UNCOMMON INFORMATION IN FIRM DISCLOSURES
By
Matthew David DeAngelis

A DISSERTATION
Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of
Business Administration – Doctor of Philosophy
2014

ABSTRACT
UNCOMMON INFORMATION IN FIRM DISCLOSURES
By
Matthew David DeAngelis
This study investigates uncommon information in firm disclosures. Using a method from computational linguistics called Latent Semantic Analysis (LSA), I measure the difference in topics
discussed in the Management’s Discussion and Analysis (MD&A) section of a firm’s 10-K filing
relative to its peers. I hypothesize that the presence of uncommon information signals to investors
that the firm’s valuation function differs from its peers’. Consistent with this hypothesis, I find
that uncommon information is associated with higher firm-specific stock returns. I also hypothesize that investors will process uncommon information slowly because uncommon information is
textually and conceptually complex and investors have little data with which to estimate its value.
Consistent with this hypothesis, I find that the market response to uncommon information is delayed. Drawing on the information theory and linguistics literature, I further hypothesize and find
that text containing uncommon information is longer and less readable, suggesting that readability
is at least partially a function of the information presented in the text. This study makes several
contributions to the literature. First, I identify a mechanism through which investors identify firmspecific components of value. Second, I link the information content of a disclosure to higher
information processing costs. In doing so, I identify a cost-benefit tradeoff faced by managers
when crafting their disclosures and by investors when they analyze those disclosures. Third, I
introduce Latent Semantic Analysis to the accounting literature as a method for comparing firm
disclosures.

Copyright by
MATTHEW DAVID DEANGELIS
2014

ACKNOWLEDGEMENTS

I am grateful to my dissertation committee, Marilyn Johnson, Brian Pentland, Alan Munn, and
especially Kathy Petroni, my chair, for their unfailing support and guidance. I also gratefully acknowledge the helpful comments of Andrew Acito, Ramji Balakrishnan, Colleen Boland, Tony
Bucaro, Ted Christensen, Elizabeth Connors, Richard Crowley, Paul Demere, Nick Dopuch, Susanna Gallani, Chris Hogan, Andy Imdieke, Raffi Indjejikian, John Jiang, Rong Jin, Ranjani Krishnan, Dan Lynch, Xiumin Martin, Miles Romney, Joe Schroeder, Karen Sedatole, Mike Shields,
Amy Swaney, Inna Voytsekhivska, Isabel Wang, Philip Wang, Dan Wangerin, participants at my
brownbag seminar and dissertation proposal at Michigan State, doctoral students at the University of Illinois and workshop participants at Washington University in St. Louis, Georgia State
University, the University of Illinois and the University of Michigan.

iv

TABLE OF CONTENTS

LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

vi

CHAPTER 1

1

INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

CHAPTER 2 LITERATURE REVIEW AND HYPOTHESIS DEVELOPMENT
2.1 Uncommon Information and Firm Valuation . . . . . . . . . . . . . . . . . .
2.1.1 Hypothesis 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.2 Hypothesis 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Difficult-to-Estimate Information . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Difficult-to-Process Information . . . . . . . . . . . . . . . . . . . . . . . .
2.3.1 Hypothesis 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.2 Hypothesis 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

. 5
. 5
. 7
. 8
. 9
. 11
. 13
. 13

CHAPTER 3 MEASURING UNCOMMON INFORMATION
3.1 Measuring the Characteristics of Language in the MD&A .
3.2 Calculating Uncommon . . . . . . . . . . . . . . . . . . .
3.3 Examples of Uncommon Information . . . . . . . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

15
16
19
22

CHAPTER 4 TESTS OF HYPOTHESES AND EMPIRICAL RESULTS
4.1 Uncommon Information and Firm-Specific Returns . . . . . . . . . .
4.2 Uncommon Information and Information Processing Costs . . . . . .
4.2.1 Uncommon Information and Delayed Market Response . . . .
4.2.2 Uncommon Information and Readability . . . . . . . . . . . .
4.3 Summary of Results . . . . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

24
24
33
33
39
44

CHAPTER 5

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

FUTURE WORK AND CONCLUSION . . . . . . . . . . . . . . . . . . 46

APPENDICES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
A
Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
B
Excerpts from Most and Least Common MD&As . . . . . . . . . . . . . . . . . . 55
BIBLIOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

v

LIST OF TABLES

Table 3.1

Sample Selection Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

Table 3.2

Top 10 Words for the Top 5 Dimensions in the Software Industry, Fiscal Year 2007 22

Table 3.3

Top 10 Words for the Most and Least Common MD&As in the Software Industry, Fiscal Year 2007 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

Table 4.1

Descriptive Statistics for Determinants Analysis . . . . . . . . . . . . . . . . . . 26

Table 4.2

Correlations between Determinants . . . . . . . . . . . . . . . . . . . . . . . . 27

Table 4.3

Regression of Uncommon on Determinants . . . . . . . . . . . . . . . . . . . . 29

Table 4.4

Descriptive Statistics for ReturnSync Regressions . . . . . . . . . . . . . . . . . 31

Table 4.5

Correlations between ReturnSync Variables . . . . . . . . . . . . . . . . . . . . 32

Table 4.6

Regression of ∆ReturnSync on Uncommon . . . . . . . . . . . . . . . . . . . . 34

Table 4.7

Descriptive Statistics for CAR Regressions . . . . . . . . . . . . . . . . . . . . 36

Table 4.8

Correlations between CAR Variables . . . . . . . . . . . . . . . . . . . . . . . 37

Table 4.9

Regression of CAR and PCAR on Uncommon . . . . . . . . . . . . . . . . . . . 40

Table 4.10 Descriptive Statistics for Readability Regressions . . . . . . . . . . . . . . . . . 42
Table 4.11 Correlations between Readability Variables . . . . . . . . . . . . . . . . . . . . 43
Table 4.12 Regression of Readability Variables on Uncommon . . . . . . . . . . . . . . . . 45
Table A.1 Variable List . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

vi

CHAPTER 1
INTRODUCTION

In this study, I examine uncommon information in firm disclosures. Uncommon information is
present in a firm’s disclosure if the disclosure discusses or emphasizes different valuation parameters, or different properties of shared parameters, than are discussed in the disclosures of the firm’s
peers. For instance, a growth firm may devote substantial space to a discussion of sales and little to
a discussion of earnings, whereas the firm’s peers predominantly discuss earnings. Or, a firm operating in a geographic market with few competitors may provide information about the relationship
between sales and earnings in that market. I predict that the presence of uncommon information in
a disclosure signals to investors that the firm’s valuation function differs from its peers’. As such,
uncommon information is useful to investors both in choosing the parameters of their valuation
models and in populating those models.
Although uncommon information is useful, prior research on investors’ categorical and contextual thinking from accounting and finance (Lee et al., 2006; Hopkins, 1996; Peng & Xiong, 2006)
and psychology and linguistics (Zipf, 1949; Becker, 1980; Pierce, 1980; Goodman et al., 1981;
Halgren et al., 2002) suggests that uncommon information may be difficult for investors to process. First, investors may have difficulty judging the value relevance and precision of uncommon
information due to a lack of data from comparable firms. Second, the collection and processing
of uncommon information may be impaired by increased cognitive and textual complexity of this
information. Since the market processes information less completely when information processing
costs are higher (Bloomfield, 2002; Li, 2008; Lee, 2012), I predict that the market response to uncommon information is delayed. As such, while uncommon information is useful for identifying
firm-specific components of value, the market incorporates that information slowly compared with
common information.
I use textual analysis to identify uncommon information in the Management’s Discussion and

1

Analysis (MD&A) section of the 10-K filing. I analyze the MD&A for several reasons. First, prior
research has found that the MD&A contains decision-relevant information for investors (Barron
et al., 1999; Clarkson et al., 1999; Brown & Tucker, 2011; Davis & Tama-Sweet, 2012). Second,
identifying uncommon information requires a significant amount of cross-sectional data. Since the
MD&A is a mandatory disclosure, it enables me to observe the information disclosed by a large
number of firms. Third, the MD&A is specifically designed to enable “investors to see through
the eyes of management" and to “provide the context within which financial information should
be analyzed” (Securities and Exchange Commission, 2003). The MD&A is therefore a rich setting
for comparing the information content of disclosures in a cross-section of firms.
In comparing MD&As across firms, it is useful to distinguish the words that managers use in
their discussion from the information conveyed by those words.1 If a firm manager spends 90%
of the MD&A discussing “earnings", while the manager of another firm talks about “profits", a
fair measure of information content would consider these MD&As to be very similar. In order to
“see through" these word choices, I use a method from computational linguistics, Latent Semantic
Analysis (LSA), to identify the topics that managers are discussing when they use particular words.
Related to factor analysis, LSA examines co-occurrences between words to determine whether
words are likely to be related to one another. I use LSA to group related words into topics and
measure the relative differences in topics discussed between a firm’s MD&A and the MD&As of
its peer firms to identify MD&As that contain uncommon information.
I measure uncommon information across a sample of 35,585 MD&As across 7,984 firms from
1994 to 2012 and examine the relationship between uncommon information and stock returns.
Specifically, I investigate whether firms with MD&As that contain uncommon information have
higher firm-specific stock returns . I hypothesize that uncommon information is related to firmspecific returns in two ways. First, the firm may possess unusual characteristics or experience
unusual events that result in both higher firm-specific returns and a higher level of disclosed uncommon information. Second, the disclosure of uncommon information provides incremental informa1 Bloomfield

(2002) distinguishes between “data”, or marks on a page, and “statistics”, the
information gleaned from processing data.
2

tion to investors about firm value. To test the first hypothesis, I regress uncommon information on
measures of firm characteristics and returns and find that firm-specific returns and characteristics
in the year prior to the disclosure are significant determinants of uncommon information. I also
find that firms are more likely to provide uncommon information when the differences between the
firm and its peers are not reflected in earnings, consistent with managers disclosing uncommon information as a supplement to earnings information. To test whether uncommon information in the
MD&A is incrementally informative, I regress the change in return synchronicity, or the change
in the portion of a firm’s returns explained by market and industry returns (Durnev et al., 2003;
Piotroski & Roulstone, 2004), on uncommon information. Consistent with uncommon information
in the MD&A providing incremental information, I find that uncommon information is associated
with an increase in firm-specific returns in the year following the disclosure.
I further hypothesize that the high information processing costs of uncommon information lead
to an incomplete market response. In order to investigate the timing of the market response to uncommon information, I regress the absolute value of market returns prior to, around, and following
the disclosure on uncommon information. Consistent with a delayed market response to uncommon information in the MD&A, I find a significant positive association between uncommon information and market returns in the two months following the disclosure, but no association between
uncommon information and stock returns in the three-day window around the disclosure. Inconsistent with the market responding to information disclosed prior to the MD&A filing, I find no
significant relationship between uncommon information and stock returns prior to the disclosure.
These results suggest that the market responds to uncommon information disclosed in the MD&A,
but that this information is difficult to process. Also consistent with high processing costs, I find
that text containing uncommon information is less readable than text containing common information. This finding supports my market tests and provides insight into sources of textual complexity
in firm disclosures.
This study contributes to the literature in several ways. First, while prior research has shown
that the inclusion of firm-specific information improves the informativeness of stock prices, lit-

