The authors thank Paul Fischer, Mirko Heinle, and audiences at Clemson, Cornell, Duke, Iowa, MIT, Northwestern, Ohio State University, Rice, and Singapore Management University for helpful comments. The authors thank Sehwa Kim for research assistance on this project. The authors gratefully acknowledge the financial support of the University of Chicago Booth School of Business (Jonathan Rogers and Sarah Zechman) and the Neubauer family (Sarah Zechman).

Do managers tacitly collude to withhold industry-wide bad news?

Jonathan L. Rogers University of Colorado

Catherine Schrand

University of Pennsylvania

Sarah L. C. Zechman University of Chicago

That managers would choose to withhold firm-specific bad news is not only intuitive, but supported by theory, observed disclosure patterns, and survey responses. When signals are uncertain and may be revised in future periods, firms may choose to withhold the news to avoid presumably costly stock return volatility. This strategy, however, may not work when news contains a common industry component. When adverse news is industry wide, traders can infer signal arrival from the disclosure of any firm in the industry which increases capital market pressure to disclose. A cooperative equilibrium of non-disclosure requires tacit collusion. We document industry characteristics associated with industry-level financial reporting opacity. We also document intra-industry clustering of increases in the annual 10-K opacity, controlling for changes in fundamentals that we suggest are driven by disclosure choice consistent with tacit collusion. These opacity episodes are more likely in industries that tend to have large, correlated, negative shocks, industries with more significant equity incentives, and greater litigation risk. The opacity episodes are less likely in industries in which observable/public macro-economic data relevant to firm valuation is available. The results have implications for understanding when economic forces are sufficient to generate voluntary disclosure of industry-wide adverse news.

First draft: February 2013

This draft: April 2014 Preliminary and incomplete

1

1. Introduction

This paper investigates determinants of financial reporting opacity related to industry-wide

information. Fundamental characteristics of the firm’s environment such as complexity, uncertainty, or

investment opportunities, affect its opacity. However, voluntary disclosure choice also plays a

significant role in determining opacity. Firms can choose to alleviate inherent opacity by aggregating

and more transparently disseminating private information. Alternatively, firms cans choose to

exacerbate inherent opacity by withholding or obfuscating information. Research on the determinants

of the strategic element of opacity frequently motivate the analysis with adverse selection models of

voluntary disclosure choice that assume firm-specific valuation signals (e.g., Dye, 1985; Jung and Kwon,

1988; Shin, 2003 and 2006). Broadly speaking, these models predict that firms with relatively favorable

news will choose to disclose it, essentially becoming more transparent. The assumption that news

arrival is uncertain, however, allows firms with relatively adverse news to maintain opacity by choosing

not to disclose.

Our analysis of opacity is distinct on two dimensions. First, we assume valuation signals

contain an industry-wide component. The assumption of an industry-wide component in valuation

signals is realistic because of correlated fundamentals due to similar production functions, demand and

supply conditions, investment opportunities, and exposure to external factors such as regulation.

Analyst specialization within industries is consistent with there being a common component to the

valuations (O’Brien, 1990; Dunn and Nathan, 2005). Second, we make the analysis a more realistic

representation of a manager’s disclosure decision by considering disclosure choice in a repeated game

where the signal is uncertain and future revisions are possible. In a single stage game with second

mover costs, coordinated opacity is not possible. Firms will move to the non-cooperative (“prisoners’

dilemma”) equilibrium in which all firms disclose. Assuming a repeated game allows us to consider

whether coordination in disclosure decisions could support a cooperative equilibrium of non-disclosure

when news is industry wide and to predict associated industry characteristics. We refer to cooperative

2

disclosure decisions for industry-wide valuation signals as “tacit collusion”, consistent with the literature

on price and quantity decisions in product markets.1

An analysis of the determinants of opacity, assuming valuation signals contain an industry-wide

component, makes a distinct contribution to existing studies on the determinants of disclosure because

the firm-specific evidence cannot simply be aggregated to make inferences about industry-level

evidence.2 The single stage adverse selection models have uncorrelated signals. They do not

accommodate industry-wide signals, and their predictions do not translate to correlated signals.

Three features of industry-wide signals affect the translation. First, when signals contain an

industry-wide component, disclosure by even one firm in the industry eliminates uncertainty about

signal arrival for other firms in the industry, which is necessary to prevent full disclosure in adverse

selection models. Thus, capital market pressure to disclose is stronger when valuation signals are

industry wide. Second, it is possible that no firms in the industry receive a firm-specific valuation signal

that is higher than the expected valuation without disclosure. In this case, no firm will have an

individual incentive to disclose and to start the unraveling. Mean shifts in the posterior distribution are

not possible in traditional adverse selection models that assume uncorrelated signals. Third, the costs

of non-disclosure are greater when signals are industry wide. In traditional adverse selection models,

non-disclosing firms suffer stock price declines when the valuation adjusts to the conditional expected

value. With industry-wide signals, however, a non-disclosing firm will face an additional cost if a peer

firm chooses disclosure. Litigation costs are one example. Plaintiffs in a 10b-5 case can use the

disclosure by the first mover in an industry as evidence to support an allegation that the second mover

omitted a material fact. Lower financial reporting credibility is another example of a cost imposed by

the first firm to disclose on a second mover. The industry-wide nature of signals generates externality 1 As described in Ivaldi et al. (2003), footnote 2: ‘ “Tacit collusion” need not involve any “collusion” in the legal sense, and in particular need involve no communication between the parties. It is referred to as tacit collusion only because the outcome … may well resemble that of explicit collusion or even of an official cartel. A better term from a legal perspective might be “tacit coordination”.’ 2 Empirical and survey evidence about firm-specific news suggests that firms withhold adverse news (Kothari, Shu, and Wysocki, 2009; Sletten, 2012; Tse and Tucker, 2010; Graham, Harvey and Rajgopal, 2005).

3

costs such that an individual firm’s costs of disclosure depend on the strategic disclosure decisions of

other firms in the industry.3

Our analysis of reporting opacity in this more realistic setting is motivated by the debate about

an appropriate framework for mandated disclosure and predicts distinct determinants relative to those

predicted by traditional single stage adverse selection models with uncorrelated signals. Both the

Financial Accounting Standards Board (FASB) and the Securities Exchange Commission (SEC) are

reviewing the current structure of disclosure mandates. The FASB cites as its goal to establish “…an

overarching framework intended to make disclosures more effective and coordinated and less

redundant.” (FASB exposure draft, 2014).4

The policymakers face two significant issues in developing a conceptual approach to disclosure

regulation. One issue is the extent to which disclosure mandates are necessary at all. Proponents of

mandates believe requirements are necessary because market forces, including capital market pressure

due to adverse selection and litigation risk, are not sufficiently strong to motivate firms to voluntarily

collect or report relevant information, in particular, adverse information. Opponents, however, have

long argued that these market forces encourage voluntary disclosure, even of adverse news, such that

disclosure mandates impose unnecessary costs on firms (Easterbrook and Fischel, 1984). A second

issue standard setters face is how to narrow the requirements in order to balance the benefits and costs,

which differ across reporting entities. One approach is tiered requirements based on firm attributes

such as size or methods of raising capital. Another approach is to set forth disclosure mandates for

specific types of transactions or events based on an assessment of the related benefits and costs.

Our empirical analysis first documents industry-level financial reporting opacity and intra-

industry clustering in opacity for the Fama-French 49 (FF49) industries. We use FOG scores to

3 Dye (1985) discusses externality costs. 4 The SEC states that its current review of disclosure 10-K requirements can lead to “simplification” and “modernization,” which are necessary because the current requirements are costly. The SEC was required to conduct a review of disclosure requirements for emerging growth companies as part of the JOBS Act. However, the SEC states that they decided a comprehensive review of the current disclosure structure could lead to suggested changes that would benefit all issuers.

4

measure opacity. FOG scores are a measure of the readability of a document, and readability of

financial documents has been associated with opacity (Bloomfield, 2002; Li, 2008). This analysis

provides descriptive evidence about firm fundamentals that show intra-industry correlation with

opacity. One notable finding is that firm size, which is a proposed category for tiered disclosure

requirements, is not systematically related to opacity across industries. Size is significantly positive in

only 35% of the industry models, and it is significantly negative in 23%. Two firm characteristics

exhibit a consistent association with opacity levels across industries – special items and the number of

non-missing items on Compustat as a measure of financial complexity both are consistently associated

with greater opacity. The consistent relations across industries between opacity and these two variables,

both of which represent the nature of specific transactions or complexity of specific firm activities,

together with the inconsistent relations between opacity and firm-level attributes such as size, informs

the standard-setting issue of how to narrow disclosure requirements.

The intra-industry analysis of opacity levels provides evidence on the inherent element of

opacity, but cannot be used to test for determinants related to disclosure choice. Correlated opacity

could reflect correlated fundamentals, and correlated fundamentals are an underlying assumption of our

predictions about tacit collusion. To distinguish the element of opacity that reflects disclosure choice

by firms in an industry, we develop a measure of “opacity episodes” that represent industry-years with a

significant increase in the intra-industry average level of 10-K FOG scores that is not explained by

changes in firms’ fundamentals. By examining increases in FOG scores after controlling for

fundamentals we attempt to mitigate concerns that correlated fundamentals, rather than coordination in

disclosure choice, determine correlated opacity.

The number of opacity episodes that we identify is small. Across the FF49 industry groups

over the 14 year period from 1995 to 2008, we find 19 episodes in which the average increase in intra-

industry disclosure opacity suggests coordinated opacity (17 when using only single segment firms). We

also identify 12 of the 19 episodes in which at least 60% of influential firms in the industry exhibit

5

positive unexplained FOG levels. The small number of episodes is consistent with collusive

equilibriums being inherently unstable due to the reliance on conjectures about other firms’ strategic

disclosure decisions. We provide weak evidence that future-period cumulative adjusted returns for the

industries with opacity episodes are significantly more negative than for matched samples starting in the

second year after the episodes. Negative subsequent returns are consistent with our interpretation of

the opacity episodes as a period in which firms did not disclose adverse valuation signals.

Our primary analysis investigates industry characteristics associated with the identified opacity

episodes. We examine characteristics motivated by our discussion of the market forces that shape

disclosure such as capital market pressure due to adverse selection and litigation risk when valuation

signals contain an industry-wide component and are uncertain. Five results emerge. First, opacity

episodes are (weakly) more likely in industries in which firms are subject to significant industry-wide

negative shocks, measured by negative tail risk. The effect is magnified in more homogenous industries

as measured by the average relative weight of industry-wide news to idiosyncratic news in the industry.

This result suggests that industries with a greater propensity for negative shocks that are felt similarly by

peer firms are more likely to conjecture that non-disclosure is a sustainable equilibrium, making tacit

collusion and resulting opacity episodes more likely.

Second, opacity episodes are more likely for industries characterized by firms with high equity

incentives, which motivate firms with adverse signals to remain opaque in order to maintain an elevated

stock price. We measure equity incentives by the proportion of firms in the industry that trade on a

major exchange (i.e., in the CRSP database) relative to the proportion filing annual reports (i.e., in the

Compustat database). An alternative interpretation of the positive association between opacity episodes

and our proxy is that industries with a lower proportion of publicly traded firms (i.e., firms file annual

reports despite not having publicly traded equity) are subject to significant disclosure regulation,

perhaps from regulators or other users that demand financial statements. Firms in such industries will

6

conjecture that other firms will disclose making tacit collusion in opacity decisions less sustainable.

Supplemental analysis favors the first interpretation.

Third, opacity episodes are less likely in industries in which valuation signals are public, or likely

to become public quickly, though this finding is significant for only one definition of opacity episodes.

We identify such industries based on the explanatory power of a returns model that includes market

returns and seven observable public macro-economic signals (i.e., short-term interest rates, default

spreads, term spreads, foreign exchange rates, producer price index, and the Fama-French SMB and

HML factors). When news is public, or the duration until it becomes public is short, the benefits of

opacity decrease (and should be common knowledge), leading to less coordinated opacity.

Fourth, opacity episodes are positively associated with our proxy for industry-level litigation

risk. We interpret the positive association as an indication that our litigation risk proxy, which is based

on a model of suits brought against firms, captures uncertainty in the industry-level propensity for large

adverse shocks. Firms in more uncertain industries are more likely to believe that future valuation

signals may increase (i.e., the news will not be as bad as anticipated), and that this potential is common

knowledge, increasing the sustainability of correlated opacity as an equilibrium. Our model of the

determinants of the opacity episodes also includes a separate proxy for industry-level uncertainty

implied by option prices. However, we find no association between more volatile industries and the

opacity episodes, and our litigation risk proxy remains significant in its presence.

Finally, opacity episodes are weakly positively related to industry concentration. We include a

revenue-based herfindahl index in the model because industry concentration could be related to several

industry characteristics affecting opacity choice. Greater concentration could increase the probability

that firms conjecture other firms have received a similar negative signal. Greater concentration could

also be correlated with fewer influential firms in the industry, which provides greater opportunities to

form beliefs that the other influential firms will cooperate, perhaps because of explicit collusion, greater

interpersonal connections, or because a focal point is more likely available to facilitate cooperation.

7

Finally, greater concentration could imply more significant existing barriers to entry, which decreases

each firm’s incentives to disclose adverse signals relative to industries with low entry barriers.

Several analytical models have analyses related to the issue of disclosure of industry-wide news.

Dye and Sridhar (1995) analyze disclosure choice when the timing of the arrival of private signals is

correlated across firms in an industry, but the signal values are uncorrelated. Jorgensen and

Kirschenheiter (2012) analyze disclosure of industry-wide news with an exogenous leader and follower

firm. The leader anticipates the disclosure decision of the follower, but there are no opportunities for

signal revisions. The closest analysis is by Acharya, DeMarzo, and Kremer (2011) who examine

disclosure when signal values have a market-wide component and predict that firms will delay

disclosure of adverse news.5 Our analysis has two distinguishing features. First, the ultimate release of

the news in their model is exogenous, and second, they do not allow for signal revisions. Acharya et al.

(2011) discuss the possibility that the news is endogenously released via another firm’s disclosure,

however, they only conjecture that the results would hold, stating: “The construction of the equilibrium

presents a significant computational challenge.”

Section 2 provides a conceptual discussion of the determinants of opacity when valuation

signals contain an industry-wide component and managers make disclosure decisions in a repeated

game anticipating the potential for signal revision. Section 3 describes our measures of intra-industry

opacity and our method for identifying opacity episodes. Section 4 presents the results of the analysis

of the determinants of the opacity episodes and Section 5 concludes.

2. Discussion of determinants of industry clustering

Two factors determine reporting opacity: 1) fundamentals or inherent opacity, and 2)

managerial choice. The notion that managerial choice affects opacity is consistent with the “incomplete

5 Both Dye and Sridhar (1995) and Acharya et al. (2011) predict a clustering of bad news announcements because disclosure by one firm will trigger disclosure by the multiple firms that withheld adverse news. Tse and Tucker (2010) provide empirical evidence of clustering in bad news announcements.

8

revelation hypothesis” of Bloomfield (2002). In a noisy rational expectations equilibrium with costly

and endogenous information acquisition, opacity hinders efficient price discovery. Managers will make

financial reporting decisions that increase the costs of extracting information that the manager does not

want aggregated into price.

Firm fundamentals that affect financial reporting opacity are interesting in their own right, and

we will analyze their association with reporting opacity at the industry level. The discussion in this

section relates to strategic disclosure choice and opacity. In Section 2.1 and 2.2, we discuss how the

assumptions of an industry-wide valuation component and potential for revised signals alter predictions

about how capital market pressure affects disclosure choice. In Section 2.3, we discuss how these two

assumptions alter predictions about how litigation risk affects disclosure choice. The purpose of the

discussion is to motivate the determinants of opacity that result from strategic interactions.

