finance 59 (4):1553–1583.
Fama, E. F., and K. R. French. 1997. ‘‘Industry Costs of Equity.’’ Journal of Financial Economics43: 153–193.
Francis, J., K. Schipper, and L. Vincent. 2002. ‘‘Earnings Announcements and Competing Informa-tion.’’ Journal of Accounting and Economics 33: 313–342.
Frankel, R., S. P. Kothari, and J. Weber. 2006. ‘‘Determinants of the Informativeness of AnalystResearch.’’ Journal of Accounting and Economics 41: 29–54.
Gibbins, M., A. Richardson, and J. Waterhouse. 1990. ‘‘The Management of Corporate Financial Dis-closure: Opportunism, Ritualism, Policies and Processes.’’ Journal of Accounting Research 28,no. 1 (Spring): 121–143.
Graham, J, C. Harvey, and S. Rajgopal. 2005. ‘‘The Economic Implications of Corporate FinancialReporting.’’ Journal of Accounting and Economics 40: 3–73.
Greene, W. 2003. Econometric Analysis. 5th ed. Upper Saddle, NJ: Prentice Hall.Hayes, R. 1998. ‘‘The Impact of Trading Commission Incentives on Analysts’ Stock Coverage Deci-
sions and Earnings Forecasts.’’ Journal of Accounting Research 36 (2): 299–320.Healy, P. M., A. P. Hutton, and K. G. Palepu. 1999. ‘‘Stock Performance and Intermediation Changes
Surrounding Sustained Increases in Disclosure.’’ Contemporary Accounting Research 16, no. 3(Fall): 485–520.
Houston, J. F, B. Lev, and J. W. Tucker. 2008. ‘‘To Guide or Not to Guide? Causes and Consequencesof Stopping Quarterly Earnings Guidance.’’ Contemporary Accounting Research. Forthcoming.
Hsieh, P., T, Koller, and D. R. Rajan. 2006. ‘‘The Misguided Practice of Earnings Guidance.’’ TheMcKinsey Quarterly (March): Web exclusive.
Hutton, A., and P. Stocken. 2006. ‘‘Effect of Reputation on the Credibility of Management Forecasts.’’Working paper, Dartmouth College.
Ivkovic, Z., and N. Jegadeesh. 2004. ‘‘The Timing and Value of Forecast and Recommendation Revi-sions.’’ Journal of Financial Economics 73: 433–463.
Kasznik, R., and B. Lev. 1995. ‘‘To Warn or Not to Warn: Management Disclosures in the Face ofan Earnings Surprise.’’ The Accounting Review 70, no. 1 (January): 113–134.
King, R., G. Pownall, and G. Waymire. 1990. ‘‘Expectations Adjustment via Timely ManagementForecasts: Review, Synthesis, and Suggestions for Future Research.’’ Journal of AccountingLiterature 9: 113–144.
Kreps, D. M., and R. Wilson. 1982. ‘‘Reputation and Imperfect Information.’’ Journal of EconomicTheory 27: 253–379.
Lang, M., and R. Lundholm. 1996. ‘‘Corporate Disclosure Policy and Analyst Behavior.’’ TheAccounting Review 71, no. 4 (October): 467–492.
McNichols, M., and P. O’Brien. 1997. ‘‘Self-selection and Analyst Coverage.’’ Journal of AccountingResearch 35: 167–199.
Mercer, M. 2005. ‘‘The Fleeting Effects of Disclosure Forthcomingness on Management’s ReportingCredibility.’’ The Accounting Review 80, no. 2: 723–744.
Milgrom, P., and J. Roberts. 1982. ‘‘Predation, Reputation, and Entry Deterrence.’’ Journal of EconomicTheory 27: 280–312.
Muth, J. F. 1961. ‘‘Rational Expectations and the Theory of Price Movements.’’ Econometrica 29:315–335.
O’Brien, P. C., and R. Bhushan. 1990. ‘‘Analyst Following and Institutional Ownership.’’ Journal ofAccounting Research 28: 55–76.
455IS SILENCE GOLDEN?
Roulstone, D. T. 2003. ‘‘Analyst Following and Market Liquidity.’’ Contemporary AccountingResearch 20, no. 3 (Fall): 551–578.
Skinner, D. J. 1994. ‘‘Why Firms Voluntarily Disclose Bad News?’’ Journal of Accounting Research32, no. 1 (Spring): 38–60.
Skinner, D. J. 1997. ‘‘Earnings Disclosures and Stockholder Lawsuits.’’ Journal of Accounting and Eco-nomics 23: 249–282.
Soffer, L. C., S. R. Thiagarajan, and B. R. Walther. 2000. ‘‘Earnings Preannouncement Strategies.’’Review of Accounting Studies 5: 5–26.
Stocken, P. C. 2000. ‘‘Credibility of Voluntary Disclosure.’’ Rand Journal of Economics 31, no. 2(Summer): 359–374.
Tucker, J. W. 2007. ‘‘Is Openness Penalized? Stock Returns around Earnings Warnings.’’ TheAccounting Review 82 (4): 1055–1087.
U. S. Chamber of Commerce. 2007. ‘‘Commission on the Regulation of U.S. Capital Markets in the21st Century: Report and Recommendations.’’ Washington, DC: Government Printing Office.
Veronesi, P. 1999. ‘‘Stock Market Overreaction to Bad News in Good Times: A Rational Expecta-tions Equilibrium Model.’’ Review of Financial Studies 12 (5): 975–1007.
Williams, P. A. 1996. ‘‘The Relation between a Prior Earnings Forecast by Management and AnalystResponse to a Current Management Forecast.’’ The Accounting Review 71, no. 1 (January):103–113.
Wilson, R. 1985. Reputations in Games and Markets: Game Theoretic Models of Bargaining, editedby A. E. Roth. Cambridge: Cambridge University Press.
456 JOURNAL OF ACCOUNTING, AUDITING & FINANCE
Copyright of Journal of Accounting, Auditing & Finance is the property of Greenwood Publishing and its
content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's
express written permission. However, users may print, download, or email articles for individual use.