3

tle is known about how investors learn about and incorporate this information into their decision
models. My study suggests that uncommon information helps investors identify variables in their
decision models that they may not have otherwise considered. Second, prior studies on the information processing costs of financial reporting focus on the structural elements of disclosure, such
as readability (Li, 2008; Lee, 2012) and report presentation (Lee et al., 2006; Hodder et al., 2008;
Bloomfield et al., 2010), and largely attribute these costs to managerial opportunism. In contrast,
this study is the first to my knowledge to directly link the information content of a disclosure to the
associated processing costs. As such, my study suggests that investors face a cost-benefit tradeoff
in choosing to utilize uncommon information. These tradeoffs may also have implications for managers’ reporting strategies. Third, this study contributes to our understanding of how information
processing constraints impact investors and market efficiency. Whereas prior research has shown
that limited attention negatively impacts market efficiency (Hirshleifer et al., 2009; Dellavigna &
Pollet, 2009; Aboody et al., 2010), my results suggest that investors may attend to information and
still process it slowly. My results may also help to shed light on why investors revisit previously
disclosed information in making their trading decisions (Tetlock, 2011; Drake et al., 2012).
The remainder of this study proceeds as follows. Section 2 reviews literature relevant to my
research questions and develops my hypotheses. Section 3 discusses the calculation of my measure
of uncommon information. In Section 4, I discuss the results of my study and in Section 5, I discuss
future work and conclude.

4

CHAPTER 2
LITERATURE REVIEW AND HYPOTHESIS DEVELOPMENT

In the next three Sections, I review literature relevant to examining the role of uncommon information in financial reporting and develop my hypotheses.

2.1

Uncommon Information and Firm Valuation

Ohlson (1995) articulates a model of firm valuation that incorporates accounting information. In
this model, a firm’s value, Pt , is expressed as a function of book value, yt , and expected future
residual income, E [x˜t+τ ]. Specifically:
∞

Pt = yt +

∑ R−τ
f E [x˜t+τ ]

(2.1)

τ=1

where R f represents the risk-free rate. Since book value is (at least in theory) readily available, an investor’s valuation task centers on the second term in this linear equation, the investor’s
expectation of future residual income. Ohlson expresses this expectation as:

E [xt+1 ] = (R f − 1)yt + ωxt + vt

(2.2)

where yt , xt and vt represent book value, residual income and information other than residual income, respectively, at time t. Together these equations suggest that value is a function of
observable accounting information, namely book value and residual income, observable other information, and unobservable expectations of future residual income and other information that
depend on observable information. In Ohlson’s model, the coefficient ω on xt and vt above are assumed to be fixed and known. However, in practice investors must form these expectations based
on the information available to them. As a result, an individual investor faces estimation risk as he
chooses the functional form of his estimation model and collects the relevant data to estimate the
model.
5

While there is some debate in the literature about whether or not estimation risk is priced
(Francis et al., 2005; Core et al., 2008), there is considerable theoretical work suggesting that, if
the market model for a firm must be estimated using incomplete information, then estimation risk
impacts investors’ estimates of firm value and risk (Handa & Linn, 1993; Clarkson et al., 1996).
Lambert et al. (2007) model a firm’s cost of capital under this framework and show that disclosure
unambiguously reduces the cost of capital by reducing investors’ estimates of the covariance of the
firm’s cash flows with the market as a whole. In contrast to prior studies, which view estimation risk
as being primarily resolved over time (as observations of a firm’s returns increase), Lambert et al.
(2007) show analytically that any information about a firm’s future cash flows reduces estimation
risk by reducing investors’ estimates of the firm’s exposure to systematic risk.
In Lambert et al. (2007), an investor estimates this exposure by estimating the covariance between a firm’s cash flows and other firms in the market. This estimate is a function of two variables:
the estimated covariance prior to the disclosure and the precision of the disclosure. In a market of
two firms, this relationship is expressed as:
Cov(V˜j , V˜k |Z j ) = Cov(V˜j , V˜k )

Var(ε˜j )
Var(Z˜j )

(2.3)

where V˜j is the expected future cash flows of firm j, V˜k is the expected future cash flows of firm k,
Z˜j is noisy information about firm j’s cash flow, and ε˜j is the noise in Z˜j . In terms of the model in
Ohlson (1995), an investor’s expectation of the future cash flows of firms j and k, V˜j and V˜k , depend
on the functional form he uses to estimate Equation 2.2, or the variables that he chooses to include
in forming his expectation. In updating these expectations, an investor uses observed financial
reports, Z˜j , to modify and populate his estimation model and update his estimated covariance.
In general, higher quality financial reporting reduces the estimated covariance between firms as
investors incorporate more precise information into their valuation models.
Empirical studies support the idea that better information about a firm leads to a lower covariance of returns between firms. Durnev et al. (2003) find that current returns contain more
information about future earnings when the proportion of a firm’s returns explained by market and
industry returns (called return synchronicity) is lower. Piotroski & Roulstone (2004) find that the
6

number of analyst forecasts are related to higher return synchronicity, indicating that analysts tend
to provide information that applies to all firms in an industry rather than precise firm-specific information. Since analysts are often used as a proxy for sophisticated outside investors (see, for
example, Bamber et al., 2011), this suggests that collecting firm-specific information is often too
costly for outsiders. Public disclosure of this information could reduce the cost of collection and
increase the proportion of investors who are informed (Kim & Verrecchia, 1994).
I predict that the presence of uncommon information in a firm’s MD&A signals to investors
that the firm’s valuation model differs from those of its peers. In particular, it suggests variables for
inclusion that the investor may not otherwise have considered. The resulting revision of investors’
decision models creates differences between the firm’s returns and the returns of peer firms. As
a result, I predict that uncommon information is associated with a lower portion of firm returns
explained by the returns of the market and the firm’s industry.

2.1.1

Hypothesis 1

I examine the relationship between uncommon information and firm-specific returns both prior
to and following the release of the MD&A. MD&A disclosure is endogenously determined by
the economics of the firm, so if the firm experiences unusual economic events or possesses unusual characteristics, those firms will disclose more uncommon information. In addition, because
accounting information tends to lag stock prices (Ball & Shivakumar, 2008), the market should respond to these events and characteristics prior to the disclosure. This leads to my first hypothesis,
stated in alternative form:
Hypothesis 1 Firm-specific returns and characteristics in year t are determinants of uncommon
information disclosed at the end of year t.

7

2.1.2

Hypothesis 2

In addition, the presence of uncommon information in a firm’s disclosure may provide incremental
information to the market. First, prior studies have found that the MD&A is incrementally informative over other disclosures (Davis & Tama-Sweet, 2012; Lee, 2012). Second, the presence
of uncommon information in a firm’s disclosures may signal to investors that the firm’s valuation
function differs from that of other firms. Managers have superior information to outside investors
(Healy & Palepu, 2001) and securities regulations require that managers report relevant and material information in their financial reports (Financial Accounting Standards Board, 2010). As such,
a decision by managers to include uncommon information in a financial report likely increases
investors’ estimates of its importance (Arya & Mittendorf, 2005). This is particularly true for certain sections of the report, such as the MD&A, in which managers are expected to “present their
disclosure so that the most important information is most prominent” (Securities and Exchange
Commission, 2003). The fact that managers of other firms decline to include this information
suggests that this information is relevant only for the reporting firm.
If uncommon information in the MD&A is incrementally informative to the market, then uncommon information should be associated with an increase in the firm-specific portion of returns
following the disclosure. This leads to my second hypothesis, stated in alternative form:
Hypothesis 2 Uncommon information disclosed at the end of year t is associated with an increase
in firm-specific returns between year t and year t + 1.
The model in Lambert et al. (2007) suggests that the valuation effect of a particular piece of
information, Z˜j , is a function of both its value relevance and precision. However, these properties
are not known and must be estimated. In Section 2.2, I discuss characteristics of information
that impose constraints on estimation. In addition, prior literature has found that, all else equal,
investors are less likely to collect and process information when the costs of doing so are high
(Grossman & Stiglitz, 1980; Bloomfield, 2002; Li, 2008; Lee, 2012). In Section 2.3, I discuss
characteristics of information that increase processing costs.

8

2.2

Difficult-to-Estimate Information

As discussed in Section 2.1, an investor faces parameter uncertainty in estimating firm value: he
is uncertain about the relevant parameters to include in his model and their relationship to value.
Hong et al. (2007) argue that, faced with a large amount of information and parameter uncertainty,
an investor chooses a parsimonious model that includes a subset of the most useful information.
This investor behaves like an econometrician, using a theoretical model to predict what information
is most relevant and then testing the model against economic outcomes. He incorrectly excludes
relevant parameters (assuming he is aware of them) under two conditions. First, the investor lacks
a strong theoretical reason to link a parameter with firm value. I discuss this condition in greater
detail in Section 2.3. Second, the investor lacks sufficient data to generate a high-quality estimate
of the parameter (Hong et al., 2007; Branch & Evans, 2006, 2010). Sufficient data can be generated
using a time-series of observations (Barry & Brown, 1985; Coles et al., 1995; Lewellen & Shanken,
2002) or a cross-section (Lambert et al., 2007). Branch & Evans (2010) show analytically that
even fully rational investors (i.e. not subject to cognitive limitations that might constrain model
selection) exclude relevant parameters from their decision models when they possess insufficient
data. When an investor is not confident in the estimates produced by including a particular variable,
he is more likely to exclude that variable from his model.
There is already substantial evidence in the accounting literature that investors incorporate
information less completely when the reliability of that information is difficult to estimate. The
accrual anomaly may be the most well-documented market inefficiency in accounting research.
Sloan (1996) and Richardson et al. (2005) argue that accruals are less persistent than cash flows
because they are more subjectively determined and thus less reliable. However, investors seem to
respond slowly to this lack of reliability. While the accounting literature has generally viewed this
as an investor overreaction to the persistence of earnings, Bloomfield (2002) notes that it is fully
consistent with an underreaction to information in accruals that is difficult to estimate. In other
words, the low reliability of accruals suppresses investor learning about and the market response
to the information contained in them.
9

Lewellen & Shanken (2002) demonstrate analytically that investor learning causes return patterns that are predictable ex post but are not visible to investors during the learning process.
The empirical specification of accruals studies supports this explanation for the accrual anomaly.
Richardson et al. (2005) estimate their model over a period of almost forty years. Xie (2001) shows
that the accrual anomaly is most likely due to abnormal accruals, which are usually calculated over
a sizable cross-section of firms or a time-series. Xie (2001) uses all firms in the same two-digit
SIC code per industry, for instance, while Francis et al. (2005) use the Fama-French 48 industry
groups and require at least 20 firms per industry. Francis et al. (2008) require that a firm have 11
years of data to be included in their model of accruals quality. If it takes over a decade of data or
20 comparable firms to estimate the reliability of accruals, it is unsurprising that investors process
this information slowly.
Experimental research in accounting confirms that individuals’ use of accounting information
depends on a critical mass of comparable data. Lipe & Salterio (2000) find that managers tend
to overweight (underweight) performance measures that are common (unique) across subordinates
when evaluating their performance. This allows the manager to reduce cognitive effort by focusing
only on the measures most useful for comparison (Slovic & MacPhillamy, 1974). Libby et al.
(2004) find that an additional reason that individuals underweight the unique measures is that
they are more uncertain about the quality of those measures. When individuals are provided with
assurance about the quality of the unique measures, or are motivated to expend a greater amount
of cognitive effort, they increase their weights on unique measures. More broadly, experiments in
a variety of settings consistently show that the ability to compare alternatives significantly reduces
the cognitive costs of choosing between them (Heneman, 1986; Zhang & Markman, 2001). This
suggests that a rare item of information is likely to impose significant cognitive costs on investors.
Even in the presence of sufficient data to estimate a parameter, an investor may exclude it
from his decision model if he lacks ex ante justification for its inclusion or the expected costs of
collecting and processing the information exceed the expected benefits. I review these conditions
in the next Section.