2.1 Industry-wide valuation signals

We first discuss capital market pressure when valuation signals contain a common industry-

wide component. Traditional adverse selection voluntary disclosure models with trader uncertainty

about signal arrival form the basis for the discussion (e.g., Dye, 1985; Jung and Kwon, 1988; Shin, 2003

and 2006).6 In these models, the manager receives a private firm-specific valuation signal and chooses

whether to disclose it. Traders respond to disclosure and silence (i.e., non-disclosure), taking into

account the strategic behavior of managers. The stock price of a firm that discloses adjusts to the

disclosed amount. The stock price of a non-disclosing firm will adjust to the expected value

conditional on non-disclosure. Adverse selection models predict a partial disclosure equilibrium in

which firms will not disclose if their valuation signal is below an endogenous threshold level that

equates the valuation given disclosure to the expected valuation given non-disclosure. The intuition is

that firms with adverse news are able to hide in the pool of firms that have no news because of the

assumption that traders are uncertain about signal arrival. 6 We refer to models with trader uncertainty about signal arrival simply as “adverse selection” models.

9

These adverse selection models assume that firms’ private signals are uncorrelated. By

definition, industry-wide signals are positively correlated across firms in an industry; the valuation signal

contains a common industry component. That valuations would have a common intra-industry

component is sensible. Firms within an industry are likely to have relatively homogeneous production

functions, to face similar supply and demand conditions, to have similar investment opportunities, and

to face similar exposure to external factors such as regulation.

Considering industry-wide rather than firm-specific valuation signals does not alter predictions

about a firm’s disclosure choice of a favorable signal. Assuming correlated signals does, however, alter

predictions about disclosure of adverse signals. With industry-wide signals, traders can infer signal

arrival for one firm in an industry based on the disclosures made by another firm in the same industry.

If even one firm in the industry receives a firm-specific valuation signal that is higher than its

unconditional expected valuation despite the adverse industry-wide component, that firm will disclose.

The disclosure will cause traders to infer that all firms in the industry received a signal, which in turn,

affects the conditional expected value for the non-disclosers. The uncertainty about signal arrival,

which generates the partial disclosure equilibrium in the threshold models, disappears and adverse

selection should lead each manager to disclose. The idea that traders would make such intra-industry

inferences is consistent with the evidence that investment analysts tend to specialize in one industry

(O’Brien, 1990) and that analysts who focus on one industry produce better forecasts than analysts

focusing on multiple industries (Dunn and Nathan, 2005).

The additional capital market pressure when signals contain an industry-wide component results

in more disclosure. If firms’ private signals contain only the industry-wide component and no

idiosyncratic component, adverse selection leads to full unraveling and full disclosure. If firms’ private

signals contain an idiosyncratic component in addition to the industry-wide component, then the

disclosure threshold will be lower than the threshold in traditional adverse selection models that assume

uncorrelated signals.

10

The discussion thus far suggests industry-wide signals will be associated with stronger capital

market pressure relative to uncorrelated signals. However, in the traditional models, at least one firm

will receive a private signal that is greater than the unconditional expected value and disclose it. When

the signals contain a common industry component, however, the private valuation signals received by

all firms in an industry could be less than their existing stock price valuations. No firm will individually

want to disclose, and if the distribution shift is common knowledge, unraveling will not start.7

At this point, one could argue that firms’ conjectures about the disclosure choices of other

firms are irrelevant. Firms that receive adverse signals choose non-disclosure. If a peer discloses, the

firm responds with disclosure. The firm may incur an immediate stock price reaction prior to the

response, which would be incorporated into the determination of the endogenous disclosure threshold.

We propose, however, that being the second mover imposes additional costs on a firm, besides the

immediate stock price reduction, when firms’ valuation signals contain a common industry component.

One example of an externality imposed on the second mover is lost financial reporting credibility. Lost

reputation is greater when a firm can be compared to a peer firm. An individual manager that

withholds materially adverse news can be subject to SEC enforcement. Individuals identified in

enforcement actions lose their jobs, face restrictions on future employment, and incur pecuniary costs

including fines and valuation losses on shareholdings (Karpoff et al., 2008a). Individuals can also face

criminal charges and jail time (Karpoff et al., 2008a; Schrand and Zechman, 2012). Karpoff et al.

(2008b) document reputational penalties at the firm level and characterize the direct legal costs as

minimal, but the loss in firm value associated with the reputational effects of the misconduct as “huge.”

Litigation costs are another example of an externality imposed on the second mover. One

element of a 10b-5 case is to show that the defendant omitted a material fact. The disclosure of

7 This observation is not unlike a point made by Dye and Sridhar (1995) in the “industry-wide common knowledge” case. They analyze disclosure choice when the timing of the arrival of private signals is correlated across firms, however, the signal values are uncorrelated. Full non-disclosure is a possible equilibrium, but only if firms “know” that other firms will not disclose (p. 166).

11

industry-wide news by a peer firm in the industry can provide useful evidence to support this claim. If

one firm in the industry discloses, plaintiffs are more likely able to establish that the non-disclosing firm

knew the information (or was reckless in not knowing it), which is otherwise difficult to prove. It may

also be easier for a plaintiff to establish materiality given that the disclosing firm assessed the

information as material. While litigation risk affects even a disclosing firm, as will be discussed later, we

propose that an industry-wide component in valuation signals creates incremental litigation costs for a

non-disclosing firm because the revelation of the adverse news by a peer firm could justify a stronger

case for the plaintiff.

In summary, when valuations contain an industry-wide component greater capital market

pressure arises because trader uncertainty about signal arrival disappears when another firm in the

industry discloses. However, this capital market pressure may not induce disclosure if the industry-wide

component is sufficiently adverse such that no firms have individual incentives to disclose their

valuations. The unraveling that occurs in adverse selection models would not start. The costs

associated with non-disclosure, including litigation and reputation costs, could be greater when news is

industry wide. The externality costs of being the second mover in the disclosure game are an additional

force that can lead to greater disclosure.

2.2 A repeated game

We next consider how the anticipation of opportunities to disclose revised signals in subsequent

periods affects managers’ incentives to disclose valuation signals. Dye and Sridhar (1995) and Acharya,

DeMarzo, and Kremer (2011), discussed previously, have multiple periods, but the future periods only

represent additional opportunities to disclose the initial signal; revised signals do not arrive. We

consider the implications of disclosure choice in a repeated game because we are interested in

identifying industry conditions and economic forces that shape disclosure in the realistic setting in

which firms make disclosure decisions. Incorporating the potential for signal uncertainty in evaluating

disclosure decisions accords with survey evidence in Graham et al. (2005) that managers hide adverse

12

news “…in hopes that the firm’s status will improve before the next required information release,

perhaps saving the company the need to ever release the bad information...” (p. 65).

Because valuation signals are uncertain, the manager can assess positive probability to an

adverse signal being lessened or never materializing. Anticipation of subsequent revisions introduces

an important tension in the manager’s disclosure decision. If the manager and industry peers do not

disclose the manager avoids costly interim-period stock price volatility. However, if the manager does

not disclose but another firm does, the manager faces externality costs associated with being the second

mover. In a single stage game with second mover costs, coordinated opacity is not possible. Firms will

move to the prisoners’ dilemma equilibrium in which all firms disclose.

In a repeated game, it is possible for firms to coordinate and collectively choose non-disclosure

or opacity despite the potential costs of being the second mover. Coordinated opacity requires each

individual manager assign positive probability to reversal, conjecture that other managers have similar

beliefs, believe they share these beliefs, and so on. As such, “softer” issues affecting beliefs about

disclosure by other firms will determine whether the cooperative equilibrium is sustainable. Focal

points make a cooperative equilibrium more sustainable (Pindyck and Rubinfeld, 1989).8 Frequent

interactions between the players also facilitate cooperation (Ivaldi et al., 2003). Perhaps the softest of

the variables is trust. Cooperation is more likely when firms can build trust over time that other firms

will make the cooperative choice. The observed coordination in firms’ disclosure decisions is referred

to as “collusion” in the product market literature, but the use of this term does not imply that firms

directly communicate their strategies (i.e., “explicitly” collude). Tacit collusion only implies that the

manager’s disclosure choice depends on her beliefs about the behavior and beliefs of other firms.

Our predictions about when correlated opacity is sustainable are similar to predictions of tacit

collusion in product markets that use a repeated game setting. Unlike product market games, however,

8 In fact, Pindyk and Rubinfeld (1989) give the prescient example that LIBOR could serve as a focal point, allowing banks to cooperate.

13

firms’ disclosure decisions do not have strategic interactions that affect the level of payoffs to

cooperation or the level of externality costs.9 The focus of our analysis is on the disclosure of the

adverse signal, not on tacit collusion in operating, investing, or financing decisions as a reaction to the

signal.10 This idea is in the background in our analysis, though, because of the assumption that firms

anticipate receiving new signals in subsequent periods and that the beliefs are common knowledge.

The reliance on conjectures makes coordinated equilibrium strategies inherently unstable, as evidenced

by the numerous empirical studies on tacit collusion or cartel formation in product markets and the

explanations for why cartels dissolve.11

2.3 Litigation risk

Adverse selection models feature capital market pressure as a market force that affects

disclosure choice; litigation risk is another. A common assertion is that firms disclose adverse news

early to mitigate litigation risk (e.g., Skinner, 1994). The empirical evidence is mixed. Skinner (1994)

finds evidence consistent with this argument, but Francis, Philbrick, and Schipper (1994) do not find

evidence that preemptive disclosure mitigates litigation risk. Skinner (1997) then examines a broader

sample and finds support consistent with Francis et al. (1994) but suggests that, while issuing forecasts

may not deter litigation, it may reduce the costs.

This line of reasoning suggests that litigation risk is a market force that reduces the need for

mandated disclosure of firm-specific news. The implications of litigation risk are unclear, however, in a

repeated game setting and when signals contain an industry-wide component. Litigation risk could be

an even stronger market force than predicted by firm-specific models when adverse signals contain an

9 Another difference between our assumed payoff structure and those in models of price or quantity decisions is that tacit collusion in product market decisions typically describes the cooperative equilibrium that arises when firms can credibly retaliate in a future period against firms that deviate (Bertomeu and Liang, 2008). The threat of retaliation does not exist in the disclosure setting. 10 Rajan (1994) is an example of a model of firms’ choices about disclosures of industry-wide news in the context of a specific strategic decision. The disclosure choice is a bank’s decision about recording credit reserves, which serves as a signal to other banks that make lending decisions about the overall level of credit quality in the industry. The analysis focuses on the implications of the recorded reserves for strategic lending and other decisions, which affect payoffs and business strategy, and the overall availability of credit. 11 See Levenstein and Suslow (2006) for a review.

14

industry wide component. If firms in an industry conjecture that at least one firm in the industry will

disclose due to litigation risk, the potential for externality costs associated with being a second mover

will favor greater disclosure. Litigation risk could mitigate or exacerbate the incentives for firms to

preemptively disclose adverse news in a repeated game relative to the predictions from a single stage

game. On one hand, as the probability that the signal will materialize or be discovered declines,

expected litigation costs also decline, favoring non-disclosure. But Skinner (1994) suggests that

litigation costs are increasing in the duration of withholding because a longer withholding period means

a longer class period.

3. Measuring intra-industry opacity and identifying opacity episodes

The variable of interest is reporting opacity. We use readability as an indication of opacity, and

our proxy for readability is the Gunning FOG index for firms’ 10-Ks. The FOG index is computed as

a function of the average number of words per sentence and the percent of words with three or more

syllables. The computation is intended to result in a score that represents the number of years of

education required to read the document. We obtain 10-K FOG scores as used in Li (2008) from Feng

Li’s website (http://webuser.bus.umich.edu/feng/). His sample is derived from the intersection of

firms available in the CRSP and Compustat databases and with electronic filings on EDGAR. Li’s

FOG score is calculated using the text (excluding heading items, paragraphs less than one line, and

tables) for all 10-K filings with more than 3,000 words or 100 lines of remaining text (Li, 2008). The

data provided on the website has been updated to include all observations meeting these requirements

that were filed with the SEC through December 2009 (or with fiscal year ends through mid-2009).

Empirical evidence provides some support for using the FOG index as a measure of opacity.

Callen, Khan, and Lu (2013) show that a higher FOG score (lower readability) is associated with greater

price reaction delay to a firm’s 10-K, where price reaction delay is measured by how much of the firm’s

return can be explained by lagged market returns. Lehavy, Li, and Merkley (2011) show that higher

15

FOG scores (lower readability) are associated with higher analyst forecast dispersion, lower forecast

accuracy, and analyst uncertainty. They interpret their evidence as consistent with FOG scores being

associated with information that is more difficult to process.

We proxy for opacity using poor readability rather than using non-disclosure for two reasons.

First, extant studies show that termination of previously disclosed information is costly (e.g., Chen,

Matsumoto, and Rajgopal, 2011), and second, Reg. S-X requires firms to discuss various topics, but the

content is discretionary. In addition, a readability index like the FOG score is objective and can be

measured on a large sample, as discussed in Lehavy et al. (2011).

The first element of our analysis examines industry fundamentals associated with opacity. For

this purpose, we estimate intra-industry models of FOG score levels. Section 3.1 describes the model

and results. In Section 3.2, we describe the model we use to identify unexplained changes in industry-

level reporting opacity. The procedure creates a set of “opacity episodes.” In Section 4, we estimate

the industry characteristics that are determinants of the opacity episodes using a logit model. The

proposed determinants are motivated by the discussion in the previous section.

3.1 Modeling levels of intra-industry opacity

We estimate the determinants of intra-industry opacity using the following model:

𝐹𝑂𝐺𝑗𝑡 = 𝛼0 + ∑ 𝛿𝑐9𝑐=1 𝐶𝑂𝑁𝑇𝑅𝑂𝐿𝑗𝑡𝑐 + 𝜖𝑗𝑡 (1)

where FOG is a firm’s 10-K FOG score. The unit of observation is firm (j) in year (t). We estimate the

model annually for each of the FF49 industries from 1994 to 2008 using two samples. The FULL

sample includes all firm-year observations with available data. The 1SEG sample excludes firms with

multiple business segments, where segments are defined at the FF49 level.12 The vector CONTROL

includes nine variables, identified by Li (2008), that represent firm fundamentals: the log of market

value of equity, the market to book ratio, special items scaled by total assets, return volatility, the

12 We use the Fama French 49 industry level to identify single segment firms in order to mitigate the effects of SFAS 131, which substantially reduced the number of single segment firms beginning in 1998.

16

number of non-missing data items in Compustat as a measure of complexity, firm age, an indicator for

Delaware incorporation, the log of the number of geographic segments plus one, and the log of the

number of business segments plus one. (See variable definitions in Appendix A.)

Table 1 reports coefficient estimates by industry for the FULL sample (Panel A) and for the

1SEG sample (Panel B). At the top of each panel, we report the number of coefficient estimates that

are significantly positive, significantly negative, and insignificant for each control variable. We present

results for 48 industry-level regressions, excluding “Tobacco Products” because of a small sample

size.13 The first notable finding is that firm size, which is a proposed category for tiered disclosure

requirements by the FASB, is not systematically related to opacity across industries. Li (2008) indicates

a positive association on average in his pooled sample across industries, but size is significantly positive

in 35% of the industry models and significantly negative in 23%.