10

2.3

Difficult-to-Process Information

Research in judgment and decision-making finds that individuals economize on effort when forming their decision models, often settling for a “good enough” model rather than a complete one
(Simon, 1955). This leads to differential search strategies depending on the decision at hand, with
extra effort made to simplify the decision as the number of possible relevant parameters grows
(Shields, 1980; Karelaia & Hogarth, 2008). Simplification strategies often involve ranking the importance of the available parameters subjectively (i.e., without calculating exact weights) and then
including only those parameters judged to be most important in the decision model (Payne et al.,
1993). Shields (1980, pg. 432) suggests that an individual’s choice of simplification strategy and
the subjective weights placed on parameters depends on the individual’s internal representation of
the decision task, with higher weights placed on parameters that are more closely “grouped” with
other available parameters and the task.
An individual’s internal representation of the world in which he lives, sometimes referred to
as a “semantic net” (Kintsch, 1988) or “concept map” (Sowa, 1984), represents knowledge as a
network of ideas with linkages of varying strength between them. When an individual encounters
a new idea, he creates a new concept in his map and forms linkages to related ideas based on the
context. On each subsequent occasion that the individual encounters the same idea, he notes the
context and augments his concept map to strengthen existing linkages (if they are present) and
adds new linkages if necessary. Concepts themselves are defined in part by these linkages (Sowa,
1984, p. 76). For example, the individual can represent the concept of a “nurse” by noting the
similarities and differences between the nurse’s role to the doctor’s in a medical context. When
the individual retrieves a particular idea, he activates the portion of this concept map in which that
idea resides, and stronger linkages lead to stronger activation of related concepts. If the individual
has to make a decision involving a set of related concepts, the decision is less costly to make if the
relevant concepts are closely linked to one another. If not, he has to search the remainder of his
concept map and form new linkages between previously unlinked concepts.1
1 For

instance, if you see a television commercial advertising a “rug doctor", you have to think
11

In this model, information processing costs occur both in the creation of new linkages and concepts (“learning”) and the retrieval of concepts from the map (“memory”). As such, information
processing costs increase when the individual is presented with unfamiliar concepts or combinations of concepts that are not closely linked within the individual’s concept map. Consistent
with this expectation, experimental research in psychology suggests that unfamiliar or unexpected
words are more costly to process (Becker, 1976; Forster & Chambers, 1973; Glanzer & Ehrenreich, 1979; Forster, 1981). These studies find that words that occur with low probability in English
text are harder for individuals to retain (Postman, 1970; Hulme et al., 1997), learn (Hall, 1954),
and recall from long-term memory (Scarborough et al., 1977; Rayner & Duffy, 1986). In addition, familiar words are more costly to process when they are placed in an unfamiliar context.
Halgren et al. (2002) finds that brain activity associated with language processing increases to a
greater degree when individuals encounter contextually unexpected words in sentences than when
they encounter words with low frequency in English text. Controlling for word frequency, Becker
(1980) finds that individuals recall words more quickly when they are accompanied by related
words (“nurse" and “doctor"), and less quickly when they are accompanied by unrelated words
(“nurse” and “furniture”), than when words are presented with a neutral stimulus such as a nonsense string of characters. Tweedy et al. (1977) adds that this effect is more pronounced when the
number of related words is increased. Context is also important in establishing the meaning of a
word: Schvaneveldt et al. (1976) finds that individuals are able to decide on the meaning of an
ambiguous word more quickly when it is presented with related words. On the whole, it seems
that a decrease in the conditional probability of observing a word in a particular context leads to a
significant increase in information processing.
about what a doctor is (someone who fixes physical problems that people have) and what kind
of physical problems a rug might have (worn out, dirty) to decide that a rug doctor is probably a
carpet cleaner or a person who replaces carpeting.

12

2.3.1

Hypothesis 3

If uncommon information imposes greater data collection and cognitive costs on investors, then
uncommon information should be less completely incorporated into market prices (Bloomfield,
2002; Li, 2008). This leads to my third hypothesis:
Hypothesis 3 Uncommon information is associated with a delayed market response.

2.3.2

Hypothesis 4

Examining uncommon information also provides an opportunity to examine the effect of conceptual complexity on textual complexity. Prior research finds that less readable language is associated with higher processing costs (Baddeley et al., 1975; Lee, 2012; Miller, 2010; Lehavy et al.,
2011; Tan et al., 2013; Rennekamp, 2012a). However, there is little evidence on whether textual
complexity in financial reports is due to the complexity of the information presented (Bloomfield,
2008; Rennekamp, 2012b) or managers’ incentives to obfuscate negative information (Li, 2008).
Information theory suggests that the cost of transmitting and processing a message is dependent
on its information content (Shannon & Weaver, 1949; Pierce, 1980). A key feature of an efficient
communication system is that messages are transmitted over a channel using the smallest number of informative units. As demonstrated by Pierce (1980, p. 94-96), this means that commonly
(uncommonly) used messages will be encoded using shorter (longer) signals. In this way, the system will transmit shorter signals with higher probability than it transmits longer signals and the
average signal length will be minimized. Thus, if we assume that it is costly to both transmit and
receive messages, an efficient communication system reserves more costly signals for less common
messages.
Research in information theory and linguistics suggests that common concepts are also more
likely to be represented by words that are shorter, and thus more readable, than less common
concepts. Zipf (1949) extends this feature to linguistic communication. Using an analogy of an
artisan (language user) selecting her tools (words) from a workbench (vocabulary), Zipf notes that

13

the artisan will minimize her effort by keeping the tools that she uses most closest to her. Since
smaller tools are easier to use, as well as easier to pack tightly so that they reduce the distance to
the next tool in line, Zipf hypothesizes that the artisan will seek to minimize the size of those tools
that she uses most frequently. Zipf then predicts and finds that there will be an inverse relationship
between the lengths of words and the frequencies of their usage (Zipf, 1949, p. 63). This so-called
Zipf’s Law has subsequently been used heavily in computational linguistics, and is also one of the
foundations of the Gunning Fog Index used in Li (2008). This leads to my fourth hypothesis:
Hypothesis 4 Text containing uncommon information is longer and less readable than text containing common information.
The above hypotheses are predicated on identifying information that is perceived as uncommon by the average investor. Although investor perceptions are not directly observable, Landauer
& Dumais (1997) find that Latent Semantic Analysis, the statistical technique used in this study,
identifies concepts and relationships in text that agree with human judgments. Landauer et al.
(1998) provides a comprehensive review of experimental research using LSA, including the use
of computer models to approximate human performance on the Test of English as a Foreign Language. I discuss my use of LSA to measure uncommon information in the next section.2

2I

use LSA to measure uncommon information because of its use in prior research to simulate
human knowledge. Another method, Latent Dirichlet Allocation (Blei et al., 2003), is gaining currency in search engine applications to identify related documents. This method is used in concurrent accounting studies (Ball et al., 2013; Huang et al., 2014) and is worthy of further exploration.
14

CHAPTER 3
MEASURING UNCOMMON INFORMATION

My measure of uncommon information uses text of the Management’s Discussion and Analysis
(MD&A) section of the annual report to determine differences between a firm’s MD&A and its
industry peers. I use a textual measure for its power in discriminating between the meanings of
messages: prior literature has shown that text has power to convey the tone (Loughran & McDonald, 2011b), uncertainty (Loughran & McDonald, 2013), riskiness Li (2006) and time interval Li
(2010). Since numerical financial information is continuous in nature, it is difficult to formulate
an interpretation of a particular realization. What constitutes a material change in a line item,
and thus a distinct message? Even if a change is certainly material, the meaning of that change
is highly contingent on the movements of other line items. Text, on the other hand, is discrete:
although theoretically language choices are infinite, in practice word choice is highly constrained
by context. Zipf (1949) finds that you need only 1,000 unique words to characterize most of the
linguistic variation in a work as subtle as James Joyce’s Ulysses; the average MD&A should be far
more limited in its word choice. In addition, textual analysis provides tools to map the words to
particular topics, thus providing a fair representation of the meaning of the text itself. Using this
representation also allows for quantification of what is omitted: to the extent that what managers
do not say is informative (Bloomfield, 2012), my measure can capture those choices as well. I
focus on the MD&A because prior research shows that the MD&A is useful for users’ decisionmaking (Barron et al., 1999; Clarkson et al., 1999; Brown & Tucker, 2011; Davis & Tama-Sweet,
2012). In addition, the MD&A provides a useful setting for quantifying the relative probabilities of
observing particular messages. The MD&A is a mandatory component of the 10-K filing, and thus
the decision to have an MD&A is not voluntary. However, although the SEC has provided some
guidance on how the MD&A should be used, there remains considerable managerial discretion in
what to discuss (Brown & Tucker, 2011). As a result, the MD&A provides a reasonably complete

15

characterization of what manager’s discuss while also providing sufficient variation in disclosure
choice.

3.1

Measuring the Characteristics of Language in the MD&A

The first paper to measure differences in language across MD&As was Brown & Tucker (2011)’s
measure of year-over-year MD&A modifications. Brown & Tucker (2011) use a vector space
model to represent the text of the MD&A as a vector in n-dimensional space, where n represents the
number of unique words in the document and each entry in the vector represents the frequency of
that word in the MD&A. Each entry is then weighted according to its “inverse document frequency"
(IDF), or the number of documents in which it occurs; words that occur less frequently across all
documents are more useful in discriminating between documents. IDF weighting is applied by
M
, where M represents the total number of documents and m
multiplying the entry by log
m
is the number of documents in which that word appears. They then measure the cosine similarity
between a firm’s MD&A and the same firm’s MD&A in the prior year. Cosine similarity represents
the distance between the two MD&As in this n-dimensional space, and is calculated as:
v1 · v2
||v1 || ||v2 ||

(3.1)

where v1 and v2 are the two document vectors, · is the dot product operator, and ||v1 || is the
vector norm of v1 . Since document-term vectors are in positive space, cosine similarity is bounded
between 0 and 1.
Brown & Tucker (2011) find that MD&As are quite sticky from year to year, with a mean
(median) cosine similarity of 84% (89%) (Brown, 2013). However, they find that modifications,
when they occur, are significantly associated with economic changes in the firm and with the
market response to the 10-K filing, suggesting that modifications are informative about changes in
firm value. Contrary to prior research (Barron et al., 1999; Clarkson et al., 1999), however, they
find that analysts do not use changes in the MD&A in their earnings forecasts. Brown and Tucker

16

present evidence that MD&A modifications are more associated with changes in liquidity and
capital resources than changes in operations, and suggest that these changes may be uninformative
for annual earnings forecasts.
While Brown & Tucker (2011)’s measure works well for year-over-year changes within the
same firm, it is not well suited for measuring differences between firms. MD&As within firms
tend to be sticky, not just in terms of the information they convey, but also the words they use to
convey it. Since most of the MD&A may be copied from year to year, observing new words is
very likely to indicate that new information is being conveyed. Across firms, however, managers
may use different words to express the same meaning (known as synonymy in linguistics) or the
same word to express different meanings (polysemy). This tendency can cause similar MD&As
to look different and different MD&As to look similar, in terms of raw word frequencies. The
profusion of word choice across firms also results in a technical problem, common to data analysis,
called the “curse of dimensionality" (Bellman, 2003). Measures of similarity tend to shrink as the
number of dimensions in the measured space gets large: in essence, there are so many directions
in which documents can be pushed that they all end up far away from one another. This results
in little differentiation between similar and different documents. In computational linguistics, it is
common to reduce the size of the space from one of words to one of meanings: from word space
to semantic space, or topic space. I use Latent Semantic Analysis (LSA) for this purpose.
LSA is a form of factor analysis that identifies sources of variation in text. A purely mathematical technique, LSA infers the relationships between words by their co-occurrences: if two words
occur in documents together often, they are likely to be related. Like other vector space models,
LSA disregards the effects of word order or syntax. However, like factor analysis, LSA is able to
uncover complex relationships from simple co-occurrences.
As an example of how LSA identifies relationships, take a collection of documents (corpus)
that contains only two topics: guns and butter. LSA would examine documents in the corpus
and identify the two main sources of variation, the topics guns and butter, and identify words that
are most highly related to each: perhaps “shoot", “aim", “fire", “metal", and “barrel" for guns

17

and “dairy", “milk", “melt", “churn" and “fattening" for butter. Since LSA is evaluating the cooccurrences for all words simultaneously, however, it also forms indirect linkages between words
that do not co-occur but share co-occurrences with other words. For instance, if the word “heat"
commonly co-occurs with the words “barrel" and “melt", “heat" will also be linked to guns and
butter, even if the word “heat" does not directly co-occur with the word “gun" or “butter" in any
of the documents. In this way, LSA constructs a representation of linkages between words that is
akin to the “concept map" discussed in Section 2.3.
LSA accomplishes this complex representation using Singular Value Decomposition (SVD),
which is a form of eigendecomposition. Provided with a matrix of documents (often represented
in rows) and counts of the words in those documents (represented in columns), SVD projects word
co-occurrences into a high-dimensional space and determines the dimensions that best represent
that space. Each word and document is then expressed as a coordinate within this high-dimensional
space, and a distance measure (usually cosine similarity, as described in Equation 3.1 above) can
be used to measure the strength of the relationship between particular words or documents. SVD
ranks the dimensions in the order of the variation they explain, so that the first singular value
corresponds to the dimension that explains the highest degree of variation. Each singular value is
accompanied by a left and right singular vector that represents the dimension in row and column
space, respectively. The original data is then projected on a subset of the singular vectors to obtain
a representation of the data in the new semantic space, and distances can be calculated between
words and documents in that space.
The right singular vectors (representing column space or word space) represent topics in the
document collection. A researcher can determine the meaning of each topic by examining the
words that have the highest association with the given topic in making his determination. Note,
however, that assigning meaning to the topics is not necessary for determining similarities between
documents and words in the collection.

18

3.2

Calculating Uncommon

I construct my sample of MD&As as follows. I download all 10-Ks posted on EDGAR between
1994 and 2012.1 There were 137,516 10-Ks posted during this period. A little over 2,000 of these
are duplicate 10-Ks filed across multiple subsidiaries of the same entity; eliminating these reduces
the sample to 135,384. I extract the Management’s Discussion and Analysis section from each 10K using a Perl procedure. My extraction procedure is not perfect: due to formatting inconsistencies
and the decision by some firms to include the MD&A by reference, I successfully identify and
extract about 62% of the MD&As from this time period, for a sample of 83,457. I remove all
10-Ks for firms that do not occur in Compustat and group the 10-Ks by industry, eliminating 10Ks for industry-years with fewer than 20 firms. This gives a sample on which Uncommon can be
calculated of 35,385.
Table 3.1 shows descriptive statistics for annual data available in Compustat from 1994 to 2012
(Panel A) and for my final sample (Panel B). The differences between my sample and Compustat
are statistically significant for most variables but few are economically significant. Unsurprisingly
given my criterion that firms have at least 20 firms in their industry in a given year, the number of
firms per industry and industry concentration are higher in my sample than in Compustat. In addition, MD&As are longer for firms in my final sample, which may be due in part to a slightly higher
average number of segments (Li, 2008). Since length tends to increase mathematical measures of
similarity between documents (Brown & Tucker, 2011), increased length in my final sample may
make it more difficult for me to identify firms disclosing uncommon information. My sample also
has a similar range and standard deviation for beta and return synchronicity, suggesting that I have
sufficient variation to test my hypotheses. Overall, it does not appear that my study suffers from
selection bias.
I construct my measure of uncommon information, Uncommon, using the following procedure.
First, I remove all HTML and section headers from the text, in order to capture only the narrative
1 Due

to an oversight, my sample does not contain filings 10-K405, a special designation prior
to 2003 indicating that the company did not file timely insider trading information.

19

Table 3.1: Sample Selection Comparison
Panel A: Compustat Sample
Statistic
Beta
ReturnSync
EarnSync
Age
Reg
Size
BT M
Leverage
NBUSSEG
NGEOSEG
RevConc
IndConc
ReturnVol
EarnVol
AnaFollow
Instown
Length
NumFirms

N

Mean

St. Dev.

Min

Median

Max

36,892
99,809
70,477
99,654
115,586
112,814
112,339
114,410
85,093
85,093
98,961
114,387
105,169
72,762
65,043
41,452
53,262
98,576

1.064
−2.194
−2.323
13.137
0.045
1,615.361
0.660
0.201
3.948
4.620
0.315
0.043
0.138
0.059
8.056
0.433
6,170.301
109.955

0.712
1.542
2.382
11.987
0.208
4,154.987
0.530
0.189
3.813
5.047
6.483
0.076
0.084
0.062
7.739
0.289
4,493.175
74.970

−0.240
−14.438
−27.875
1
0
1.361
0.0003
0.000
1
1
0.020
0.001
0.029
0.001
1
0.001
200
20

0.950
−2.053
−1.925
9
0
219.505
0.532
0.159
3
3
0.232
0.005
0.116
0.038
5
0.405
5,175
93

3.620
2.103
5.701
67
1
35,718.200
3.862
0.777
33
79
1,703.142
1.000
0.521
0.349
68
1.000
171,467
326

St. Dev.

Min

Median

Max

0.141
0.728
1.504
2.373
11.337
0.195
3,386.329
0.549
0.195
3.701
5.029
9.734
0.088
0.086
0.065
7.630
0.290
4,491.641
75.603

−0.998
−0.240
−14.438
−21.043
1
0
1.364
0.0004
0.000
1
1
0.030
0.025
0.029
0.001
1
0.001
229
20

−0.787
0.981
−1.820
−1.929
10
0
254.317
0.530
0.145
3
3
0.167
0.107
0.124
0.042
6
0.434
5,946
102

Panel B: Uncommon Information Sample
Statistic
Uncommon
Beta
ReturnSync
EarnSync
Age
Reg
Size
BT M
Leverage
NBUSSEG
NGEOSEG
RevConc
IndConc
ReturnVol
EarnVol
AnaFollow
Instown
Length
NumFirms
Note:

N

Mean

35,385
24,524
31,827
24,646
31,824
35,385
35,020
34,508
34,969
25,879
25,879
30,840
35,385
32,616
23,589
23,209
27,897
35,385
35,373

−0.775
1.095
−2.006
−2.333
12.941
0.040
1,313.116
0.669
0.196
4.298
5.128
0.315
0.128
0.145
0.064
8.284
0.454
6,818.828
117.355

***
***
***
***
***
***
***
***
***
***
***
***
***
***
***
***

***
***
***
***
***
***
***
***
***
***
***
***
***
***
***
***

−0.289
3.620
2.003
5.509
67
1
35,503.660
3.861
0.777
30
73
1,703.142
1.000
0.521
0.349
68
1.000
70,950
326

∗∗∗ indicates that mean or median in the Uncommon sample is significantly
different from the Compustat sample at less than the 1% level.
All variables defined in Table A.1

20

portion of the MD&A. Second, I remove all stopwords (the, and, of, etc.) and stem the remaining
words to their roots.2 Third, to reduce error from typos and proper nouns, I remove all word stems
that occur in fewer than 1% of the MD&As in each Fama-French industry and in the MD&A of
only one CIK code per industry, for all years. Fourth, I construct a document-term matrix for each
industry-year. The document-term matrix contains the number of times that each word appears in
each MD&A. Following Brown & Tucker (2011), I apply a normalized TF-IDF weighting to each
entry. I modify this weighting in two ways. First, as recommended by Loughran & McDonald
(2011b), I scale the raw count of each word stem in each MD&A by the total number of words in
that MD&A. Second, I multiply the result by the inverse document frequency of that word stem
within each industry-year instead of the full sample, to be consistent with calculating distances on
an industry-year basis. Formally, my normalized TF-IDF weighting is calculated as:

(WordCount/Length) ∗ log

M
m

(3.2)

where WordCount is the number of times a particular word stem occurs in a particular MD&A,
Length is the total word count in that MD&A, M is the total number of MD&As in the same
industry-year and m is the number of MD&As in the same industry-year in which this particular
word stem occurs. I also center and scale each column to have a mean of zero and a variance
of one to normalize the data. Third, I perform SVD on the matrix. Like factor analysis, SVD
produces as many dimensions in the projected space as there are in the original data. In order to
obtain meaningful distances, it is necessary to discard dimensions that have low explanatory power.
However, there is no clear rule of thumb for selecting these dimensions. I find that distances
1 of the size of the sample
are well-behaved when the dimensions in column space are about 10

to be distanced, which provides enough observations per dimension to be meaningful while still
capturing a substantial portion of the variation in the dimensions. As such, I select a number of
dimensions equal to 10% of the number of firms in the industry-year. Prior studies have shown
2 Stemming

removes the suffix of words in order to group related words by their root. For
example, “represent", “represented", and “representing" are all reduced to “repres".