The only determinants that have relatively consistent explanatory power across industries are

special items scaled by total assets and the number of non-missing data items in Compustat. Firms

with larger special items, which are generally negative, have consistently less readable financial reports

(higher FOG scores) across industries, consistent with the findings in Li (2008). Firms with more non-

missing items, which are presumably more complex, also have consistently higher FOG scores or lower

readability across industries. This result is opposite to the negative association in the pooled sample in

Li (2008), but it is consistent with his prediction that greater financial complexity is associated with

greater opacity. Special items and non-missing items are distinct from the other determinants we

explore in this specification, none of which show consistent patterns across industries, in that they

represent specific types of transactions or activities rather than attributes of the firms. The intra-

industry analysis therefore suggests that tiered disclosure requirements based on firm attributes may not

identify the disclosure conditions that are most at risk of opacity. Instead, an approach focusing on

13 The Tobacco Products industry (Fama-French industry 5) has 54 industry-year observations in the FULL sample panel (an average of less than four per year in the 15 year sample period) and only 14 observations in the 1SEG sample panel.

17

characteristics of events or transaction would be more likely to alleviate sources of opacity that are

consistent across industries.

Table 2 reports measures of intra-industry dispersion in FOG for each of FF49 industries. We

measure dispersion as the intra-industry coefficient of variation (standard deviation/mean) and as the

quartile coefficient of dispersion (Q3 - Q1/median). We measure dispersion separately for the FULL

sample and the single segment firms (1SEG) over the period 1994 through 2008. Lower dispersion

implies greater clustering of readability within the industry. We evaluate whether the industry clustering

is unusually high or low by testing the equality of the variance in each industry against the variance for

the pooled sample of firms in all industries. The coefficient of variation for the pooled sample

reference group in the FULL sample is 0.0742 (Panel A) and 0.0734 in the 1SEG sample (Panel B).

The last column of Table 2 indicates whether the industry exhibits more significant clustering/less

dispersion (“c” or “cc”) or more significant dispersion (“x” or “xx”) than the reference group.14

The industries with the most significant clustering are Consumer Goods (9), Healthcare (11),

Precious Metals (27), Computer Software (36), and Insurance (46). Additional industries that exhibit

weaker evidence of clustering are Apparel (10), Pharmaceutical Products (13), Chemicals (14),

Communication (32), and Business Services (39).

3.2 Identifying opacity episodes

The object of interest in the primary analysis is coordinated opacity that results from voluntary

disclosure choice rather than correlated fundamentals. Li (2008) concludes that the FOG measure, in

levels, may capture intentional opacity. He documents that higher FOG score levels are associated with

lower earnings and lower earnings persistence for profitable firms and with earnings decreases. Based

on this evidence, he concludes that managers may be strategically disclosing to “hide adverse news.”

14 We compute three statistics to test the equality of the variances: Bonferroni, Scheffe, and Sidak. The Bonferroni and Sidak statistics provide identical results. The Scheffe statistic rejects equality of variances for fewer industries. An “x” or “c” means that equality of variances is rejected by the Bonferroni and Sidak statistics. An “xx” or “cc” means that the equality of variances is also rejected by the Scheffe statistic.

18

While his evidence is consistent with FOG scores representing opacity choice, fundamentals such as

complexity or realized poor performance that are omitted from his FOG model could also drive these

results (Bloomfield, 2008).

In an effort to distinguish the correlated opacity that results because of strategic disclosure

choice given an industry-wide valuation component and a repeated game, we identify a set of industry-

year observations that exhibit significant unexplained increases in intra-industry FOG scores. We

modify eqn. (1) to measure industry changes in FOG scores, controlling for changes in fundamentals.

We estimate eqn. (2) for rolling two-year windows from 1994 through 2008

𝐹𝑂𝐺𝑗𝑡 = 𝛼0 + 𝛼1𝑌𝐸𝐴𝑅2 + ∑ 𝛿𝑐9𝑐=1 𝐶𝑂𝑁𝑇𝑅𝑂𝐿𝑗𝑡𝑐 + 𝜀𝑗𝑡 (2)

where YEAR2 is an indicator for the later of the two years. This model allows us to identify industry-

years with unexplained changes in FOG. An industry-year is considered an “opacity episode” in the

latter of the two years if the coefficient estimate on the YEAR2 indicator ( 1α ) is positive and

significant (p-value < 0.10). For use in later falsification tests, we also identify “transparency episodes.”

An industry-year is considered a transparency episode (TRANSP) in year 2 if 1α is negative and

significant (p-value < 0.10). This analysis generates 19 industry-year observations of opacity episodes

for the years 1995 through 2008 and 13 transparency episodes.

We identify a second set of opacity and transparency episodes by estimating eqn. (2) using only

single segment firms based on the FF49 industries (“1SEG” sample). This method identifies 17 opacity

episodes and 12 transparency episodes. We expect the parameter estimates when estimating eqn. (2)

with the 1SEG sample to provide better controls for the fundamental firm characteristics that affect

FOG, thus providing a cleaner measure of the change in abnormal FOG.

Table 3 reports the industry distribution of opacity and transparency episodes estimated using

both the FULL and 1SEG samples. Tables 3 and 4 report episodes by industry and year, respectively.

Technology-related industries (35, 36, and 37; hardware, software, and chips), are over-represented in

19

the opacity episodes, with all three industries showing industry-wide decreases in readability in 2002 and

2003. There is less obvious industry clustering in the transparency episodes. The opacity episodes

generally exhibit time clustering in 2002 and 2003. We do not expect that required disclosures related

to the Sarbanes-Oxley Act of 2002 (SOX) are the source of this clustering given that the specific

provisions were not effective until later and the general provisions should have affected all industries.

However, as a precaution, we estimate the logit model of the determinants of the opacity episodes

including non-opacity observations only for years in which there is at least one opacity episode.

We also identify a third set of opacity episodes that is a subset of the FULL sample episodes.

In addition to the on-average increase in industry-wide unexplained FOG, we require the majority of

influential firms in the industry exhibit significant unexplained increases in FOG. To identify influential

firms in each industry-year, we find the lowest number of firms such that the sales concentration ratio

is greater than or equal to 50%. For example, if the concentration ratio for the largest four (six) firms is

40% (55%), then the number of influential firms in the industry is six. We apply the coefficient

estimates from eqn. (1) for industry i in year y-1 to the control variables for the influential firms in

industry i in year y to estimate their predicted FOG. A positive abnormal FOG, equal to actual FOG

less predicted FOG, indicates relatively high disclosure opacity for that year. This procedure identifies

12 of the original 17 opacity episodes that also have at least 60% of influential firms with abnormally

high FOG levels. We expect these episodes to provide a cleaner measure of strategic opacity.

Our indirect approach to identifying periods of opacity is preferred to other more direct

approaches. We cannot rely on observed ex post adverse outcomes and work backwards to examine

whether the adverse signal was withheld. Such a sample selection method would identify only those

episodes in which the adverse signals eventually materialized. But the discussion in Section 2 suggests

that correlated opacity is more likely when there is greater uncertainty about the eventual realization of

the signal. Thus, such sampling would be biased against identifying periods in which firms chose

20

opacity. We also cannot examine reporting decisions using concurrent performance – if the firm is

recognizing the concurrent performance it is effectively disclosing the signal.

3.3 Returns tests

We use a traditional portfolio method to investigate the relation between opacity episodes and

subsequent returns. We compare the returns of firms in industry i following an opacity episode to

those of a matched sample described below. We expect to observe poor returns, on average, following

opacity episodes if our episodes capture opacity that obfuscates adverse industry-wide valuation signals.

This analysis is inherently weak because opacity episodes are more likely when firms believe the adverse

signal will not materialize (or will be better than originally anticipated). Poor returns on average are

consistent with strategic opacity, although not finding poor subsequent returns does not preclude

strategic opacity; the initial adverse valuation signal may not have materialized. Also, we do not make

predictions about the timing of the subsequent poor returns, although we do not expect them to be

immediate. Strategic opacity is positively related to the duration that the stock price remains artificially

elevated. If firms had rationally anticipated that the valuation signal would be made public quickly, we

should not observe correlated opacity.

We measure subsequent returns as the cumulative abnormal size-decile adjusted returns (CARs)

in months 1, 3, 6, 12, 24 and 36 after month four following a year end. We compute CARs for an

opacity sample and a matched sample. The opacity sample consists of firm year observations in

industry i and year y if the i-y industry-year combination is an opacity episode based on the 1SEG

sample. The matched sample consists of all other firm-year observations, excluding firms in industry i

in years y+1 through y+3. For example, if an industry has an opacity episode in 2002, firms in the

industry are excluded from the matched sample in 2003-2005. This exclusion ensures that returns for

the opacity sample firms do not confound the y+1 and y+2 returns for the matched sample. We

examine return performance for all firms in the industry, including firms that were not in the

21

determination of the opacity episodes due to data availability, and we do not require firms to have data

for all three subsequent years.

Table 5 reports the average firm-level CARs for the opacity sample and matched sample

portfolios for the 15 industry-year opacity episodes (OPACITY1SEG = 1). Panel A presents average

CARs for portfolios that include only single segment firms and exclude new entrants to the industry,

defined as firms with returns available as of month four following the opacity episode year (y) but with

missing returns as of December year y.15 Panel B presents average CARs for single and multi-segment

firms, which significantly increases the sample size in our tests, but at the cost of less specific returns

related to the industry segment subject to the adverse valuation signal.

In the opacity sample portfolios in Panels A and B, there are 1,750 (2,246) observations for

single segment (single and multi-segment) firms with data available to compute CARs in the first month

following the opacity episodes aggregated across the 15 opacity episodes. This number declines to

1,295 (1,708) by month +36. In both panels, the CARs of firms in the opacity sample turn negative

(but not significant) by the end of the second year following the episode. The CARs for the opacity

sample are significantly greater than for the matched sample at each date up to one year following the

episode (Ret12mon). By Ret24mon, however, the opacity sample CARs are significantly lower than

those of the matched sample.

Panels C and D report CARs for cross-sections based on the propensity of the industry to

experience adverse valuation signals. We identify such industries based on expected return skew

implied by option prices. The procedure designates 17 high negative tail risk (positive skew) industries,

or approximately 35% of the FF49 industries. The industries are designated with an asterisk in

Appendix B. Bifurcating the opacity and matched samples based on the likelihood of an industry-wide

adverse signal is meant to reduce noise in the matched sample by eliminating industries that are unlikely

15 We also create a set of firms that includes single segment new entrants to the industry. All results (untabulated) are qualitatively the same, suggesting that new entrants and existing firms were similarly affected by the industry-wide valuation component.

22

to have had adverse signals. We expect the differences in average CARs to be more pronounced when

comparing only high negative tail risk industries.

Panel C reports CARs for single segment firms in high negative tail risk industries (comparable

to Panel A). Except in month +1, the differences between the opacity and matched sample portfolios

are insignificant in the first two years, in contrast to the higher CARs for the opacity sample reported in

Panel A. The CARs turn significantly negative for the opacity sample, but not for the matched sample,

in Ret24mon, and the difference in CARs between the two samples is significantly negative in

Ret36mon. Panel D shows that the low tail risk industries drive the positive differences in CARs

between the opacity and matched sample portfolios in months 1, 3, 6, and 12 in Panel A. In fact, in the

low tail risk industries, the opacity sample firms have significantly higher CARs than the matched

sample in Ret36mon.

In untabulated analysis, we adjust CARs for performance-related delistings. These missing

observations will create a bias in our sample that would work against finding subsequent poor returns

in the opacity sample if the missing observations are firms with the most egregious adverse news. We

replace observations missing due to performance-related delistings with the 5th percentile CAR for the

same FF49 industry in year y. The replacement of missing values with the 5th percentile is made for

both the opacity and matched sample portfolios. The delisting codes (from CRSP) we use to identify

performance-related delistings are 400 to 500 and 520 to 584 (Shumway, 1997). We assume delistings

for other reasons (e.g., changes in exchanges) do not create a bias in our portfolio returns. In Panels A

and B, the CARs are lower, by construction, for both the opacity and matched sample portfolios, but

the differences in CARs are similar in magnitude and significance.

Overall, we find some evidence of poor subsequent returns for firms following opacity

episodes. The analysis shows the episodes can empirically distinguish industries that realize poor

returns two to three years following the episode, which provides some validation that the ex ante

identified episodes are associated with yet unrealized adverse signals. Our analysis of average realized

23

returns understates the amount of adverse news if the signal was either partially mitigated or never

realized. Firms may have chosen opacity anticipating that the adverse signal would not materialize.

4. Determinants of opacity episodes

We estimate the determinants of opacity episodes identified in Section 3 using a logit model.16

The discussion in Section 2 motivates the industry characteristics that we include in the model as

explanatory variables, all of which are measured at the industry or industry-year level.

OPACITYiy=α0 + δ1TAILRISKi + δ2IND_WIDEiy + δ3TAILRISKi*IND_WIDEiy +

δ4VOLATILITYiy + δ5PUBLICi + δ6%CRSPiy + δ7LIT_RISKiy + δ8HERFiy +εiy (3)

We estimate the model separately for opacity episodes identified using the FULL sample

(OPACITY=1 and OPACITYINFL=1) and the 1SEG sample (OPACITY1SEG=1). The dependent

variables equal zero for remaining industry-year observations and are missing if data were not available

to estimate the FOG model. An industry-year is also set to missing for industry i in years y+1 through

y+3 if the industry was defined as an opacity episode in year y. This exclusion reduces noise in the non-

opacity observations, since opacity in year y may continue in subsequent years. The logit model

includes only years with at least one opacity episode. These restrictions yield 17 opacity episodes (out

of the original 19) for the FULL sample, 12 opacity episodes with influential firms, and 14 opacity

episodes (out of the original 17) for the 1SEG sample. Two OPACITY1SEG episodes are lost because of

repeat years and the 1995 episode is lost because the litigation risk proxy is available starting in 1996.

4.1 Explanatory variables and predictions

We measure all explanatory variables at the industry or industry-year level consistent with our

goal of characterizing industries likely to have opacity episodes. Whether opacity is a sustainable

equilibrium depends not only on each firm’s own incentives and costs, but also on the firm’s

16 We also estimate the model using a probit specification with consistent results.

24

conjectures about other firms’ disclosure decisions. If firms conjecture that it is in the best interest of

all firms in the industry to maintain opacity, then opacity is a sustainable equilibrium. If a firm

anticipates that a peer firm will deviate and disclose the adverse signal, no firm will be willing to

withhold the industry-wide news, fearing that other firms will disclose it first and impose externality

costs on the non-discloser. Because the sustainability of an opacity equilibrium depends on conjectures

about the costs and benefits for all firms in the industry, we measure the explanatory variables at the

industry or industry-year level. Appendix A provides details of the variable construction. Appendix B

reports the variables by industry.

TAILRISK: The first explanatory variable is an indicator equal to one for industries in which

firms are more likely to have adverse or complex news. We identify industries with high negative tail

risk based on expected return skew implied by option prices as described previously. We use an

industry indicator rather than an industry-year indicator because we use market data to construct the

variable. While industry tail risk may be time variant, creating an annual measure requires assumptions

about when the market realizes the news that is reported opaquely. We expect firms in high tail risk

industries to have a greater propensity to receive an adverse industry-wide signal, which increases an

individual firm’s incentives to maintain opacity. If this industry characteristic is common knowledge,

capital market pressure may not induce disclosure, making opacity episodes more likely.