21

that, since there is considerable variation within industries, comparisons are best made between a
firm and a few comparable firms instead of the entire industry (Bhojraj & Lee, 2002; DeFranco
et al., 2011; Gong et al., 2013). As such, I calculate the average cosine similarity between each
MD&A and the 4 most similar MD&As in the same industry-year. Unlike Brown & Tucker (2011),
my transformed matrix contains both positive and negative values, so cosine similarity is bounded
between one (completely similar) and negative one (completely dissimilar) instead of one and
zero. Since uncommon information should be contained in those MD&As that are dissimilar,
Uncommon, is this average cosine similarity multipled by negative one.

3.3

Examples of Uncommon Information

The following examples illustrate how Uncommon represents the contents of documents and distances between them. Table 3.2 shows the top 5 dimensions for the Software Industry for Fiscal
Year 2007.
Table 3.2: Top 10 Words for the Top 5 Dimensions in the Software Industry, Fiscal Year 2007

1
2
3
4
5
6
7
8
9
10

Competitive
Advantage
skill
context
expertis
sap
cultur
rigor
solv
creat
problem
analyst

Labor Force

Benchmarking

group
aim
worker
stoppag
colleg
council
serv
vice
degre
timeli

excel
innov
window
compel
portal
highest
mda
lowest
reader
year

Growth/
Change
frequent
requir
rapid
intens
rigor
expertis
sap
bargain
solv
secret

Contracting
defer
arrang
evid
valuat
recogn
consist
establish
compens
undeliv
residu

To aid in the interpretation of this example, I have provided my interpretation of the topic represented by each dimension, based on the word list and observing these words in context. As you
can see from the first and second dimensions, the software industry is focused on solving problems
and attracting top talent. The focus on words associated with growth and meeting benchmarks

22

suggest that it is also a rapidly changing and competitive industry.
Table 3.3 shows the top 10 word stems (by TF-IDF weighting) for the most and least common
MD&As in this industry year. The most common MD&A belongs to NCI Information Systems,
Inc. This company provides tech support to mostly government agencies. NCI is particularly
concerned with software contracts, discussing “backlog" and “subcontractor" a great deal. To see
many of these words in context, see B. It is not surprising that a company that produces primarily
stable, commonplace services would discuss common topics in its MD&A.
Table 3.3: Top 10 Words for the Most and Least Common MD&As in the Software Industry,
Fiscal Year 2007
1
2
3
4
5
6
7
8
9
10

NCI
Borland
backlog
vice
client
serv
subcontractor counsel
percent
degre
civilian
interoper
fring
stoppag
subcontract
document
contract
hold
predetermin
colleg
unfund
council

The least common MD&A in the Software Industry for 2007 belongs to Borland Software
Corporation. Borland provides services to support software developers. The prominence of the
word stem “serv" in its MD&A makes this apparent. Borland states in its MD&A that it is "a
leading vendor of Open Application Lifecycle Management solutions, or ALM... a new, customercentric approach to helping IT organizations transform software delivery" (emphasis mine). It
“differentiate[s] our products and solutions from those of our competitors based on cross-platform
interoperability", hence the word stem “interoper" in their top words list. Presumably to give
shareholders confidence, they spend a significant amount of time discussing the qualifications of
their top management, leading to word stems such as “vice", “counsel", “degree" and “college".
Borland appears to be pioneering in a niche market, leading to a highly uncommon MD&A. For
an excerpt of this MD&A, see B.

23

CHAPTER 4
TESTS OF HYPOTHESES AND EMPIRICAL RESULTS

4.1

Uncommon Information and Firm-Specific Returns

In Hypothesis 1, I predict that firm-specific returns and characteristics are determinants of uncommon information. Observing an association between uncommon information and firm-specific
returns and characteristics prior to the disclosure is consistent with economic events affecting both
variables. To this end, I examine the relationship between Uncommon and two measures of firmspecific returns in the year prior to the disclosure. I examine market beta (Beta) because the model
in Lambert et al. (2007) predicts that investors respond to the presence of firm-specific information by reducing their estimates of beta. I also examine return synchronicity, which measures
the portion of firm returns explained by both the market (as does beta) and firms in the same
industry. I calculate return synchronicity, ReturnSync, following Piotroski & Roulstone (2004):
R2
, where R2 is the the R-squared from a regression of weekly firm returns on conlog
2
1−R
current and lagged value-weighted market and industry returns over the fiscal year. I hypothesize
that uncommon information is associated with the prior incorporation into stock prices of a higher
amount of firm-specific information, leading to a negative association between Uncommon and the
market beta and return synchronicity in the year prior to the disclosure.
I also examine other factors that may cause a firm to differ economically from other firms.
Young firms (Age) tend to have higher product differentiation than other firms (Khan & Manopichetwattana, 1989), which might lead to uncommon information, while regulated firms (Reg), tend to
be more homogenous. In addition to market beta, I examine firm size (Size), book-to-market
ratio (BT M) and leverage (Leverage). These common sources of risk may be associated with
Uncommon if risky firms are more likely to provide uncommon information.
In Section 1, I suggest that managers make a cost-benefit tradeoff in deciding whether to pro-

24

vide uncommon information. As such, firms might be more likely to provide uncommon information if the benefits of doing so are higher. As such, I predict that uncommon information is more
likely to be present when other sources of information, such as the firm’s earnings, do not fully
reflect the economic differences between the firm and its peers (for instance, if the differences
are forward-looking). I include earnings synchronicity, EarnSync, and interact it with Beta and
ReturnSync, to see if uncommon information is more prevalent among firms with economic differences that are not reflected in earnings. Summary earnings is also a poorer reflection of a firm’s
economics when the firm’s income is more diverse (Piotroski & Roulstone, 2004), so I include the
number of business and geographical segments (NBUSSEG and NGEOSEG); RevConc, a firmspecific, revenue-based Herfindahl index to reflect the diversification of the firm across multiple
segments; and earnings volatility, EarnVol. Analysts and institutional owners are also significant
sources of information about the firm, so firms might make more uncommon information when
they have lower analyst coverage and institutional ownership. On the other hand, analysts and institutional owners are attracted to better disclosures, so the reverse could also be true. Since firms
are sensitive to proprietary costs (Verrecchia, 1983), they may be less likely to provide uncommon
information when their industry is less concentrated (IndConc).
I also examine factors that might mechanically impact Uncommon: the number of words in the
MD&A (Length) and the number of firms in the industry (NumFirms). Brown & Tucker (2011)
show that the cosine similarity measure causes documents that are longer to be more similar to all
other documents, reducing differences between them. Although I adjust my similarity measure for
document length, I include Length to see if any of this mechanical relationship remains. NumFirms
might impact Uncommon in two ways. More firms in the industry might make it easier to find close
peers for the firm, reducing Uncommon. On the other hand, if industry classification is noisy, a
larger industry might indicate diverse firms being grouped together, increasing Uncommon.
Table 4.1 shows descriptive statistics for Uncommon and the above determinants and Table
4.2 shows univariate Pearson (above diagonal) and Spearman (below diagonal) correlations. The
average (median) firm has a highly similar MD&A to its peers, with a cosine similarity of 0.795

25

(0.813) (keeping in mind that Uncommon is cosine similarity multiplied by −1). In the univariate
case, Uncommon is positively associated with Beta and unassociated with ReturnSync, contrary
to predictions. Results on riskiness are mixed: firms providing uncommon information are lower
growth and less leveraged, but also younger and in less competitive industries.
Table 4.1: Descriptive Statistics for Determinants Analysis
Statistic
Uncommon
Beta
ReturnSync
EarnSync
Age
Reg
Size
BT M
Leverage
NBUSSEG
NGEOSEG
RevConc
EarnVol
Ana f ollow
Instown
IndConc
Length
NumFirms

N

Mean

St. Dev.

Min

Median

Max

8,276
8,276
8,276
8,276
8,276
8,276
8,276
8,276
8,276
8,276
8,276
8,276
8,276
8,276
8,276
8,276
8,276
8,276

−0.795
1.186
−1.539
−2.364
17.495
0.068
2,070.291
0.590
0.201
5.000
5.981
0.209
0.065
8.049
0.595
0.124
7,140.683
95.674

0.138
0.727
1.365
2.386
12.114
0.252
4,299.256
0.429
0.191
4.050
5.604
0.369
0.063
7.261
0.254
0.080
4,305.407
61.095

−0.998
−0.217
−11.839
−20.829
3
0
2.681
0.001
0.000
1
1
0.037
0.001
1
0.001
0.024
280
20

−0.813
1.079
−1.341
−1.957
13
0
527.195
0.494
0.168
3
3
0.163
0.044
6
0.624
0.109
6,417
78

−0.290
3.691
1.699
4.605
66
1
35,412.500
3.798
0.777
30
66
31.699
0.349
47
1.000
0.716
57,467
326

All variables defined in Table A.1

26

Table 4.2: Correlations between Determinants
1
1. Uncommon
2. Beta
3. ReturnSync
4. EarnSync
5. Age
6. Reg
7. Size
8. BT M
9. Leverage
10. NBUSSEG
11. NGEOSEG
12. RevConc
13. EarnVol
14. Ana f ollow
15. Instown
16. IndConc
17. Length
18. NumFirms

0.11
-0.02
0.02
-0.09
-0.06
0.01
-0.11
-0.17
-0.01
0.08
-0.05
0.13
0.02
-0.02
-0.22
0.09
0.64

2
0.11
0.18
0.07
-0.17
-0.20
-0.08
-0.04
-0.19
-0.10
0.19
-0.06
0.38
0.12
0.09
0.04
0.12
0.18

3
-0.02
0.15
0.03
0.23
0.07
0.49
-0.01
0.05
0.18
0.18
-0.17
-0.19
0.39
0.42
-0.10
0.37
0.01

4
0.02
0.06
0.02
-0.04
-0.07
0.03
-0.03
-0.05
0.01
0.07
-0.05
0.15
0.07
0.04
0.03
0.01
0.01

5
-0.10
-0.18
0.22
-0.04
0.24
0.25
0.08
0.10
0.19
0.07
-0.12
-0.28
0.08
0.10
-0.08
0.08
-0.17

6
-0.07
-0.19
0.07
-0.07
0.32
0.15
0.07
0.18
0.13
-0.15
0.05
-0.25
0.04
-0.12
-0.30
0.10
-0.08

7
0.04
-0.07
0.26
0.06
0.27
0.12
-0.35
0.13
0.26
0.18
-0.22
-0.30
0.73
0.52
-0.05
0.34
0.00

8
-0.08
0.00
-0.06
-0.01
0.01
0.02
-0.16
0.13
0.05
-0.08
0.04
-0.18
-0.23
-0.07
-0.05
0.05
-0.14