IND_WIDE: The second explanatory variable proxies for the degree to which valuation signals

in an industry have a common component. When the industry-wide component is likely to be

significant relative to the idiosyncratic component, individual firms can more easily infer other firms’

incentives. Even if this industry characteristic is common knowledge, however, the implication for the

sustainability of coordinated opacity is ambiguous. On one hand, it is more likely that managers

receiving an adverse signal will conjecture that other firms in the industry have received an adverse

signal, and conjecture that the industry-wide component is common knowledge. This line of reasoning

suggests that homogeneity makes coordinated opacity more likely. However, homogeneity also increases

25

the probability that traders will infer industry-wide signal arrival from the disclosure of any individual

firm. In less homogenous industries, traders infer less from disclosure by one firm in the industry

about signal arrival for another firm, thus uncertainty about news arrival remains.17

Our proxy for the level of industry homogeneity (or degree of industry-wide news) is the

negative value of industry-year heterogeneity. We compute heterogeneity as the annual average of the

standard deviation of residuals from monthly within-industry standard market models. This industry-

year measure (IND_WIDE) represents the average industry-wide component of returns relative to the

idiosyncratic component. The mean (median) reported in Table 6 Panel A is -0.16 (-0.16). Equation

(3) also includes the interaction of IND_WIDE with the industry-level indicator for high tail risk as it is

likely that the intersection of these characteristics has a stronger effect on coordinated opacity.

VOLATILITY: The third explanatory variable, industry return volatility, is included in the

model for two reasons. First, we expect volatility to be correlated with fundamental uncertainty and

thus with valuations that are inherently more difficult to describe. Second, we expect industry-wide

return volatility to affect disclosure incentives. The more uncertain the valuation signals, the more

likely it is that future signals will at least partially reverse. An adverse signal may never be revealed if the

source of the shock is not realized (e.g., a decline in average customer credit quality reverses prior to

significant credit losses.) Uncertainty increases the incentives for each firm to withhold adverse news.

If this industry characteristic is common knowledge, capital market pressure may not induce disclosure,

making opacity episodes more likely in high-volatility industries. We measure industry return volatility

as the average implied volatility of at-the-money options for firms in an industry following Van Buskirk

(2011). We use an industry indicator rather than industry-year indicator because we use market data to

construct the variable. 17 This discussion mirrors a case considered in Dye and Sridhar (1995). In their main analysis, the timing of the arrival of private signals is correlated, but the signal values are uncorrelated, and firms do not learn whether other firms have received a signal. In their comparative static analysis related to the number of firms (n) in the industry (Theorem 2a), two forces are at play. As n increases, traders infer less from the disclosure by any one firm about the probability that a non-discloser has received a private signal, but adverse selection motivates more firms to disclose. The second effect always dominates and the threshold is decreasing in n.

26

PUBLIC: The fourth explanatory variable is a proxy for the expected duration until the

valuation signal is revealed. The longer the expected duration of opacity, the greater is the expected

value of the artificially elevated stock price. The expected duration depends on two factors. First, the

duration depends on the uncertainty of the news. The more uncertain the valuation signal, the less

likely the signal will be revealed. The VOLATILITY variable, described previously, represents this

input to duration. Second, the duration depends on whether, and if so how quickly, the withheld news

is likely to be made public by external parties (e.g., an analyst or the media) or by a public

announcement of macro-economic data (e.g., GDP figures that are relevant to firm valuation).

To compute PUBLIC, we estimate a factor model for each of the FF49 industries that includes

the equal-weighted market index and the equal-weighted industry index plus seven macro-economic

risk factors: short-term interest rates, default spreads, term spreads, a foreign exchange factor, a

producer price index, and the Fama-French SMB and HML factors. The factors represent observable

public signals. The power of these observable factors to explain returns for firms in the industry

provides a proxy for the degree to which news about the industry is likely to be public. The variable

PUBLIC is the adjusted-R2 from the factor model. Table 6 reports that the average (median) adjusted-

R2 used to measure the amount of information in industry returns conveyed through observable public

sources is 12.2% (11.0%). Appendix B (final column) reports the PUBLIC variable by FF49 industry.

Seven industries stand out as having substantially higher adjusted R2s: coal, precious metals, tobacco

products, defense, shipbuilding/railroad equipment, petroleum and natural gas, and electronic

equipment. The seven industries with the smallest adjusted R2s are: food products, entertainment,

consumer goods, healthcare, construction materials, personal services, and wholesale. We expect

opacity episodes are decreasing in PUBLIC.

%CRSP: The fifth explanatory variable proxies for industry-level incentives for firms to

maintain an artificially elevated stock price. In industries in which more managers have equity

incentives, any individual manager is more likely to have incentives to remain opaque. If this industry

27

characteristic is common knowledge, capital market pressure may not induce disclosure, making opacity

episodes more likely in high-incentive industries. Our industry-year proxy for equity incentives is the

number of firms in the industry with data on CRSP divided by the number of firms with data on

Compustat (%CRSP). For CRSP, we count the firms in each industry-year with greater than 200 daily

return observations. For Compustat, we count the firms in each industry-year in the fundamentals

annual file with an SIC code, sales, and total assets. Firms in industries with a higher percent of firms

on CRSP are less likely to disclose adverse signals and more likely to conjecture that other firms in the

industry have similar incentives. Appendix B presents the average annual %CRSP measures by

industry. Industries with low percentages are: Shipbuilding/Railroad Equipment, Rubber and Plastic

Products, Utilities, and Apparel, ignoring the “Almost Nothing” industry. Industries with high

percentages are: Banking, Precious Metals, and Trading. The average (median) reported in Table 6 is

122.4% (110.3%) and the inter-quartile range is 88.9% to 125.8% (not reported).

LIT_RISK: The litigation risk variable is motivated by several factors. Litigation risk encourages

individual managers to disclose adverse signals. If litigation risk pressure is common knowledge,

opacity episodes are less likely ceteris paribus. Litigation risk, however, is likely to be a function of the

uncertainty of valuation signals. Thus, in a repeated game setting the implications of litigation risk are

ambiguous. Expected litigation costs decline as the probability that the shock will materialize or be

discovered declines, increasing the sustainability of tacit collusion. If, however, litigation costs are

increasing in the duration that the signal is withheld, and this is common knowledge, opacity episodes

are less likely. Our proxy for litigation risk (LIT_RISK) uses Model 3 from Table 7 of Kim and Skinner

(2012), which is intended to measure the ex ante probability of litigation. See Appendix A for a

description of their model and our sample. We estimate LIT_RISK at the firm level for fiscal years

between 1996 and 2008 and create an industry-year average.

HERF: The model includes a revenue-based herfindahl index as a measure of industry-year

concentration. We include concentration in the model as a summary measure of various industry

28

characteristics that could be associated with incentives or opportunities for coordinated opacity.18

Greater industry concentration indicates that there are a smaller number of larger and perhaps more

influential firms in an industry, which may enable easier coordination due to explicit collusion

opportunities or because it is more likely that a focal point is available.19 Greater concentration may

also be associated with firms’ conjectures that other firms have received a similar signal and believe

each firm has similar conjectures, and that the signal affects all other firms in the same direction. In

addition, greater industry concentration can be associated with high entry barriers, in which case firms

collectively have incentives to remain opaque.

Table 6 Panel B reports the correlation matrices for the explanatory variables separately for the

observations that will be included in the FULL sample logit analysis and in the 1SEG sample logit

analysis. The correlations with industry size are also reported although industry size is not included in

the model.20 Our measures of uncertainty (VOLATILITY) and homogeneity (IND_WIDE) are highly

correlated in both samples. We run the analyses without the VOLATILITY measure and results

remain qualitatively the same using all three measures of opacity episodes (untabulated). In addition,

our control variable for litigation risk (LIT_RISK) is highly correlated with a number of the other

independent variables. We consider the implications of these correlations in the analysis and discuss

the results in the next section.

4.2 Results

Table 7 reports the marginal effects from estimating the logit model of determinants. Panel A

presents results when we define opacity episodes based on significant increases in average industry

18 We also compute an asset based herfindahl index and the concentration of the top four, six, or eight firms in each industry and year based on market shares of total sales and assets. See Appendix A for details. We report the herfindahl index as it gives some weight to “medium-sized” firms, however, results are consistent with concentration ratios. 19 Ivaldi et al. (2003) note this explanation for a relation between industry concentration and collusion even in product markets but claim that there is little evidence. Instead, concentration ratios are commonly predicted to be associated with tacit collusion in product markets because concentration determines the profitability of a collusive price (or quantity) strategy relative to the non-collusive strategy, and the short-term profits from deviating, both of which affect firms’ equilibrium beliefs about whether the collusive product-market strategy is sustainable. 20 We estimate industry size based on the number of firms listed in Compustat for the respective industry-year.

29

FOG using the FULL sample (OPACITY). Panel B column (1) presents results when we define

opacity episodes as the subset of opacity episodes in which the influential firms have significant positive

unexplained FOG (OPACITYINFL). Panel B column (2) presents results when we identify opacity

episodes using the 1SEG sample that excludes firms with multiple business segments (OPACITY1SEG).

In Panel A, there are 17 industry-year opacity episodes and 120 non-opacity industry-year observations.

In Panel B, there are 12 (14) industry-year opacity episodes and 125 (112) non-opacity industry-year

observations using the OPACITYINFL (OPACITY1SEG) measures.

The first result is that opacity episodes are (weakly) positively related to the industry’s

propensity for significant industry-wide negative shocks (TAILRISK) using both OPACITY and

OPACITYINFL. This relation is magnified when the firms in the industry are more homogenous as

measured by the interaction between TAILRISK and IND_WIDE. Opacity episodes are significantly

negatively related to the main effect of intra-industry homogeneity (IND_WIDE) in Panel A, though

the predicted association was ambiguous and we find no results in either specification in Panel B. On

one hand, valuation signals in homogeneous industries are more likely to have a common component

that affects all firms similarly and have the effect be common knowledge, which makes withholding

more sustainable. On the other hand, intra-industry homogeneity increases the likelihood that investors

will infer industry-wide news from the disclosure by one firm, increasing the costs of withholding and

making coordinated opacity less sustainable.

The second result is that opacity episodes are more likely in industries with greater equity

incentives (%CRSP). We observe this result using both OPACITY and OPACITYINFL. One

explanation for the positive association is a numerator effect – industries characterized by greater equity

incentives, and thus greater benefits of maintaining an artificially inflated stock price, exhibit

coordinated opacity. A second explanation, however, is that the result is driven by the denominator

effect – industries in which firms file annual reports despite not having publicly traded equity. If such

industries are subject to significant mandated disclosure requirements, and this industry characteristic is

30

common knowledge, firms in these industries are less likely to view opacity as a sustainable equilibrium.

Therefore, opacity will (endogenously) not be sustainable. Additional analysis in Section 4.3 attempts

to disentangle these explanations.

The third result is that opacity episodes are less likely when industry information is more

publicly available, though this finding is only significant when using the OPACITYINFL proxy for

opacity episodes. Greater availability of public information shortens the expected duration of the

elevated stock price, thus decreasing the likelihood that correlated opacity is sustainable. In short, more

observable public information deters firms that receive adverse valuation signals from coordinating

their disclosure decisions and collectively remaining opaque.

The propensity for an opacity episode is increasing in litigation risk. This result is not

consistent with the assertion that litigation risk is a market force that encourages disclosure. Our proxy

for litigation costs (LIT_RISK) is based on a model of suits that were brought against firms in an

industry. Although the litigation risk model controls for fundamentals including return skew and return

volatility, the positive sign is consistent with the proxy for litigation risk reflecting greater uncertainty

about valuation signals. Higher uncertainty will lead firms to believe that future valuation signals may

increase and that this outcome and these beliefs are common knowledge, which increases the

sustainability of opacity as an equilibrium.

As noted previously, LIT_RISK is highly correlated with several explanatory variables. We

estimate eqn. (3) excluding LIT_RISK (untabulated). The results for these other variables remain

qualitatively similar with the following exceptions. Using the FULL sample (OPACITY), the results

related to TAILRISK, IND_WIDE, and the interaction term are more significant and the results related

to %CRSP weaken slightly, although they are still significant at conventional levels.

Opacity episodes are positively related to industry concentration (HERF) for the episodes

identified by the single segment firm sample in Panel B. There are several explanations for a positive

association. Higher concentration could reflect a smaller number of influential firms such that firms are

31

more likely to conjecture that other firms will share similar adverse signals and that these conjectures

are common knowledge. Higher concentration could also be correlated with softer factors that can

sustain cooperation, such as the existence of a focal point. Finally, higher concentration could imply

greater barriers to entry, which decrease firms’ incentives to disclose.

Opacity episodes are not related to industry uncertainty as measured by return volatility

(VOLATILITY). We expected that increased uncertainty could increase managers’ collective beliefs

that an adverse valuation signal would reverse and thus increase the duration of the artificially inflated

stock price.

4.3 Additional analyses

Our first additional analysis is meant to evaluate whether changes in correlated fundamentals

omitted from the FOG model identifying opacity episodes explain our findings. We conduct a

falsification test using the transparency episodes (TRANSP) described in Section 3, which represent

periods of industry-wide significant decreases in FOG (i.e., increases in transparency). If omitted

correlated fundamentals explain our findings, we expect the explanatory variables in the logit model to

explain transparency episodes as well as opacity episodes. The industry-level explanatory variables that

we investigate are predicted to be associated with increases in correlated opacity that reflect disclosure

choice, but they should not predict increases in correlated transparency unless they measure correlated

fundamentals.

There are 13 significant industry-year decreases or transparency episodes and 167 non-

transparency observations. Table 8 column (2) reports the marginal effects from estimating the logit

model of determinants. The results for the opacity episodes are reported in column (1) for convenient

comparison. None of the variables that have explanatory power for the opacity episodes have

explanatory power for the transparency episodes.

Our second additional analysis is meant to distinguish two explanations for the positive

association between %CRSP and opacity episodes reported in Table 7. Our motivation for

32

investigating %CRSP as a determinant of opacity episodes is that firms in industries with high equity

incentives are more likely to tacitly collude. While the positive association between opacity and %CRSP

is consistent with this explanation, a second explanation is that firms in industries with low %CRSP are

less likely to collude because of existing mandatory disclosure regulation. If firms believe that peer firms

will make the valuation signals public due to mandatory disclosure, then each individual firm has

incentives to disclose and coordinated opacity is not sustainable.

In an attempt to distinguish these two explanations, we analyze whether the positive relation

between %CRSP and opacity episodes comes from the low end or the high end of the distribution of

%CRSP (or both). Table 9 reports the results for opacity episodes defined using the FULL sample

(Panel A) and for the subset of opacity episodes for which influential firms have abnormal FOG (Panel

B). We do not present the results for the single segment sample using OPACITY1SEG due to the small

sample and lack of results for the %CRSP variable in this specification in Table 7.

We use two model specifications. In the first test specification, we expand the logit model to

include an interaction term that distinguishes the low end of the %CRSP distribution. We identify high

and low equity incentives using an indicator variable (LOW) equal to 1 for observations that are less

than or equal to the median industry-year level of %CRSP. The results in Column (1) show the

marginal effect of %CRSP declines but remains positive and significant. The marginal effect of the

interaction term (%CRSP*LOW) is negative but not significant. It is worth noting that the marginal

probabilities for the other explanatory variables in the analysis in Column (1) of both panels are similar

to those in Table 7 in terms of sign, magnitude, and significance.

In the second specification, we estimate the logit model separately for industry years with low

(LOW = 1) and high (LOW = 0) equity incentives.21 We use this specification because the

interpretation of interaction terms in logit models can be unreliable (Ai and Norton, 2003). In both

panels, the marginal effect of %CRSP is significantly positive for industry-years with high equity 21 The results are the same if we define low (high) equity incentives as %CRSP<1 (%CRSP≥1).

33

incentives (Column 3) and insignificant for industry years with low equity incentives (Column 2). These

results suggest that the explanatory power for the opacity episodes derives from the high end of the

%CRSP distribution, which represents managers who are more likely to benefit from keeping the

adverse signal hidden from equity markets, and who believe other firms in the industry share those

benefits and beliefs.

5. Conclusion

This study empirically examines the determinants of financial reporting opacity. The analysis is

distinct from existing studies in that we assume firms’ valuation signals contain an industry-wide

component and we assume managers make disclosure decisions anticipating future opportunities to

disclose and the potential for signal revision. These two assumptions provide a realistic depiction of

the setting in which managers operate and inform the debate on disclosure mandates. In particular, the

assumption of an industry-wide component in valuation signals would suggest that capital market

pressure to disclose adverse signals is strong enough such that disclosure mandates are unnecessary.

However, in a repeated game setting that allows opportunities for tacit collusion, opacity can be an

optimal disclosure choice. Opacity, therefore, is determined by strategic interactions in disclosure

choice among firms in an industry. Because tacit collusion depends on the relative magnitudes of the

costs and benefits of disclosure for each individual firm and the firm’s conjectures about these costs

and benefits for all firms in the industry, whether we expect to see withholding of industry-wide bad

news in practice is an empirical question.

We do observe periods of correlated opacity that are not explained by changes in correlated

fundamentals, which suggests that market forces are not entirely sufficient to induce full disclosure.

The number of “opacity episodes,” however, is small. Of 500 (444) possible industry-year

combinations for the full (single segment) sample of firms, we find between 12 and 17 opacity episodes

(depending on the sample and method of identification) in which an increase in intra-industry

34

disclosure opacity suggests tacit collusion. Two years after the identified opacity episodes, we see

evidence of significantly lower adjusted returns for the samples of opaque firms. These results suggest

that the opacity represents adverse news.

Our analysis of the determinants of the identified opacity episodes suggests the following. First,

the opacity episodes are more likely in industries that face a greater likelihood of a significant negative

shock, especially if the firms in these industries are more homogenous. Second, withholding is more

likely in industries with a greater proportion of firms having equity incentives, in which the benefits of

maintaining an elevated stock price are more likely to exceed the costs of withholding, and this would

be common knowledge. Third, there is weak evidence that withholding is less likely in industries in

which the news is likely to be public or become public quickly. Finally, withholding is more likely in

industries with greater litigation risk.

35

References

Acharya, Viral V., Peter DeMarzo and Ilan Kremer, 2011, Endogenous information flows and the clustering of announcements, American Economic Review 101: 2955-2979.

Ai, Chunrong and Edward C. Norton, 2003, Interaction terms in logit and probit models, Economics Letters 80: 123-129.

Bertomeu, Jeremy and Pierre Jinghong Liang, 2008, Dynamic voluntary disclosure with endogenous proprietary costs, Working paper.

Bloomfield, Robert J., 2002, The “incomplete revelation hypothesis” and financial reporting, Accounting Horizons 16: 233-243.

Bloomfield, Robert, 2008, Discussion of “Annual report readability, current earnings, and earnings persistence,” Journal of Accounting and Economics 45 (248-252).

Callen, Jeffrey L., Mozaffar Khan, and Hai Lu, 2013, Accounting quality, stock price delay, and future stock returns, Contemporary Accounting Research 30: 269–295.

Chen, Shuping, Dawn Matsumoto, and Shiva Rajgopal, 2011, Is silence golden? An empirical analysis of firms that stop giving quarterly earnings guidance, Journal of Accounting and Economics 51 (102): 134-150.

Dunn, Kimberly and Siva Nathan, 2005, Analyst industry diversification and earnings forecast accuracy, Journal of Investing 14: 7-14.

Dye, Ronald A., 1985, Disclosure of nonproprietary information, Journal of Accounting Research 23: 123 -145.

Dye, Ronald A. and Sri S. Sridhar, 1995, Industry-wide disclosure dynamics, Journal of Accounting Research 33, (1): 157-174.

Easterbrook Frank H. and Daniel R. Fischel, 1984, Mandatory disclosure and the protection of investors, Virginia Law Review 70: 669-715.

FASB exposure draft, 2014, Proposed statement of financial accounting concepts: Conceptual framework for financial reporting. Chapter 8: Notes to financial statements. Exposure Draft issued March 4, 2014. File Reference No. 2014-200.

Francis, Jennifer, Donna Philbrick, and Katherine Schipper, 1994, Shareholder litigation and corporate disclosures, Journal of Accounting Research 32: 137-164.

Graham, John R., Campbell R. Harvey, and Shiva Rajgopal, 2005, The economic implications of corporate financial reporting, Journal of Accounting and Economics 40 (1-3): 3-73.

Ivaldi, Marc, Bruno Jullien, Patrick Rey, Paul Seabright, and Jean Tirole, 2003, The economics of tacit collusion, Final report for DG competition, European commission.

Jorgensen, Bjorn N. and Michael T. Kirschenheiter, 2012, Interactive discretionary disclosure, Contemporary Accounting Research 29, 382-397.

Jung and Kwon, 1988, Disclosure when the market is unsure of information endowment of managers, Journal of Accounting Research 26: 146-153.

Karpoff, Jonathan M., D. Scott Lee, Gerald S. Martin, 2008a, The consequences to managers for financial misrepresentation, Journal of Financial Economics 88: 193–215.

36

Karpoff, Jonathan M., D. Scott Lee, Gerald S. Martin, 2008b, The cost to firms of cooking the books, Journal of Financial and Quantitative Analysis 43: 581-611.

Kim, Irene and Douglas J. Skinner, 2012, Measuring securities litigation risk, Journal of Accounting and Economics 53 (1-2): 290-310.

Kothari, S.P., Susan Shu, and Peter D. Wysocki, 2009, Do managers withhold bad news?, Journal of Accounting Research 47: 1-36.

Lehavy, Reuven, Feng Li, and Kenneth Merkley, 2011, The effect of annual report readability on analyst following and the properties of their earnings forecasts, The Accounting Review 86, 1087-1115.

Li, Feng, 2008, Annual report readability, current earnings, and earnings persistence, Journal of Accounting and Economics 45: 221-247.

Levenstein, Margaret C. and Valerie Y. Suslow, 2006, What determines cartel success?, Journal of Economic Literature 44 (1): 43-95.

O’Brien, Patricia C.,1990, Forecast accuracy of individual analysts in nine industries, Journal of Accounting Research 28 (2): 286-304.

Pindyck, Robert S. and Daniel L. Rubinfeld, 1989, Microeconomics, MacMillan Publishing Company, New York.

Rajan, Raghuram G., 1994, Why bank credit policies fluctuate: A theory and some evidence, The Quarterly Journal of Economics 109 (2): 399-441.

Rogers, Jonathan L. and Andrew Van Buskirk, 2009, Shareholder litigation and changes in disclosure behavior, Journal of Accounting and Economics 47 (1/2): 136-156.

Schrand, Catherine M. and Sarah L.C. Zechman, 2012, Executive overconfidence and the slippery slope to financial misreporting, Journal of Accounting and Economics 53: 311-329.

Shin, Hyun S., 2003, Disclosures and asset returns, Econometrica 71 (2003):105 - 133.

Shin, Hyun S., 2006, Disclosure risk and price drift, Journal of Accounting Research 44: 351 - 379.

Shumway, Tyler, 1997, The delisting bias in CRSP data, Journal of Finance 52: 327-340.

Skinner, Douglas J., 1994, Why firms voluntarily disclose bad news?, Journal of Accounting Research 32, 38-60.

Skinner, Douglas J., 1997, Earnings disclosure and stockholder lawsuits, Journal of Accounting and Economics, 23: 249-282.

Sletten, Ewa, 2012, The effect of stock price on discretionary disclosure, Review of Accounting Studies 17: 96-133.

Tse, Senyo and Jennifer Wu Tucker, 2010, Within-industry timing of earnings warnings: do managers herd?, Review of Accounting Studies 15: 879-914.

Van Buskirk, Andrew, 2011, Volatility skew, earnings announcements, and the predictability of crashes, Available at SSRN: http://ssrn.com/abstract=1740513.

37

Table 1. Determinants of FOG scores by industry

OLS estimates from pooled regressions of eqn. (1) by Fama-French 49 industry over the years 1994 through 2008. Firm-year FOG scores are regressed on: Log of market value of equity (MVE), Market to book ratio (MTB), = Number of years a firm has been listed on CRSP (Firm age), Special items scaled by total assets (Special items), Standard deviation of monthly stock returns for the 12 months ending in the third month of the current year (Return volatility), Number of non-missing items in Compustat as a measure of financial complexity (Non-missing items), an indicator = 1 if the firm was incorporated in Delaware (Delaware), log of the number of geographic segments plus one (GEO segments), and log of the number of business segments plus one (BUS segments). Panel A presents results for the FULL sample. Panel B presents results for the 1SEG sample. *, **, and *** equal 10%, 5%, and 1% significance at the 2-tailed level. Panel A: FULL sample

Intercept MVE MTB Firm age Special items

Return volatility

Non-missing items Delaware

GEO segments

BUS segments

Number of estimates: Positive, (p-value<0.10) 17 6 12 1 12 32 20 4 9 Negative, (p-value<0.10) 11 11 12 13 6 1 5 18 8 Not significant 20 31 24 34 30 15 23 26 31 By industry: Agriculture 13.092 *** 0.170 * -0.062 0.005 -4.661 ** 2.408 * 0.011 ** -0.042 0.52 1.637 *** Food Products 16.444 *** -0.048 -0.058 0.001 -1.516 0.167 0.009 *** -0.021 0.117 0.439 ** Candy & Soda 15.707 *** 0.478 *** -0.297 ** 0.047 -4.487 * 6.048 * -0.009 2.100 *** 0.414 - Beer & Liquor 16.455 *** -0.134 * 0.246 *** 0.005 -3.012 -1.212 0.009 ** -0.409 0.076 1.126 *** Recreation 18.169 *** 0.125 *** 0.050 0.013 ** -0.317 1.407 * 0.003 -0.371 ** -0.275 -0.305 Entertainment 19.448 *** 0.177 *** -0.117 ** 0.015 ** -1.466 ** 0.984 0.000 0.476 *** -0.486 ** -0.670 ** Printing/Publishing 19.040 *** -0.265 *** -0.202 ** 0.008 0.315 -1.244 0.007 *** 0.431 ** -0.144 0.349 Consumer Goods 17.312 *** 0.018 0.008 0.002 -2.125 *** 1.197 0.006 *** -0.174 * -0.272 * 0.368 * Apparel 18.987 *** -0.068 * 0.054 0.007 * -0.017 -0.844 0.001 0.079 0.129 -0.109 Healthcare 19.449 *** 0.045 -0.051 * -0.002 -0.556 0.419 0.003 0.081 0.304 -0.521 ** Medical Equipment 18.079 *** -0.069 *** 0.032 ** -0.014 *** -0.778 * 1.018 ** 0.008 *** 0.008 0.02 -0.036 Pharma. Products 18.487 *** -0.048 *** 0.001 -0.002 0.454 * -0.733 *** 0.007 *** 0.207 *** -0.178 ** 0.022 Chemicals 17.533 *** 0.013 -0.051 -0.006 ** -1.119 * 1.565 * 0.008 *** 0.209 ** -0.456 *** 0.349 * Rubber/Plastic Prods 17.859 *** -0.022 0.048 -0.002 -1.088 0.304 0.009 *** 0.421 *** -0.170 -0.636 ** Textiles 20.482 *** 0.287 ** -0.462 ** 0.016 0.848 5.341 *** -0.010 * -0.236 -0.677 0.318 Construction Mats 17.491 *** 0.089 ** 0.041 0.004 -0.125 1.421 0.005 *** 0.262 ** -0.264 * -0.109 Construction 19.123 *** -0.081 ** 0.083 0.008 * -2.200 * -1.819 ** 0.004 * 0.216 * -0.337 ** 0.332 Steel Works Etc. 17.235 *** -0.074 ** -0.009 0.009 *** 0.155 -0.561 0.009 *** 0.081 0.071 -0.062 Fabricated Products 18.931 *** -0.001 0.314 0.013 * -1.551 0.51 0.003 0.135 -0.981 *** -0.158 Machinery 17.670 *** -0.059 *** 0.096 *** 0.005 ** -1.049 * 0.784 * 0.005 *** 0.176 *** 0.112 -0.197 Electrical Equipment 17.013 *** 0.055 -0.030 -0.011 ** -0.252 0.739 0.010 *** 0.314 *** -0.182 0.009 Automobiles/Trucks 17.086 *** -0.084 * -0.024 0.001 -0.323 0.944 0.008 *** -0.135 0.309 * 0.219

38

Aircraft 15.940 *** 0.274 *** 0.364 * -0.008 0.537 0.272 0.009 *** 0.646 *** -0.552 * -0.403 Ship/Railroad Equip. 16.571 *** -0.148 -0.116 0.006 0.224 -0.571 0.013 *** 0.374 0.457 -0.465 Defense 16.990 *** 0.302 *** -0.033 0.002 0.322 -1.413 0.005 -1.299 *** 0.684 -0.974 Precious Metals 17.831 *** -0.015 0.063 -0.017 *** 0.508 0.242 0.006 * -0.007 -0.349 -0.076 Mining 18.396 *** 0.029 0.137 -0.005 1.264 -0.838 0.006 -0.678 ** -0.957 *** 0.950 * Coal 13.377 *** -0.022 0.665 * 0.033 * -5.387 * -9.711 0.001 1.139 2.022 *** 2.386 *** Petroleum/Nat Gas 18.574 *** 0.013 0.038 0.007 ** 0.858 -0.484 0.004 *** -0.049 -0.261 *** 0.190 Utilities 17.984 *** 0.179 *** -0.564 *** -0.002 -3.389 * 2.351 ** 0.005 *** -0.058 -0.087 0.007 Communication 18.859 *** 0.040 * -0.070 ** -0.004 0.389 0.494 0.005 *** -0.089 0.100 -0.425 ** Personal Services 18.367 *** 0.033 -0.073 0.005 -1.522 -0.198 0.005 ** 0.311 ** -0.362 * 0.441 * Business Services 19.227 *** -0.002 -0.012 -0.003 -0.259 0.142 0.004 *** 0.174 *** -0.289 *** -0.048 Computers 18.378 *** 0.077 *** -0.023 -0.014 *** 0.102 -0.324 0.004 *** 0.047 -0.269 ** 0.570 *** Computer Software 18.478 *** -0.023 * 0.006 -0.007 *** 0.253 -0.980 *** 0.007 *** 0.152 *** 0.174 *** -0.279 ** Electronic Equip. 18.714 *** 0.037 ** 0.015 -0.009 *** -0.084 -0.498 * 0.005 *** -0.083 -0.157 *** -0.071 Meas/Control Equip. 17.605 *** -0.027 0.074 *** -0.001 -0.431 0.058 0.009 *** 0.196 *** -0.473 *** 0.063 Business Supplies 18.239 *** 0.055 -0.184 * 0.007 * -0.682 -0.073 0.004 * 0.518 *** -0.572 *** 0.234 Shipping Containers 19.106 *** -0.110 -0.020 0.025 *** -4.129 * 1.938 -0.001 0.098 -0.221 -0.054 Transportation 18.712 *** 0.035 0.007 0.005 * -2.857 *** -0.973 0.001 0.305 *** -0.357 *** 0.099 Wholesale 18.691 *** 0.142 *** -0.085 ** -0.006 ** -1.606 *** 1.494 *** 0.001 0.12 0.086 -0.368 *** Retail 17.328 *** 0.062 ** -0.062 * -0.014 *** -0.501 0.079 0.008 *** 0.242 *** -0.267 * 0.214 Restaurants, Hotels 18.748 *** 0.059 -0.025 0.006 -0.145 0.348 0.002 -0.166 -0.008 0.002 Banking 18.730 *** 0.115 *** -0.186 *** -0.017 ** 0.129 1.376 * 0.007 *** 0.099 -0.258 -0.579 ** Insurance 19.334 *** -0.042 ** 0.078 -0.001 -1.060 -1.982 *** 0.006 *** 0.116 ** -0.049 -0.368 *** Real Estate 17.229 *** 0.102 ** -0.004 -0.004 -0.970 3.073 ** 0.004 0.04 0.865 ** -0.167 Trading 18.550 *** 0.077 *** 0.015 -0.011 *** 0.009 -0.833 * 0.004 *** 0.322 *** 0.193 -0.057 Almost Nothing 19.209 *** 0.057 0.026 -0.010 * -0.343 1.477 * 0.004 -0.342 ** -0.144 -0.168