27

9
-0.15
-0.15
0.04
-0.04
0.09
0.16
0.02
0.11
0.09
-0.18
0.07
-0.33
0.03
0.02
-0.01
0.11
-0.19

10
-0.01
-0.12
0.17
0.00
0.27
0.16
0.22
0.01
0.07
0.40
-0.76
-0.19
0.11
0.14
-0.07
0.30
-0.06

11
0.06
0.21
0.15
0.07
0.09
-0.13
0.15
-0.07
-0.16
0.25
-0.79
0.10
0.15
0.22
0.03
0.22
0.07

12
-0.01
-0.02
-0.08
-0.01
-0.03
0.01
-0.05
0.01
0.01
-0.19
-0.19
0.02
-0.13
-0.21
0.01
-0.26
-0.03

13
0.12
0.33
-0.14
0.11
-0.25
-0.16
-0.14
-0.12
-0.24
-0.19
0.04
0.01
-0.06
-0.15
0.12
-0.13
0.18

14
0.03
0.09
0.32
0.05
0.11
0.02
0.62
-0.19
-0.01
0.10
0.14
-0.04
-0.04
0.50
-0.03
0.29
0.02

15
-0.03
0.08
0.42
0.04
0.10
-0.11
0.23
-0.11
0.01
0.11
0.18
-0.07
-0.17
0.42

16
-0.17
0.04
-0.10
0.03
-0.09
-0.20
-0.01
-0.02
-0.01
-0.07
0.00
0.01
0.10
-0.03
-0.04

17
0.08
0.08
0.31
0.00
0.17
0.16
0.28
0.05
0.12
0.30
0.13
-0.08
-0.11
0.26
0.25
-0.10

18
0.58
0.21
0.01
0.00
-0.20
-0.09
-0.01
-0.11
-0.15
-0.07
0.07
-0.03
0.20
0.04
0.00
-0.24
0.07

0.01
0.31 -0.08
0.01
-0.31
0.11
All variables defined in Table A.1
Pearson correlations are above the diagonal, Spearman below.
Correlations in bold are significant at below the 5% level.

Table 4.3 shows the results of the regression of Uncommon on its determinants. Consistent
with expectations and contrary to the univariate results, Uncommon is negatively associated with
both Beta and ReturnSync at less than the 1% level. This suggests that it is important to control
for other factors in evaluating these relationships. In addition, Uncommon is significantly positively associated with EarnSync in one of the two specifications, indicating that firms make more
uncommon information when earnings do not reflect differences between firms. Also consistent
with this prediction, the interactions between Beta and EarnSync and between ReturnSync and
EarnSync have negative coefficients that are significant at less than the 1% level, suggesting that
a firm’s disclosures are more uncommon when there is a mismatch between return synchronicity and earnings synchronicity. Firms with uncommon information also have more business and
geographic segments, as predicted. However, these firms are also less diversified across their segments, inconsistent with predictions. Uncommon is negatively associated with analyst following
and institutional ownership, adding to the evidence that uncommon information is more likely
when investors have fewer sources for firm-specific information.1
It does not appear that firms with more uncommon information are riskier along the traditional
dimensions of risk. Uncommon is negatively associated with Beta, positively associated with Size,
unassociated with BT M and negatively associated with leverage. In addition, Uncommon is positively associated with Length, allaying fears that the measure is mechanically affected by document
length and providing preliminary evidence for Hypothesis 4. However, as is clear from Table 4.2
and the coefficient on NumFirms, Uncommon is highly positively associated with NumFirms,
suggesting that it is important to control for NumFirms in all regressions that contain Uncommon.
The results above suggest that disclosures of uncommon information are affected by the underlying economics and characteristics of the firm. To test Hypothesis 2, I examine the association
between uncommon information and the change in return synchronicity following the disclosure,
controlling for firm characteristics. Finding an association in this analysis suggests that the disclosure of uncommon information affects the firm’s information environment. I regress ∆ReturnSync,
1 It

is also possible that analysts are more likely to follow firms with common information.

28

Table 4.3: Regression of Uncommon on Determinants
Dependent variable:
Uncommon
Beta

−0.008∗∗∗
(0.002)

−0.004∗∗∗
(0.001)
−0.0004
(0.001)
0.0002
(0.0002)
0.024
(0.018)
0.008∗∗∗
(0.002)
0.002
(0.003)
−0.025∗∗
(0.010)
0.014∗∗∗
(0.004)
0.010∗∗∗
(0.003)
0.025∗∗∗
(0.009)
−0.008
(0.028)
−0.008∗∗∗
(0.002)
−0.021∗∗∗
(0.007)
−0.007
(0.005)
0.012∗
(0.006)
0.111∗∗∗
(0.012)

ReturnSync
EarnSync
Age
Reg
Size
BT M
Leverage
NBUSSEG
NGEOSEG
RevConc
EarnVol
Ana f ollow
Instown
IndConc
Length
NumFirms
Beta × EarnSync

0.003∗∗∗
(0.001)
0.0002
(0.0002)
0.023
(0.018)
0.007∗∗∗
(0.002)
0.002
(0.003)
−0.024∗∗
(0.010)
0.014∗∗∗
(0.004)
0.010∗∗∗
(0.003)
0.025∗∗∗
(0.009)
0.006
(0.028)
−0.007∗∗∗
(0.002)
−0.022∗∗∗
(0.007)
−0.007
(0.005)
0.012∗∗
(0.006)
0.111∗∗∗
(0.012)
−0.002∗∗∗
(0.001)

−0.001∗∗
(0.0003)
−1.325∗∗∗
(0.040)
Yes
Yes
8,276
0.391

ReturnSync × EarnSync
Constant
Year fixed effects
Industry fixed effects
Observations
Adjusted R2
Note:

−1.314∗∗∗
(0.042)
Yes
Yes
8,276
0.391

∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01
All variables defined in Table A.1
Standard errors in parentheses are clustered at the firm and year level.

29

the change in return synchronicity from the year prior to the disclosure to the year following, on
Uncommon and ∆Uncommon, the level and change in uncommon information, and control variables. All variables in this analysis are defined in detail in Table A.1. The coefficient on Uncommon
(∆Uncommon) reflects the association between the level of (change in) uncommon information and
a future decline in return synchronicity. To control for other characteristics of the MD&A, I include the year-over-year change in the MD&A, Change, after Brown & Tucker (2011) obtained
from Stephen Brown’s website (Brown, 2013), and ∆Length, the change in the total number of
words in the MD&A. Following Piotroski & Roulstone (2004), I control for known determinants
of return synchronicity. I include ∆EarnSync, the change in earnings synchronicity (calculated
similarly to return synchronicity using return on assets in place of returns and calculated over the
last 12 quarters); Reg, whether the firm is in a regulated industry (two-digit SIC code of 49 or 62);
∆Instown, the change in the percentage of the firm owned by institutions from the 13-F filings;
IndConc, industry concentration (the log of an industry-specific revenue-based Herfindahl index);
∆RevConc, the change in the diversification of the firm (the log of a firm-specific revenue-based
Herfindahl index); ∆Ana f ollow, the change in the number of analysts issuing a forecast for the
firm’s earnings during the year; ∆Size, the change in the size of the firm (the log of the market
value of equity at year-end); ∆EarnVol, the change in operating earnings volatility (the standard
deviation of earnings over the last five years). I also include ∆NumFirms, the change in the log of
the number of firms in the industry, since it is highly correlated with Uncommon.
Table 4.4 shows descriptive statistics for these variables. Consistent with the determinants analysis, MD&As across firms in the same industry are quite similar. In addition, Uncommon changes
little from year-to-year, with a mean (median) absolute value of 0.085 (0.061) and a standard deviation of 0.081. There is, however, signicant variation, with absolute changes up to 0.586 relative
to a mean level of -0.791 for Uncommon. The year-over-year change in MD&A is consistent with
Brown & Tucker (2011), with a mean and median between 10% and 15%.
Table 4.5 shows univariate Pearson (Spearman) correlations above (below) the diagonal for
the regression variables. Consistent with Hypothesis 2, both Uncommon and ∆Uncommon are

30

Table 4.4: Descriptive Statistics for ReturnSync Regressions
Statistic
∆ReturnSync
Uncommon
∆Uncommon
|∆Uncommon|
Change
∆Length
∆EarnSync
Reg
∆Instown
∆IndConc
∆RevConc
∆Ana f ollow
∆Size
∆EarnVol
∆NumFirms

N
3,272
3,272
2,567
2,567
3,272
3,272
3,272
3,272
3,272
3,272
3,272
3,272
3,272
3,272
3,272

Mean
0.070
−0.791
0.002
0.085
0.148
533.586
0.008
0.057
0.033
−0.018
0.0001
0.273
226.259
−0.004
−0.762

St. Dev.
1.421
0.140
0.118
0.081
0.138
2,271.784
2.937
0.232
0.143
0.079
0.105
2.517
1,412.057
0.031
18.293

Min
−10.549
−0.998
−0.586
0.0001
0.002
−16,427
−16.160
0
−0.612
−0.575
−0.831
−14
−9,459.574
−0.281
−53

Median
0.071
−0.809
0.001
0.061
0.102
433.5
0.008
0
0.025
−0.005
0.000
0
34.196
−0.001
−2

Max
9.953
−0.299
0.576
0.586
0.858
16,160
14.501
1
0.672
0.371
0.678
19
23,132.040
0.172
128

All variables defined in Table A.1

significantly associated with a decline in return synchronicity in the year following the disclosure.
In addition, the MD&A is more likely to contain a higher level and an increase in uncommon
information as it increases in length over the prior year.

31

Table 4.5: Correlations between ReturnSync Variables
1
∆ReturnSync
Uncommon
∆Uncommon
Change
∆Length
∆EarnSync
Reg
∆Instown
∆IndConc
∆RevConc
∆Ana f ollow
∆Size
∆EarnVol
∆NumFirms

-0.05
-0.05
-0.02
-0.07
-0.02
0.01
0.04
-0.02
-0.03
0.03
0.04
0.04
0.03

2
-0.04

3
-0.05
0.44

4
5
-0.02 -0.05
0.05 0.07
0.40
0.03 0.10
0.05 0.03
0.06
0.09 0.11 0.07
-0.01 -0.03 -0.01 -0.02
-0.09 0.00 -0.04 0.03
0.00 -0.05 -0.01 -0.03
0.09 0.04 -0.05 0.06
0.05 -0.01 -0.03 -0.01
-0.01 -0.03 0.00 -0.04
0.02 -0.01 -0.06 0.01
-0.04 0.00 0.03 -0.02
-0.17 0.01 0.07 0.01

6
7
-0.01 0.01
-0.02 -0.09
-0.03 -0.01
-0.01 -0.05
-0.06 0.03
0.00
0.00
0.02 -0.03
0.00 0.01
0.01 -0.03
-0.01 -0.03
-0.01 0.05
0.01 0.02
-0.03 0.04

8
9
10
11
12
13
14
0.04 -0.01 -0.06 0.01 -0.01 0.04 -0.01
-0.01 0.08 0.04 0.00 0.06 -0.03 -0.12
-0.06 0.02 -0.04 -0.02 0.00 0.02 0.02
0.01 -0.06 -0.06 0.03 -0.03 0.02 0.10
0.00 0.03 -0.02 -0.03 0.01 0.01 0.03
0.02 -0.01 0.00 -0.02 0.00 0.00 -0.01
-0.03 0.03 -0.03 -0.04 0.02 0.03 0.02
-0.01 0.01 0.08 0.01 0.02 0.00
-0.02
0.04 0.00 0.02 -0.06 -0.19
0.04 0.00
0.02 -0.02 -0.08 -0.07
0.10 -0.03 0.05
0.00 0.03 0.01
0.15 0.07 -0.01 0.07
0.00 -0.04
0.03 -0.06 -0.04 0.03 -0.01
0.11
0.00 -0.29 -0.06 0.02 -0.06 0.06
All variables defined in Table A.1
Pearson correlations are above the diagonal, Spearman below.
Correlations in bold are significant at below the 5% level.