39

Panel B: 1SEG sample

Intercept MVE MTB Firm age Special items

Return volatility

Non-missing items Delaware

GEO segments

Number of estimates: Positive, (p-value<0.10) 12 7 9 2 8 27 19 2 Negative, (p-value<0.10) 10 6 10 8 8 0 4 13 Not significant 26 35 29 38 32 21 24 33 By industry: Agriculture 19.262 *** 0.068 0.175 -0.014 -2.643 1.497 -0.002 -0.569 * -0.673 Food Products 16.361 *** -0.060 -0.037 -0.001 -1.711 0.398 0.011 *** 0.061 0.222 Candy & Soda 15.707 *** 0.478 *** -0.297 ** 0.047 -4.487 * 6.048 * -0.009 2.100 *** 0.414 Beer & Liquor 18.414 *** -0.399 *** 0.195 * 0.030 ** -0.875 -1.868 0.009 * -0.251 0.007 Recreation 18.030 *** 0.082 0.022 0.034 *** 0.151 1.172 0.003 -0.286 * -0.312 Entertainment 18.898 *** 0.245 *** -0.151 *** 0.010 -2.100 *** 2.326 *** -0.001 0.380 *** -0.391 Printing/Publishing 19.452 *** -0.271 *** -0.079 0.031 ** 0.494 -1.474 0.002 0.856 *** 0.138 Consumer Goods 18.010 *** 0.017 0.003 0.004 -1.920 ** 1.143 0.004 ** -0.170 -0.172 Apparel 18.909 *** -0.105 ** 0.108 * 0.006 0.572 -1.367 0.002 0.311 ** 0.005 Healthcare 19.155 *** 0.069 ** -0.053 * -0.005 -0.692 0.771 0.002 0.191 0.280 Medical Equipment 18.144 *** -0.102 *** 0.033 ** -0.020 *** -0.546 0.704 0.008 *** 0.063 0.002 Pharma. Products 18.572 *** -0.050 *** -0.001 -0.003 0.563 ** -0.772 *** 0.007 *** 0.181 *** -0.187 ** Chemicals 17.518 *** 0.015 -0.016 -0.002 -1.356 * 2.067 ** 0.008 *** 0.141 -0.392 ** Rubber/Plastic Prods 18.303 *** 0.019 0.053 -0.006 -0.225 -0.592 0.008 *** 0.390 ** -0.859 *** Textiles 19.686 *** 0.325 ** -0.335 0.015 0.984 3.620 -0.010 -0.055 -0.149 Construction Mats 16.738 *** 0.021 0.406 *** 0.011 -1.582 0.627 0.005 * 0.400 ** 0.208 Construction 19.191 *** -0.149 *** 0.115 0.005 -2.698 * -2.060 ** 0.006 ** 0.459 *** -0.587 *** Steel Works Etc. 17.813 *** -0.085 -0.012 0.008 * 0.828 0.522 0.008 *** 0.029 -0.298 Fabricated Products 19.245 *** -0.088 0.447 -0.007 -1.407 -1.411 0.006 -0.314 -0.872 * Machinery 17.219 *** -0.120 *** 0.123 *** 0.002 -0.516 0.707 0.007 *** 0.225 *** 0.235 * Electrical Equipment 17.642 *** -0.020 -0.024 0.003 0.090 1.297 0.007 *** 0.292 * -0.152 Automobiles/Trucks 17.585 *** -0.073 0.021 0.016 *** -0.572 1.710 0.004 -0.098 0.288 Aircraft 14.485 *** 0.184 0.113 -0.017 3.024 1.365 0.012 ** 1.053 *** 0.465 Ship/Railroad Equip. 18.554 *** -0.200 -0.177 0.017 -1.196 -5.511 ** 0.013 * 0.026 -0.877 Defense 25.183 ** -1.333 -0.018 0.069 1.373 -8.907 0.015 0.918 -1.878 Precious Metals 17.523 *** -0.034 0.071 -0.021 *** 0.726 0.551 0.007 * 0.049 -0.202 Mining 19.556 *** -0.212 * 0.199 0.013 -1.373 -2.846 * 0.007 -0.295 -0.803 * Coal 20.665 *** -0.093 0.444 0.033 -0.805 -6.960 -0.014 - 1.884 Petroleum/Nat Gas 18.752 *** 0.005 0.044 0.010 *** 0.924 -0.702 0.003 * -0.058 -0.202 * Utilities 17.375 *** 0.229 *** -0.166 -0.004 -4.206 5.666 *** 0.004 -0.077 -0.018

40

Communication 19.470 *** -0.021 -0.065 * -0.001 0.848 ** 0.185 0.003 ** -0.097 0.016 Personal Services 18.585 *** 0.072 -0.089 0.001 -0.983 -0.582 0.008 *** 0.056 -0.863 *** Business Services 19.274 *** -0.014 -0.013 -0.007 ** -0.081 0.095 0.003 *** 0.188 *** -0.187 ** Computers 19.023 *** 0.065 *** -0.024 -0.020 *** 0.002 -0.402 0.003 ** 0.080 -0.304 *** Computer Software 18.410 *** -0.027 ** 0.006 -0.007 ** 0.254 -0.889 *** 0.007 *** 0.170 *** 0.136 ** Electronic Equip. 18.741 *** 0.032 * 0.013 -0.009 *** 0.095 -0.678 ** 0.005 *** -0.130 ** -0.217 *** Meas/Control Equip. 17.672 *** -0.028 0.072 *** 0.003 -0.650 0.468 0.009 *** 0.213 ** -0.588 *** Business Supplies 19.351 *** 0.135 ** -0.164 0.015 ** -1.175 1.471 -0.001 0.720 *** -1.290 *** Shipping Containers 19.884 *** -0.169 -0.043 0.043 *** -4.266 * 2.416 -0.004 0.567 -0.412 Transportation 18.529 *** 0.047 -0.034 0.012 *** -2.262 ** 0.126 0.001 0.195 ** -0.186 Wholesale 18.549 *** 0.123 *** -0.120 ** -0.006 -1.745 ** 1.841 *** 0.001 0.116 0.130 Retail 17.556 *** 0.043 -0.028 -0.016 *** -0.360 -0.204 0.008 *** 0.213 ** -0.195 Restaurants, Hotels 18.879 *** 0.045 -0.002 0.007 -0.154 -0.804 0.002 -0.259 * 0.024 Banking 18.097 *** 0.131 *** -0.169 *** -0.023 ** -0.117 1.748 * 0.006 ** 0.119 -0.037 Insurance 19.153 *** -0.039 ** 0.045 0.002 -0.547 -1.837 *** 0.005 *** 0.247 *** 0.065 Real Estate 18.373 *** 0.094 0.038 -0.015 * -2.321 4.735 *** -0.002 -0.161 0.752 Trading 18.724 *** 0.068 *** 0.017 * -0.011 *** -0.061 -0.917 ** 0.003 *** 0.348 *** 0.156 Almost Nothing 18.737 *** 0.112 ** 0.037 0.017 0.201 1.897 * 0.002 -0.263 -0.094

41

Table 2: Measures of FOG score dispersion by industry

Panel A (Panel B) presents measures of dispersion for each FF49 industry for the FULL sample (the 1SEG sample that excludes firms with multiple business segments). CV is the coefficient of variation, which is the sample standard deviation/sample mean. QD is the quartile coefficient of dispersion, which is quartile 3 – quartile 1 divided by the median. The last column indicates the significance of the clustering. “c” indicates that equality of variances is rejected for the industry vs. the pooled sample reference group (n = 46,242 in the FULL sample and 34,172 in the 1SEG sample) and the industry exhibits greater clustering. “x” indicates that equality of variances is rejected and the industry FOG scores are more disperse. “cc” and “xx” are the same but based on a more conservative test statistic.

Panel A: FULL sample n CV QD SIG

Pooled sample reference group 46,242 0.0742 0.0901

1 Agriculture 145 0.0801 0.1049 x 2 Food Products 575 0.0726 0.0900 cc 3 Candy & Soda 54 0.0852 0.1307 4 Beer & Liquor 122 0.0632 0.0733 5 Tobacco Products 54 0.0870 0.0924 6 Recreation 415 0.0741 0.1033 c 7 Entertainment 706 0.0769 0.0961 x 8 Printing and Publishing 367 0.0778 0.1070 x 9 Consumer Goods 889 0.0743 0.0926 xx 10 Apparel 592 0.0729 0.0829 cc 11 Healthcare 950 0.0729 0.0875 cc 12 Medical Equipment 1,569 0.0681 0.0778 13 Pharmaceutical Products 3,024 0.0659 0.0767 cc 14 Chemicals 879 0.0705 0.0870 c 15 Rubber and Plastic Products 444 0.0769 0.0946 16 Textiles 146 0.0898 0.1048 17 Construction Materials 903 0.0853 0.1077 x 18 Construction 625 0.0690 0.0785 19 Steel Works Etc. 684 0.0755 0.0890 xx 20 Fabricated Products 195 0.0790 0.0951 21 Machinery 1,641 0.0710 0.0836 c 22 Electrical Equipment 662 0.0728 0.0882 23 Automobiles and Trucks 679 0.0768 0.0914 xx 24 Aircraft 201 0.0798 0.0977 25 Shipbuilding, Railroad Equip. 91 0.0655 0.0658 26 Defense 91 0.0835 0.0997 x 27 Precious Metals 151 0.0603 0.0686 cc 28 Non-Metallic/Ind. Metal Mining 149 0.0755 0.0834 29 Coal 53 0.0861 0.1202 30 Petroleum and Natural Gas 1,710 0.0772 0.0901 31 Utilities 1,359 0.0757 0.0888 x 32 Communication 1,353 0.0729 0.0832 c 33 Personal Services 538 0.0710 0.0850 34 Business Services 2,571 0.0702 0.0873 cc 35 Computers 1,294 0.0685 0.0769 36 Computer Software 4,047 0.0648 0.0781 cc 37 Electronic Equipment 2,705 0.0665 0.0779 38 Measuring & Control Equip. 1,060 0.0624 0.0741 39 Business Supplies 556 0.0767 0.0904 x 40 Shipping Containers 149 0.0665 0.0659 c 41 Transportation 1,215 0.0783 0.0961 xx 42 Wholesale 1,810 0.0778 0.0926 43 Retail 1,725 0.0795 0.0952 xx 44 Restaurants, Hotels, Motels 542 0.0747 0.0889 45 Banking 577 0.0780 0.0963 x 46 Insurance 2,090 0.0657 0.0804 cc 47 Real Estate 364 0.0756 0.0907 48 Trading 2,995 0.0783 0.0986 xx 49 Almost Nothing 407 0.0720 0.0968 c

42

Table 2, continued.

Panel B: 1SEG sample n CV QD SIG

Pooled sample reference group 34,172 0.0734 0.0892

1 Agriculture 77 0.0669 0.1006 c 2 Food Products 403 0.0754 0.0928 xx 3 Candy & Soda 54 0.0852 0.1307 4 Beer & Liquor 66 0.0639 0.0796 5 Tobacco Products 14 0.1152 0.1515 6 Recreation 301 0.0751 0.1025 x 7 Entertainment 564 0.0771 0.0981 x 8 Printing and Publishing 174 0.0705 0.0866 x 9 Consumer Goods 625 0.0727 0.0906 cc

10 Apparel 444 0.0720 0.0830 c 11 Healthcare 743 0.0725 0.0908 cc 12 Medical Equipment 1,363 0.0681 0.0780 13 Pharmaceutical Products 2,790 0.0658 0.0763 c 14 Chemicals 557 0.0689 0.0892 c 15 Rubber and Plastic Products 259 0.0695 0.0781 16 Textiles 95 0.0894 0.0937 x 17 Construction Materials 398 0.0831 0.1039 x 18 Construction 354 0.0730 0.0793 19 Steel Works Etc. 398 0.0756 0.0891 x 20 Fabricated Products 108 0.0778 0.0930 21 Machinery 1,043 0.0714 0.0847 cc 22 Electrical Equipment 349 0.0698 0.0924 23 Automobiles and Trucks 367 0.0752 0.0821 x 24 Aircraft 75 0.0852 0.1225 25 Shipbuilding, Railroad Equip. 50 0.0724 0.0596 26 Defense 21 0.0807 0.0559 27 Precious Metals 136 0.0621 0.0714 cc 28 Non-Metallic/Ind. Metal Mining 84 0.0703 0.0788 29 Coal 23 0.0592 0.0832 30 Petroleum and Natural Gas 1,260 0.0773 0.0929 x 31 Utilities 494 0.0719 0.0858 32 Communication 940 0.0725 0.0842 c 33 Personal Services 361 0.0705 0.0780 34 Business Services 1,882 0.0698 0.0865 c 35 Computers 1,062 0.0654 0.0751 36 Computer Software 3,570 0.0637 0.0770 cc 37 Electronic Equipment 2,256 0.0658 0.0751 38 Measuring & Control Equip. 809 0.0617 0.0724 39 Business Supplies 277 0.0772 0.0890 x 40 Shipping Containers 59 0.0735 0.0935 x 41 Transportation 967 0.0800 0.1010 xx 42 Wholesale 1,157 0.0775 0.0875 43 Retail 1,394 0.0791 0.0933 xx 44 Restaurants, Hotels, Motels 400 0.0733 0.0852 45 Banking 407 0.0828 0.1022 x 46 Insurance 1,554 0.0631 0.0781 cc 47 Real Estate 241 0.0796 0.0955 48 Trading 2,775 0.0779 0.0992 x 49 Almost Nothing 275 0.0688 0.0944 c

43

Table 3: Summary of opacity and transparency episodes by Fama-French 49 industry The opacity (transparency) episodes represent a significant industry-level increase (decrease) in FOG identified in two-year rolling regressions. Columns (1) and (3) show the episodes identified using the full sample (FULL); Columns (2) and (4) show the episodes identified using the sample that excludes firms with multiple business segments (1SEG).

Opacity Episodes Transparency Episodes FF 49 Industry FULL (1) 1SEG (2) FULL (3) 1SEG (4)

1 Agric Agriculture 2 Food Food Products 1 2 3 Soda Candy & Soda 4 Beer Beer & Liquor 5 Smoke Tobacco Products 6 Toys Recreation 7 Fun Entertainment 1 8 Books Printing and Publishing 1 1 1 9 Hshld Consumer Goods 1 1

10 Clths Apparel 1 11 Hlth Healthcare 1 12 Medeq Medical Equipment 1 1 13 Drugs Pharmaceutical Products 1 1 1 1 14 Chems Chemicals 1 1 15 Rubbr Rubber and Plastic Products 1 1 16 Txtls Textiles 17 Bldmt Construction Materials 1 18 Constr Construction 19 Steel Steel Works Etc. 2 20 Fabpr Fabricated Products 1 21 Mach Machinery 1 22 Elceq Electrical Equipment 1 1 2 1 23 Autos Automobiles and Trucks 24 Aero Aircraft 1 25 Ships Shipbuilding, Railroad Equip. 26 Guns Defense 27 Gold Precious Metals 1 28 Mines Non-Metallic/Ind. Metal Mining 29 Coal Coal 30 Oil Petroleum and Natural Gas 31 Util Utilities 32 Telcm Communication 1 1 1 1 33 Persv Personal Services 34 Bussv Business Services 1 1 35 Hardw Computers 1 1 36 Softw Computer Software 2 2 1 1 37 Chips Electronic Equipment 2 2 38 Labeq Measuring & Control Equip. 1 39 Paper Business Supplies 40 Boxes Shipping Containers 41 Trans Transportation 42 Whlsl Wholesale 1 1 1 43 Rtail Retail 44 Meals Restaurants, Hotels, Motels 45 Banks Banking 46 Insur Insurance 1 47 Rlest Real Estate 1 1 48 Fin Trading 1 1 49 Other Almost Nothing 1 Total 19 17 13 12

44

Table 4: Summary of opacity and transparency episodes by year The opacity (transparency) episodes represent a significant industry-level increase (decrease) in FOG identified in two-year rolling regressions. Columns (1) and (3) show the episodes identified using the full sample (FULL); Columns (2) and (4) show the episodes identified using the sample that excludes firms with multiple business segments (1SEG). Opacity Episodes Transparency Episodes FULL (1) 1SEG (2) FULL (3) 1SEG (4) Year Fama-French 49 industry Fama-French 49 industry Fama-French 49 industry Fama-French 49 industry 1995 42 (whlsl) 1996 27 (gold) 1997 15 (rubbr)

1998 22 (elceq)

7 (fun) 15 (rubbr) 22 (elceq)

42 (whlsl)

1999

9 (hshld) 12 (medeq) 13 (drugs) 22 (elceq) 36 (softw)

2 (food) 12 (medeq) 13 (drugs) 22 (elceq) 36 (softw)

2000 8 (books)

47 (rlest) 8 (books) 19 (steel) 47 (rlest)

2001 32 (telcm) 19 (steel) 32 (telcm)

2002

24 (aero) 32 (telcm) 34 (bussv) 35 (hardw) 36 (softw) 37 (chips) 46 (insur)

32 (telcm) 34 (bussv) 35 (hardw) 36 (softw) 37 (chips) 49 (other)

2003

8 (books) 11 (hlth) 13 (drugs) 14 (chems) 20 (fabpr) 36 (softw) 37 (chips) 38 (labeq) 42 (whlsl) 48 (fin)

13 (drugs) 14 (chems) 21 (mach) 36 (softw) 37 (chips) 48 (fin)

2004 9 (hshld) 22 (elceq)

2005 2 (food) 10 (clths)

2 (food)

2006 2007 17 (bldmt) 2008 Total 19 17 13 12

45

Table 5: Returns following the opacity episodes This table presents average cumulative abnormal size-decile adjusted returns (CARs) for portfolios of firms in the opacity sample and the matched sample. The opacity sample consists of firm year observations in industry i in year y if industry i-year y is an opacity episode determined by a significant industry-level increase in FOG in two-year rolling regressions for the 1SEG sample (OPACITY1SEG). The matched sample consists of all other firm year observations, excluding firms in industry i in years y+1 through y+3 if industry i-year y is an opacity episode. Returns are presented for months 1-, 3-, 6-, 12-, 24- and 36 after month four (April) of the opacity episode year y. Panels A and B present average CARs for single segment firms and single and multi-segment firms, respectively. Panels C and D present average CARs for single segment firms in high (negative) TAILRISK and low TAILRISK industries, respectively. *, **, and *** equal 10%, 5%, and 1% significance at the 2-tailed level. Ret1mon Ret3mon Ret6mon Ret12mon Ret24mon Ret36mon Panel A: Single segment firms Opacity sample: Return 0.0367*** 0.0583*** 0.0961*** 0.0691*** -0.0205 0.0318 N 1,750 1,716 1,667 1,596 1,452 1,295 Matched sample: Return 0.0043*** -0.0015 -0.0042* 0.0036 0.0214*** 0.0259*** N 28,523 28,020 27,285 25,929 23,437 20,578 Difference in returns 0.0324*** 0.0598*** 0.1003*** 0.0655*** -0.0419** 0.0059 t-value 6.72 7.62 8.77 4.02 2.13 0.23 Panel B: Single+multi-segment firms Opacity sample: Return 0.0319*** 0.0529*** 0.0825*** 0.0660*** -0.0148 0.0301 N 2,246 2,204 2,142 2,060 1,893 1,708 Matched sample: Return 0.0033*** -0.0002 0.0004 0.0012 0.0101*** 0.0116** N 44,141 43,403 42,350 40,365 36,720 32,269 Difference in returns 0.0286*** 0.0531*** 0.0821*** 0.0648*** -0.0249 0.0185 t-value 6.98 8.07 8.55 4.73 1.53 0.87 Panel C: High TAILRISK Single segment firms Opacity sample: Return 0.0625** 0.0627 0.0439 -0.0260 -0.2979* -0.3654*** N 27 26 25 23 19 18 Matched sample: Return 0.0007 -0.0041 -0.0289*** -0.0188 -0.0042 -0.0329 N 1,977 1,936 1,873 1,760 1,549 1,285 Difference in returns 0.0618* 0.0668 0.0728 -0.0072 -0.2937 -0.3325*** t-value 1.92 1.23 0.93 0.04 1.58 3.11 Panel D: Low TAILRISK Single segment firms Opacity sample: Return 0.0589*** 0.0989*** 0.1640*** 0.1549*** 0.0194* 0.1256*** N 982 952 919 870 790 695 Matched sample: Return 0.0044*** -0.0013 -0.0021 0.0049 0.0226*** 0.0282*** N 28,128 27,649 26,938 25,642 23,239 20,208 Difference in returns 0.0545*** 0.1002*** 0.1661*** 0.1500*** -0.0032 0.0974*** t-value 7.65 8.83 9.87 6.18 0.11 2.60

46

Table 6: Descriptive statistics for industry-level characteristics

This table provides descriptive statistics on industry-level characteristics that are the explanatory variables in the logit model. Industry size is the average number of firms in the industry with non-missing sales and assets data on Compustat. TAILRISK is an indicator variable that equals one for industries that have high negative expected return skewness computed based on the methods in Van Buskirk (2011), and zero otherwise (see Appendix A). IND_WIDE is the negative measure of industry heterogeneity, which is measured as the average monthly standard deviation of the idiosyncratic component of returns within an industry in year y. %CRSP proxies for industry-level equity incentives and is the average annual number of firms with CRSP data relative to the number of firms with Compustat data (see Appendix B). VOLATILITY is the average implied volatility of at-the-money options for firms in the industry following Van Buskirk (2011). PUBLIC proxies for the availability of public information and is the adjusted R2 values from estimation of a standard market model that includes seven macro-economic factors described in Appendix A for each firm i within an industry, estimated using monthly return and factor data over the period 1994-2008. LIT_RISK is an industry-year measure of litigation risk (see Appendix A). HERF is the revenue-based herfindahl index for the 50 largest firms in the industry. Panel B presents the correlations between these variables by sample. Statistical significance in the correlation matrices at 10%, 5%, and 1% are denoted by *, **, and ***, respectively. Panel A: Average annual measures (477 industry-year observations in the FULL sample) Mean Median Std. Dev. Industry size 150 105 124.79 Negative shock (TAILRISK) 0.3774 0.0000 0.4852 Homogeneous (IND_WIDE) -0.1610 -0.1552 0.0506 Equity incentives (%CRSP) 1.0940 0.9861 0.5444 Uncertainty (VOLATILITY) 0.4474 0.4465 0.0772 Public information (PUBLIC) 0.1224 0.1103 0.0360 Litigation risk (LIT_RISK) 0.0460 0.0428 0.0176 Industry concentration (HERF) 0.1012 0.0732 0.1757

Panel B: Correlation matrix (FULL sample)

Industry size TAILRISK IND_ WIDE %CRSP VOLATIL-

ITY PUBLIC LIT_RISK

TAILRISK 0.007 IND_WIDE -0.223*** -0.014 %CRSP 0.093** 0.061 0.215*** VOLATILITY 0.315*** 0.206*** -0.587*** -0.042 PUBLIC 0.237*** -0.140*** -0.068 0.281*** 0.320*** LIT_RISK 0.344*** 0.137*** -0.234*** -0.077 0.391*** 0.296*** HERF -0.253*** 0.035 -0.031 0.074 0.134*** 0.200*** -0.042

(1SEG sample)

Industry size TAILRISK IND_ WIDE %CRSP VOLATIL-

ITY PUBLIC LIT_RISK

TAILRISK -0.077 IND_WIDE -0.214*** 0.039 %CRSP 0.093* 0.053 0.246*** VOLATILITY 0.271*** 0.122** -0.606*** -0.062 PUBLIC 0.336*** -0.123** -0.114** 0.247*** 0.356*** LIT_RISK 0.318*** 0.078 -0.253*** -0.087* 0.407*** 0.340*** HERF -0.131*** 0.152*** -0.124** 0.107** 0.245*** 0.191*** 0.048

47

Table 7: Characteristics of industries that have opacity episodes This table presents estimated marginal effects from a logit regression model of opacity episode occurrence on industry and industry-year level correlates. The unit of observation is at the industry-year level. Panel A presents results for opacity episodes determined by a significant industry-level increase in FOG in two-year rolling regressions (OPACITY). Panel B presents results for two alternative measures of opacity episodes. Column 1 defines opacity episodes as the subset of episodes from Panel A for which influential firms have a significant positive unexplained FOG (OPACITY

INFL). Column 2 defines withholding similarly to Panel A, but excludes firms with multiple business segments from the measure of withholding (OPACITY1SEG) and from the regression. See Table 6 for definitions of the independent variables. Statistical significance (one-sided if predicted, two-sided otherwise) at the 10%, 5% and 1% level is denoted by *, **, and ***, respectively. Panel A: Opacity episodes identified by significant increase in industry-average FOG for the full sample

• 17 opacity episodes; 120 non-opacity industry-year observations

Predicted OPACITY Sign Coefficient z-statistic Intercept -0.5116*** -2.81 Negative shock (TAILRISK) + 0.2791* 1.61 Homogeneous (IND_WIDE) ? -1.6853** -1.97 TAILRISK * IND_WIDE + 1.6756** 1.81 Equity incentives (%CRSP) + 0.0942*** 2.39 Uncertainty (VOLATILITY) + -0.3466 -0.74 Public information (PUBLIC) - -0.7592 -0.96 Litigation risk (LIT_RISK) ? 3.2978** 2.16 Industry concentration (HERF) ? 0.3419 1.25 Pseudo R2 22.73% Wald χ2 test statistic 13.49 Wald χ2 p-value 0.10 Panel B: Alternative measures of opacity episodes

• OPACITYINFL: 12 opacity episodes; 125 non-withholding industry-year observations • OPACITY1SEG: 14 opacity episodes; 112 non-withholding industry-year observations

Predicted (1) OPACITYINFL (2) OPACITY1SEG Sign Coefficient z-statistic Coefficient z-statistic Intercept -0.3397** -2.33 -0.3772** -2.40 Negative shock (TAILRISK) + 0.1648* 1.42 -0.0467 -0.30 Homogeneous (IND_WIDE) ? -0.5304 -0.88 -1.0126 -1.51 TAILRISK * IND_WIDE + 0.9414* 1.50 0.0214 0.03 Equity incentives (%CRSP) + 0.0677*** 2.32 0.0450 1.23 Uncertainty (VOLATILITY) + 0.1232 0.33 -0.2004 -0.49 Public information (PUBLIC) - -1.0298** -1.63 -0.2907 -0.42 Litigation risk (LIT_RISK) ? 2.0731* 1.91 1.7524 1.46 Industry concentration (HERF) ? 0.0611 0.26 0.6057* 1.84 Pseudo R2 22.29% 23.31% Wald χ2 test statistic 10.77 12.01 Wald χ2 p-value 0.22 0.15

48

Table 8: Falsification test: Characteristics of industries that have transparency episodes This table presents estimated marginal effects from a logit regression model of transparency episodes (TRANSP) on industry and industry-year level correlates. The unit of observation is at the industry-year level. Column (1) reports results from Table 7 for opacity episodes for convenient comparison. The transparency (opacity) episodes are determined by a significant industry-level decrease (increase) in FOG in two-year rolling regressions using the FULL sample. See Table 6 for definitions of the independent variables. Statistical significance (one-sided if predicted for opacity episodes, two-sided otherwise) at the 10%, 5% and 1% level is denoted by *, **, and ***, respectively. Predicted (1) OPACITY (2) TRANSPARENCY Sign Coefficient z-statistic Coefficient z-statistic Intercept -0.5116*** -2.81 -0.1565 -1.21 Negative shock (TAILRISK) + 0.2791* 1.61 0.0532 0.48 Homogeneous (IND_WIDE) ? -1.6853** -1.97 -0.3638 -0.87 TAILRISK * IND_WIDE + 1.6756** 1.81 0.2717 0.49 Equity incentives (%CRSP) + 0.0942*** 2.39 -0.0388 -0.66 Uncertainty (VOLATILITY) + -0.3466 -0.74 0.1844 0.67 Public information (PUBLIC) - -0.7592 -0.96 -0.3236 -0.63 Litigation risk (LIT_RISK) ? 3.2978** 2.16 -0.3392 -0.29 Industry concentration (HERF) ? 0.3419 1.25 -0.1835 -0.72 Pseudo R2 22.73% 6.63% Wald χ2 test statistic 13.49 6.57 Wald χ2 p-value 0.10 0.58