32

Table 4.6 shows the multivariate results. Consistent with Hypothesis 2, both the level and the
change in Uncommon are significantly associated (at the 10% level or greater) with a decline in
return synchronicity. With the exception of the change in earnings volatility, which is significantly
associated with an increase in return synchronicity, the other variables in the regression are not
significant.
The above results are consistent with disclosed uncommon information being both a result of
uncommon firm economics and characteristics and a source of information about the firm. However, an alternative explanation for my findings is that uncommon information released prior to
the MD&A disclosure is processed slowly, leading to a decline in return synchronicity following
the disclosure. In the next Section, I examine the timing of the market response to uncommon
information to determine whether it precedes or follows disclosure.

4.2
4.2.1

Uncommon Information and Information Processing Costs
Uncommon Information and Delayed Market Response

In this Section, I examine the effect of uncommon information on the firm’s returns around the
MD&A disclosure. In Hypothesis 3, I predict that a significant portion of the market response
to uncommon information is delayed, leading to abnormal returns in the window following its
release. To test Hypothesis 3, I regress the absolute value of the firm’s cumulative abnormal returns
in various windows on Uncommon and control variables. I calculate cumulative abnormal returns
using a market model estimated over 255 days ending 63 days prior to the disclosure. I choose 63
days because 95% of firms report earnings fewer than 63 days prior to the 10-K filing. I take the
absolute value of returns because I do not know the sign of the news that is reported in uncommon
information. This is consistent with theoretical research (e.g. Kim & Verrecchia, 1991) that uses
the absolute value of returns as a measure of the absolute value of news. I measure the absolute
value of cumulative abnormal returns (henceforth CAR) over four windows. The first two windows
examine the market response around the disclosure: CAR−1,1 for the day before to the day after the

33

Table 4.6: Regression of ∆ReturnSync on Uncommon
Dependent variable:
∆ReturnSync
Uncommon

−0.297∗
(0.177)
−0.421∗∗
(0.211)
0.116
(0.149)
−0.00001
(0.00001)
−0.006
(0.009)
−0.243
(0.161)
0.290
(0.254)
0.336
(0.450)
−0.694
(0.525)
0.007
(0.007)
−0.00003
(0.00003)
1.791∗∗
(0.832)
−0.003
(0.002)
−1.230

∆Uncommon
Change
∆Length
∆EarnSync
Reg
∆Instown
∆IndConc
∆RevConc
∆Ana f ollow
∆Size
∆EarnVol
∆NumFirms
Constant
Year fixed effects
Industry fixed effects
Observations
Adjusted R2

0.122
(0.212)
−0.00002∗
(0.00001)
−0.006
(0.009)
−0.237
(0.182)
0.245
(0.260)
0.270
(0.358)
−0.724
(0.561)
0.007
(0.007)
−0.00002
(0.00002)
1.871∗∗
(0.788)
−0.002
(0.002)
−1.440∗∗∗
(0.343)
Yes
Yes

Yes
Yes

3,272
0.065

2,567
0.056
∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01

Note:

All variables defined in Table A.1
Standard errors in parentheses are clustered at the firm and year level.

34

filing and CAR2,60 for the post-disclosure period from 2 to 60 days after the disclosure. In order to
provide evidence that an observed market response is not due to unobserved characteristics of the
firm, I also examine a pseudo-event period prior to the earnings announcement. This examination
also provides evidence that the market is responding to uncommon information in the MD&A, as
opposed to previously disclosed information. PCAR−1,1 measures CAR over the three days starting
63 days prior to the earnings announcement (-63 to -61) and PCAR2,60 measures CAR over the 58
days leading up to the earnings announcement window (-60 to -2). Results for CAR−1,1 or CAR2,60
that are not mirrored in the corresponding pseudo-event period increases confidence that the result
is due to a characteristic of the MD&A disclosure and not the firm. All variables in this analysis
are defined in detail in Table A.1.
I control for the year-over-year change (Change) and length (Length) of the MD&A in order to
separate from Uncommon the effect of additional information or information processing costs that
might be contained in changed or long MD&As. I also control for other factors that impact returns.
Since prior research has found significant drift associated with earnings announcements (i.e. Ball
& Brown, 1968), I include CAR calculated during the earnings announcement window (EA−1,1 )
and the log of the absolute value of unexpected earnings (UE) to control for news contained in
earnings. I also include known sources of risk: market beta (Beta) and the firm’s size (Size),
book-to-market ratio (BT M) and leverage (Leverage). To control for uncertainty surrounding the
firm’s economics, I also control for the volatility of returns (ReturnVol) and earnings (EarnVol).
Following Brown & Tucker (2011), I also include FileLate, an indicator variable equal to one if
the 10-K was filed more than 90 days after the fiscal year-end and zero otherwise, and NumItems,
the log of the number of non-missing Compustat items. To control for the effects of transaction
costs, I include trading volume for each window (Volume) and the stock price at the end of the
fiscal year (Price). Finally, consistent with the findings in the determinants analysis, I control for
the impact of the number of firms per industry on Uncommon.
Tables 4.7 and 4.8 provide descriptive statistics and univariate correlations between these variables. In the univariate case, Uncommon has significant Pearson correlations with CAR in all

35

windows, emphasizing the importance of controlling for sources of risk and other firm-specific
factors in the multivariate case.
Table 4.7: Descriptive Statistics for CAR Regressions
Statistic
CAR−1,1
CAR2,60
PCAR−1,1
PCAR2,60
Volume−1,1
Volume2,60
PVolume−1,1
PVolume2,60
Uncommon
Change
Length
EA−1,1
VolumeEA−1,1
UE
Beta
Size
BT M
Leverage
ReturnVol
EarnVol
FileLate
NumItems
Price
NumFirms

N

Mean

St. Dev.

Min

Median

Max

5,767
5,767
5,767
5,767
5,767
5,767
5,767
5,767
5,767
5,767
5,767
5,767
5,767
5,767
5,767
5,767
5,767
5,767
5,767
5,767
5,767
5,767
5,767
5,767

0.034
0.169
0.036
0.159
2.585
5.697
2.407
5.578
−0.779
0.144
6,390.743
0.056
3.001
−6.061
1.057
1,714.738
0.565
0.205
0.126
0.066
0.168
292.687
24.113
108.145

0.043
0.173
0.043
0.177
1.324
0.977
1.471
0.989
0.143
0.136
3,830.657
0.061
1.292
1.722
0.737
3,521.883
0.408
0.184
0.073
0.065
0.374
43.494
24.656
71.251

0.00002
0.0001
0.00000
0.00000
−31.841
0.935
−33.964
0.660
−0.998
0.002
259
0.00001
−3.326
−11.864
−0.217
2.786
0.0004
0.000
0.029
0.001
0
133
0.320
20

0.021
0.118
0.023
0.108
2.691
5.754
2.504
5.627
−0.791
0.098
5,689
0.037
3.111
−6.202
0.919
508.615
0.480
0.176
0.110
0.044
0
296
19.550
89

0.516
1.960
0.579
3.635
6.335
9.343
6.955
8.971
−0.288
0.883
37,819
0.746
6.591
2.258
3.683
34,929.220
3.770
0.776
0.521
0.349
1
391
769.400
326

All variables defined in Table A.1

36

Table 4.8: Correlations between CAR Variables
1
1. CAR−1,1
2. CAR2,60
3. PCAR−1,1
4. PCAR2,60
5. Volume−1,1
6. Volume2,60
7. PVolume−1,1
8. PVolume2,60
9. Uncommon
10. Change
11. Length
12. EA−1,1
13. VolumeEA−1,1
14. UE
15. Beta
16. Size
17. BT M
18. Leverage
19. ReturnVol
20. EarnVol
21. FileLate
22. NumItems
23. Price
24. NumFirms

0.20
0.16
0.21
0.19
0.11
0.08
0.10
0.02
0.11
-0.13
0.21
0.09
0.12
0.20
-0.19
-0.01
-0.09
0.35
0.26
0.15
-0.10
-0.24
0.01

2
0.23
0.15
0.24
0.10
0.16
0.11
0.13
0.05
0.13
-0.10
0.18
0.11
0.17
0.21
-0.24
0.01
-0.08
0.40
0.30
0.14
-0.03
-0.28
0.04

3
0.17
0.21
0.19
0.08
0.11
0.12
0.14
0.06
0.08
-0.07
0.15
0.10
0.07
0.18
-0.13
-0.02
-0.08
0.34
0.25
0.10
-0.07
-0.19
0.02

4
0.26
0.27
0.21
0.13
0.16
0.18
0.22
0.04
0.11
-0.10
0.19
0.17
0.14
0.22
-0.20
-0.03
-0.06
0.44
0.30
0.13
-0.08
-0.26
0.03

5
0.19
0.08
0.07
0.12
0.87
0.71
0.80
0.05
-0.06
0.21
0.22
0.78
-0.14
0.36
0.39
-0.31
-0.07
0.22
0.27
-0.08
0.35
0.24
0.04

37

6
0.10
0.16
0.10
0.16
0.78
0.78
0.88
0.05
-0.04
0.21
0.24
0.84
-0.15
0.41
0.40
-0.33
-0.08
0.28
0.31
-0.10
0.37
0.24
0.04

7
0.06
0.06
0.10
0.16
0.52
0.63
0.88
0.05
-0.02
0.18
0.20
0.74
-0.13
0.38
0.36
-0.31
-0.08
0.29
0.33
-0.07
0.33
0.19
0.04