49

Table 9: Supplemental analysis of %CRSP as a determinant of opacity episodes This table presents estimated marginal effects from logit regression models of opacity episode occurrence on industry and industry-year level correlates allowing for a separate relation between firms with low and high equity incentives as measured by %CRSP. The unit of observation is at the industry-year level. Panels A and B present results for opacity episodes determined by a significant industry-level increase in FOG in two-year rolling regressions (OPACITY) and by a significant industry-level increase in FOG combined with positive abnormal FOG for influential firms (OPACITYINFL), respectively. Column (1) in each panel reports results for the first model from Table 7 with an additional interaction term of %CRSP and LOW, which is an indicator variable for industry-year observations that have a %CRSP less than or equal to the median industry-year level. Columns (2) and (3) in each panel report results for the subsample of observations with low equity incentives (LOW = 1) and high equity incentives (LOW = 0), respectively. Statistical significance (one-sided if predicted, two-sided otherwise) at the 10%, 5% and 1% level is denoted by *, **, and ***, respectively. Panel A: OPACITY used to identify opacity episodes22 Predicted (1) Interaction (2) LOW %CRSP (3) HIGH %CRSP Sign Coefficient z-stat Coefficient z-stat Coefficient z-stat Intercept -0.4854*** -2.61 -0.4366** -1.99 -0.5554* -1.64 Negative shock (TAILRISK) + 0.2789** 1.64 0.3218* 1.54 0.3760* 1.28 Homogeneous (IND_WIDE) ? -1.5763* -1.83 -0.7063 -0.73 -3.5577** -2.09 TAILRISK * IND_WIDE + 1.6458** 1.81 1.7930* 1.51 2.3492* 1.52 Equity incentives (%CRSP) + 0.0811** 1.78 0.2228 1.12 0.1387** 1.91 %CRSP * LOW -0.0349 -0.55 Uncertainty (VOLATILITY) + -0.2615 -0.54 -0.4435 -0.76 -0.9340 -1.19 Public information (PUBLIC) - -0.7962 -1.02 0.4584 0.42 -2.1740* -1.62 Litigation risk (LIT_RISK) ? 3.1663** 2.07 2.0593 1.42 3.9447 1.47 Industry concentration (HERF) ? 0.3031 1.09 0.4927 1.16 0.8151 1.60 Pseudo R2 23.01% 27.63% 27.88% Wald χ2 test statistic 13.83 5.51 8.80 Wald χ2 p-value 0.13 0.70 0.36 Panel B: OPACITYINFL used to identify opacity episodes23 Predicted (1) Interaction (2) LOW %CRSP (3) HIGH %CRSP Sign Coefficient z-stat Coefficient z-stat Coefficient z-stat Intercept -0.3082** -2.09 -0.0556 -0.70 -0.4990 -1.48 Negative shock (TAILRISK) + 0.1733* 1.53 0.1809 1.10 0.2089 1.14 Homogeneous (IND_WIDE) ? -0.4760 -0.81 0.1085 0.47 -1.1504 -1.10 TAILRISK * IND_WIDE + 0.9530* 1.58 1.2249 1.15 1.0700 1.12 Equity incentives (%CRSP) + 0.0524* 1.60 0.0382 0.52 0.0658* 1.32 %CRSP * LOW -0.0370 -0.85 Uncertainty (VOLATILITY) + 0.1589 0.44 -0.0093 -0.06 0.3166 0.71 Public information (PUBLIC) - -1.0153* -1.64 -0.3075 -0.81 -0.5962 -0.91 Litigation risk (LIT_RISK) ? 2.0177* 1.87 0.8240 0.91 0.9252 0.63 Industry concentration (HERF) ? 0.0407 0.18 0.0945 0.53 -0.0801 -0.29 Pseudo R2 23.16% 29.38% 42.64% Wald χ2 test statistic 10.63 1.63 2.92 Wald χ2 p-value 0.30 0.99 0.94

22 Using the OPACITY measure includes 17 collusion episodes and 120 non-collusion industry-year observations. The LOW (HIGH) %CRSP sample has 7 (10) collusion episodes; 74 (46) non-collusion industry-year observations. 23 Using the OPACITYINFL includes 12 collusion episodes and 125 non-collusion industry-year observations. The LOW (HIGH) %CRSP sample has 5 (7) collusion episodes; 76 (49) non-collusion industry-year observations.

50

Appendix A: Variable definitions

This appendix provides detailed definitions of variables that are included in the analysis. Control variables used in the FOG model

MVE = Log of market value of equity (CSHO * PRCC_F), winsorized at 1%. MTB = Market to book ratio (MVE + LT / AT), winsorized at 1%. Special items = Special items scaled by total assets (SPI/AT), winsorized at 1%. Return volatility = Standard deviation of monthly stock returns for the 12 months ending in the third month of the

current year (i.e., from current month three back to lag nine), winsorized at 1%. This variable is set to missing if there are less than 11 months of return data.

Non-missing items = Number of non-missing items in Compustat, winsorized at 1%, to proxy for complexity. The total number of items possible is based on all data items for the balance sheet, income statement, and cash flow statement per the WRDS Compustat web interface.

Firm age = Number of years a firm has been listed on CRSP. Delaware = An indicator = 1 if the firm was incorporated in Delaware (INCORP= “DE”), 0 otherwise. GEO Segments = Log of the number of geographic segments plus one, obtained from the Compustat Segment file (but

collapsed into FF49 industries). The variable is set to one if missing in the datafile. BUS Segments = Log of the number of business segments plus one, obtained from the Compustat Segment file (but

collapsed based on FF49 industries). The variable is set to one if missing in the datafile. Explanatory variables used in the logit model of the opacity episodes

TAILRISK: Industry tail risk Using option data from OptionMetrics, we calculate the Van Buskirk (2011) measure of expected volatility skew before each I/B/E/S earnings announcement that meets his data requirements. This measure is calculated as the average implied volatility for out-of-the-money puts less the average implied volatility for at-the-money calls. We compute skew for 85,293 firm-quarters between 1995 and 2010. We calculate the firm-year average skew so that firms with more populated data do not influence the measurement. The average (and median) firm-year have volatility skew of approximately 0.03 (higher values indicate more negative skewness). On average, the industries have 32.8 firms per year (median = 20.2). The minimum (maximum) is 1.7 (131.2) for industry 1 (industry 36). The inter-quartile range is 5.9 to 50.9. For each industry, we obtained the median and maximum skew over the 13 years. Industries with a maximum greater than the median industry maximum (0.061) and a median greater than the global industry median of 0.03 are designated as having high positive skew or high negative tailrisk (TAILRISK = 1). INDUSTRY_WIDE: Intra-industry homogeneity We estimate a single factor market model using monthly returns for each FF49 industry from January 1994-December 2008. We then calculate the standard deviation of residuals for each industry-month. Industry-year heterogeneity measure equals the average of the monthly standard deviations across all months in each calendar year for each industry. Industry-year homogeneity (INDUSTRY_WIDE) is equal to the negative of the heterogeneity industry-year measure. %CRSP: Stock price incentives to collude For each FF49 industry-year, we count the firms with greater than 200 daily return observations on CRSP and divide by the number of firms in the Compustat fundamentals annual file with an SIC code, sales, and total assets data. VOLATILITY: Industry uncertainty Using option data from OptionMetrics, we calculate the Van Buskirk (2011) measure of implied volatility for at-the-money options before each I/B/E/S earnings announcement that meets the data requirements. We calculate the firm average implied volatility by year to convert this measure to firm-year observations so that firms with more populated data do not influence the measurement. PUBLIC: Availability of public information For each FF49 industry, we estimate a factor model (A1) using monthly data from January 1994 to December 2008: 𝑟𝑖𝑚 = 𝛼𝑗 + 𝛽𝑗𝑚𝑘𝑡𝑟𝑚𝑘𝑡,𝑚 + 𝛽𝑗𝑖𝑛𝑑𝑟𝑖𝑛𝑑,𝑚 + ∑ 𝛿𝑗𝑧7

𝑧=1 𝐹𝐴𝐶𝑇𝑂𝑅𝑚𝑧 + 𝜀𝑖𝑚 (A1)

where imr is the monthly return for firm i in industry j (j = 1 to 49); 𝑟𝑚𝑘𝑡,𝑚 is the monthly return on the CRSP equally-weighted market index; 𝑟𝑖𝑛𝑑,𝑚 is the monthly return on the equally-weighted industry index; and 𝐹𝐴𝐶𝑇𝑂𝑅𝑚𝑧 is the

51

monthly value of one of seven macro-economic risk factors identified in prior research and described below. PUBLIC is the adjusted R2 value of the factor model. A greater adjusted R2 for an industry suggests that more information about the industry is available through public signals. The seven factors in Equation (A1) are: (1) Short-term interest rates = the 3-month treasury bill rate from CRSP. (2) Default premium = Moody's Seasoned Baa Corporate Bond Yield (available from

http://www.federalreserve.gov/releases/h15/data.htm) minus the 10-year constant maturity government bond yield (from http://research.stlouisfed.org/fred2/).

(3) Term premium = the yield on 10-year constant maturity government bonds minus the yield on one-year constant maturity government bonds (from http://research.stlouisfed.org/fred2/).

(4) Foreign Exchange Rates = weighted average of the foreign exchange value of the U.S. dollar against a subset of the broad index currencies that circulate widely outside the country of issue (from http://research.stlouisfed.org/fred2/).

(5) Producer price index = the total finished goods producer price index from the Bureau of Labor Statistics. (6) Small minus Big (SMB) = the Fama-French monthly benchmark factor for the performance of small stocks relative to

big stocks. (7) High – low book-to-market (HML) = the Fama-French monthly benchmark factor for the performance of value

stocks relative to growth stocks.

SMB and HML are from Kenneth French’s website: (http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html).

LIT_RISK: Litigation risk Our proxy for litigation risk is based on Model 3 from Table 7 of Kim and Skinner (KS, 2012). This model is intended to capture the factors that make a firm more vulnerable to securities litigation prior to the revelation of “triggering events” like major price declines. In this sense, the model provides an ex ante probability of litigation risk. Specifically, we estimate the following model:

𝑆𝑈𝐸𝐷 = 𝑏0 + 𝑏1(𝐹𝑃𝑆𝑡) + 𝑏2(𝐿𝑁𝐴𝑆𝑆𝐸𝑇𝑆𝑡−1) + 𝑏3(𝑆𝐴𝐿𝐸𝑆_𝐺𝑅𝑂𝑊𝑇𝐻𝑡−1) + 𝑏4(𝑅𝐸𝑇𝑈𝑅𝑁𝑡−1)+ 𝑏5(𝑅𝐸𝑇𝑈𝑅𝑁_𝑆𝐾𝐸𝑊𝑁𝐸𝑆𝑆𝑡−1) + 𝑏6(𝑅𝐸𝑇𝑈𝑅𝑁_𝑆𝑇𝐷_𝐷𝐸𝑉𝑡−1) + 𝑏7(𝑇𝑈𝑅𝑁𝑂𝑉𝐸𝑅𝑡−1) + 𝑒

We estimate LIT_RISK for fiscal years between 1996 and 2008 for a sample of 48,844 firm-years (including 2,667 lawsuits). The firm-year observations are averaged to create the industry-year observations. Our sample of sued firms is generated from a database provided by Woodruff-Sawyer, a San Francisco-based insurance brokerage firm. As explained in Rogers and Van Buskirk (2009), Woodruff-Sawyer aggregates data from a number of sources including Securities Class Action Clearinghouse (SCAC), which is the source of the sued firm sample in KS. Estimating the model on this sample allows us to relax some of KS’s sample requirements that are not necessary for our study (e.g., they require return data to be available for three years prior to the beginning of each fiscal year). As expected, the sign and significance of each of the independent variables is similar to that reported by KS (see KS for specific variable definitions). HERF: Industry concentration We compute a revenue-based herfindahl index for each FF49 industry for each year as:

𝐻𝐸𝑅𝐹𝑗𝑦 = ���𝑆𝑎𝑙𝑒𝑠𝑖 �𝑆𝑎𝑙𝑒𝑠𝑖

𝑚

𝑖=1

� �2𝑚

𝑖=1

�

where Salesi is revenues for each firm i in industry j in year y and m equals 50 if there are at least 50 firms in the industry and equals the number of firms in industry j in year y if the number is less than 50. We similarly compute a herfindahl index based on market share of total assets. We also compute industry-year level concentration ratios as the sum of the market shares of sales or total assets for the n largest firms in an industry: 𝐶𝑛 = ∑ (𝑀𝑎𝑟𝑘𝑒𝑡 𝑆ℎ𝑎𝑟𝑒)𝑖𝑛

𝑖=1 We compute the concentration ratio for the top four, six, and eight firms. Concentration ratios are commonly used in the cartel literature, but there is no consensus on the appropriate n to include.

52

Appendix B: Summary of explanatory variables across the Fama-French 49 industries

An asterisk after the industry name indicates that the industry is designated as having high tail risk (TAILRISK=1). The remaining columns report the average annual number of firms in each FF49 industry over the period 1994-2008 with available data on Compustat and CRSP, the average annual percent of firms on CRSP relative to Compustat (%CRSP), the PUBLIC score, and the average industry-year HERF.

Average annual # of firms Industry Compustat CRSP %CRSP PUBLIC HERF

1 Agric Agriculture 18.4 16.7 92.0% 0.128 0.232 2 Food Food Products 89.5 81.8 92.9% 0.064 0.065 3 Soda Candy & Soda 11.1 20.1 191.9% 0.135 0.306 4 Beer Beer & Liquor 12.5 23.7 198.1% 0.096 0.329 5 Smoke Tobacco Products 5.9 8.2 143.3% 0.203 0.822 6 Toys Recreation * 53.3 52.7 106.2% 0.103 0.156 7 Fun Entertainment * 102.3 87.5 85.2% 0.090 0.143 8 Books Printing and Publishing 45.9 58.5 129.2% 0.102 0.064 9 Hshld Consumer Goods * 111.5 95.7 85.7% 0.089 0.127

10 Clths Apparel * 79.1 62.4 80.0% 0.107 0.070 11 Hlth Healthcare 112.3 119.5 105.6% 0.080 0.089 12 Medeq Medical Equipment 186.9 187.4 104.1% 0.103 0.074 13 Drugs Pharmaceutical Products * 316.7 324.1 104.1% 0.160 0.092 14 Chems Chemicals 92.0 103.3 113.4% 0.103 0.059 15 Rubbr Rubber and Plastic Products 57.8 43.8 73.3% 0.101 0.070 16 Txtls Textiles 32.8 29.3 91.6% 0.122 0.155 17 Bldmt Construction Materials 103.5 88.3 84.2% 0.087 0.071 18 Constr Construction * 69.8 70.9 103.8% 0.100 0.061 19 Steel Steel Works Etc. 75.6 82.2 112.4% 0.185 0.068 20 Fabpr Fabricated Products 22.5 19.7 103.4% 0.131 0.141 21 Mach Machinery 183.8 177.3 98.2% 0.126 0.055 22 Elceq Electrical Equipment * 79.7 135.8 166.7% 0.133 0.230 23 Autos Automobiles and Trucks * 79.9 90.1 113.4% 0.166 0.220 24 Aero Aircraft 22.2 26.3 119.2% 0.139 0.256 25 Ships Shipbuilding, Railroad Equip. 10.7 8.5 81.8% 0.245 0.425 26 Guns Defense * 9.8 9.5 96.4% 0.227 0.747 27 Gold Precious Metals 21.3 64.5 308.7% 0.246 0.501 28 Mines Non-Metallic/Industrial Metal Mining 18.2 29.1 163.5% 0.184 0.192 29 Coal Coal * 7.1 12.2 174.5% 0.361 0.251 30 Oil Petroleum and Natural Gas 198.9 242.7 124.6% 0.194 0.121 31 Util Utilities 223.8 170.5 78.6% 0.133 0.026 32 Telcm Communication * 193.0 238.4 133.6% 0.161 0.067 33 Persv Personal Services * 68.8 68.1 100.6% 0.086 0.066 34 Bussv Business Services 334.3 387.0 122.9% 0.109 0.037 35 Hardw Computers 163.1 137.5 87.3% 0.171 0.136 36 Softw Computer Software 524.1 446.7 87.8% 0.186 0.150 37 Chips Electronic Equipment 337.7 350.0 106.5% 0.213 0.066 38 Labeq Measuring & Control Equip. 123.9 122.4 100.2% 0.124 0.093 39 Paper Business Supplies 65.7 67.1 106.9% 0.113 0.074 40 Boxes Shipping Containers 16.0 17.5 111.3% 0.148 0.140 41 Trans Transportation * 139.2 152.3 113.4% 0.110 0.045 42 Whlsl Wholesale 234.3 250.7 109.6% 0.078 0.057 43 Rtail Retail * 312.9 287.7 93.8% 0.103 0.066 44 Meals Restaurants, Hotels, Motels * 115.2 131.3 116.7% 0.112 0.075 45 Banks Banking 337.1 723.0 246.2% 0.122 0.056 46 Insur Insurance * 207.0 200.3 99.7% 0.108 0.045 47 Rlest Real Estate 56.3 49.5 90.3% 0.103 0.146 48 Fin Trading * 315.5 1085.2 364.6% 0.138 0.110 49 Other Almost Nothing 79.0 25.5 31.5% 0.097 0.261