8
0.09
0.11
0.14
0.25
0.70
0.88
0.71

9
0.04
0.07
0.06
0.03
0.04
0.04
0.03
0.04

10
0.11
0.14
0.08
0.12
-0.03
-0.02
0.00
0.00
0.03

11
-0.12
-0.10
-0.05
-0.10
0.16
0.18
0.14
0.17
0.14
-0.20

12
0.29
0.19
0.14
0.24
0.18
0.22
0.15
0.20
0.04
0.10
-0.06

0.06
-0.02 0.04
0.19 0.16 -0.24
0.23 0.04 0.11 -0.06
0.84 0.06 -0.03 0.19 0.39
-0.15 -0.04 0.08 -0.08 0.10
0.42 0.11 0.05 0.04 0.25
0.39 0.03 -0.15 0.36 -0.10
-0.34 -0.10 -0.01 -0.03 -0.07
-0.10 -0.18 -0.04 0.08 -0.10
0.32 0.07 0.18 -0.17 0.33
0.35 0.05 0.16 -0.11 0.29
-0.09 0.02 0.16 -0.17 0.09
0.33 -0.05 -0.15 0.39 0.06
0.22 -0.02 -0.13 0.18 -0.16
0.05 0.66 0.00 0.16 0.01
Continued on next page

Table 4.8 (cont’d)

1. CAR−1,1
2. CAR2,60
3. PCAR−1,1
4. PCAR2,60
5. Volume−1,1
6. Volume2,60
7. PVolume−1,1
8. PVolume2,60
9. Uncommon
10. Change
11. Length
12. EA−1,1
13. VolumeEA−1,1
14. UE
15. Beta
16. Size
17. BT M
18. Leverage
19. ReturnVol
20. EarnVol
21. FileLate
22. NumItems
23. Price
24. NumFirms

13
0.08
0.08
0.09
0.17
0.70
0.84
0.61
0.84
0.05
-0.01
0.17
0.38

14
0.18
0.21
0.13
0.17
-0.14
-0.15
-0.10
-0.15
-0.03
0.09
-0.07
0.12
-0.14

15
0.15
0.19
0.16
0.19
0.31
0.40
0.31
0.41
0.10
0.04
0.05
0.21
0.38
0.07

16
-0.10
-0.13
-0.07
-0.12
0.16
0.18
0.14
0.18
0.04
-0.07
0.24
-0.07
0.17
-0.27
-0.02

17
0.07
0.11
0.06
0.04
-0.25
-0.29
-0.21
-0.29
-0.08
0.02
-0.05
0.03
-0.29
0.35
-0.13
-0.20

18
19
-0.05 0.33
-0.05 0.40
-0.05 0.36
-0.04 0.46
-0.04 0.18
-0.06 0.25
-0.04 0.23
-0.07 0.30
-0.16 0.07
-0.04 0.17
0.10 -0.14
-0.08 0.29
-0.09 0.22
-0.13
0.14 0.26
0.40 0.06
-0.22 0.43
0.36 -0.48 0.01
0.01 -0.18
-0.33 0.31 -0.18 -0.40
0.11 0.06
-0.12 0.11 -0.24 0.10 0.15
-0.13
0.26 0.25 0.49 -0.35 -0.04 -0.17
0.31 0.21 0.49 -0.23 -0.19 -0.29 0.60
-0.08 0.19 0.04 -0.29 0.09 0.01 0.22
0.37 -0.06 0.23 0.27 -0.18 -0.14 -0.06
0.22 -0.54 -0.21 0.76 -0.40 0.07 -0.46
0.04 -0.05 0.11 0.00 -0.14 -0.19 0.05

20
0.22
0.26
0.21
0.27
0.19
0.25
0.22
0.29
0.08
0.13
-0.07
0.21
0.24
0.19
0.42
-0.10
-0.11
-0.22
0.48

21
0.17
0.16
0.12
0.13
-0.07
-0.11
-0.08
-0.09
0.03
0.15
-0.14
0.11
-0.09
0.20
0.02
-0.13
0.15
0.01
0.20
0.14

22
-0.07
-0.02
-0.04
-0.06
0.31
0.38
0.28
0.35
-0.08
-0.11
0.26
0.04
0.38
-0.04
0.25
0.15
-0.16
-0.08
-0.01
0.13
-0.17

23
-0.14
-0.17
-0.11
-0.16
0.15
0.16
0.10
0.15
-0.02
-0.07
0.13
-0.11
0.15
-0.35
-0.14
0.38
-0.28
0.02
-0.26
-0.19
-0.16
0.05

24
0.01
0.05
0.04
0.03
0.00
0.00
0.01
0.02
0.61
-0.01
0.14
0.03
0.01
-0.07
0.12
-0.02
-0.12
-0.16
0.07
0.07
0.00
-0.20
-0.03

0.16
0.14 -0.22
-0.37 -0.26 0.11
0.03 0.00 -0.11 -0.04
All variables defined in Table A.1
Pearson correlations are above the diagonal, Spearman below.
Correlations in bold are significant at below the 5% level.

38

Table 4.9 shows the results of the regression. Consistent with the MD&A being a source of
uncommon information, Uncommon is unassociated with market returns prior to the disclosure.
In addition, consistent with a delayed market response to uncommon information, the coefficient
on Uncommon is positive and significant at less than the 1% level in the post-disclosure window
but insignificant in the three-day window around the 10-K filing. In addition, the coefficient on
Change is positive and significant in the post-disclosure window, suggesting that a change in the
MD&A also imposes significant information processing costs on investors. Change is likewise
insignificant in the pre-disclosure period. Contrary to the findings of Brown & Tucker (2011),
Change is insignificant in the disclosure window. However, due to data limitations of calculating Uncommon and including additional control variables, my sample size is significantly smaller
(5,767 firm years to their 23,083), which may limit the power of my test to detect an effect. Note,
however, that significance in the disclosure window for either Change or Uncommon does not contradict the hypothesis that changes in the MD&A and uncommon information impose significant
information processing costs on investors. It would simply indicate that some investors are faster
at processing this information than others.

4.2.2

Uncommon Information and Readability

Hypothesis 4 predicts that uncommon information will be longer and less readable than common
disclosure. To test this hypothesis, I examine two common readability measures from the linguistics and accounting literature: the Gunning Fog index (Fog) and the total number of words
(Length). Following Li (2008), I test if Uncommon has explanatory power for these two measures
after controlling for known determinants. Since the Fog index is calculated as the summary of two
components, I also examine whether Uncommon explains each of these components: percentage
of complex words (where a complex word has three syllables or more) and words per sentence.
All variables in this analysis are defined in detail in Table A.1.
I calculate the Fog index following Li (2008), with one modification. Prior literature has calculated the Fog index using the Fathom package, which counts the syllables per word and words

39

Table 4.9: Regression of CAR and PCAR on Uncommon

CAR−1,1
CAR−1,1
Volume−1,1

0.011∗∗∗
(0.002)

Volume2,60

Dependent variable:
CAR2,60
PCAR−1,1
0.158∗∗∗
(0.057)
−0.009
(0.007)
0.071∗∗∗
(0.012)

PCAR−1,1
0.001∗∗
(0.001)

PVolume−1,1
PVolume2,60
Uncommon
Change
Length
EA−1,1
VolumeEA−1,1
UE
Beta
Size
BT M
Leverage
ReturnVol
EarnVol
FileLate
NumItems
Price
NumFirms
Constant
Year fixed effects
Industry fixed effects
Observations
Adjusted R2

0.004
(0.005)
0.004
(0.004)
0.0004
(0.001)
0.148∗∗∗
(0.021)
−0.006∗∗∗
(0.002)
0.001∗∗
(0.0003)
−0.001
(0.001)
−0.001∗∗
(0.001)
−0.004∗∗
(0.002)
−0.005
(0.004)
0.065∗∗∗
(0.013)
0.010
(0.013)
0.004∗
(0.003)
0.009
(0.012)
−0.008∗∗∗
(0.002)
−0.001
(0.004)
−0.004
(0.069)
Yes
Yes
5,767
0.245

0.042∗∗∗
(0.016)
0.036∗∗
(0.016)
0.004
(0.004)
0.161∗∗
(0.067)
−0.024∗∗∗
(0.007)
0.002∗
(0.001)
−0.007∗
(0.004)
−0.016∗∗∗
(0.003)
−0.005
(0.006)
0.006
(0.013)
0.384∗∗∗
(0.052)
0.075∗
(0.040)
0.012∗∗
(0.006)
0.051
(0.037)
−0.024∗∗∗
(0.005)
0.011
(0.017)
−0.340
(0.223)
Yes
Yes
5,767
0.251

0.006
(0.005)
−0.003
(0.003)
0.003∗∗∗
(0.001)
0.004
(0.007)
0.002∗∗∗
(0.0005)
0.001
(0.001)
−0.002
(0.001)
−0.0005
(0.001)
−0.001
(0.002)
−0.002
(0.004)
0.143∗∗∗
(0.024)
0.019
(0.016)
0.003
(0.002)
0.011
(0.015)
−0.005∗∗∗
(0.002)
0.005
(0.005)
−0.070
(0.084)
Yes
Yes
5,767
0.152

PCAR2,60

0.052
(0.096)
−0.004
(0.003)
0.063∗∗∗
(0.013)
−0.017
(0.019)
0.010
(0.017)
0.003
(0.003)
0.249∗
(0.142)
−0.012∗∗∗
(0.004)
0.002
(0.002)
−0.015∗
(0.008)
−0.010∗∗∗
(0.002)
−0.013∗∗∗
(0.004)
0.00002
(0.010)
0.639∗∗∗
(0.092)
0.001
(0.050)
0.0001
(0.003)
−0.0004
(0.036)
−0.027∗∗∗
(0.006)
0.004
(0.018)
−0.109
(0.201)
Yes
Yes
5,767
0.288

∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01

Note:

All variables defined in Table A.1
Standard errors in parentheses are clustered at the firm and year level.

40

per sentence and then calculates the relevant summary statistics. However, the Fathom package
is designed for general use in (often noisy) machine-readable English text and makes certain assumptions about the structure of the text in calculating its metrics. One of these assumptions is
that a period (.) ends a sentence if it occurs between two word characters (word characters in Perl
are capital and lowercase letters, the digits 0 through 9 and the underscore) with an optional space.
This means that the Fathom package identifies a new sentence in almost all cases where it observes
a period, including periods that represent a decimal point in a number. As a result, text that contains
more decimalized numbers is calculated as having more, shorter sentences and is considered more
readable. I modify the Fathom package to exclude periods that precede digits in my calculation of
Fog. I calculate Length following the prior literature, since this calculation excludes numbers and
thus is not affected by this issue.
To test Hypothesis 4, I regress Fog, percentage of complex words (%Complex), words per sentence (W PS), and Length on uncommon information, the year-over-year change in the MD&A and
known determinants of Fog and Length. These determinants include the log of the market value of
equity (Size), the book-to-market ratio (BT M), the ratio of special items to total assets (Special),
return and earnings volatility (ReturnVol and EarnVol), the number of business and geographic
segments (NBUSSEG and NGEOSEG), the financial complexity of the firm, calculated as the log
of the number of non-missing items in Compustat (NumItems), an indicator variable if the firm
is incorporated in Delaware (DE), and indicator variables if the firm has had a seasoned equity
offering or merger/acquisition during the year (SEO and MA). Consistent with the determinants
analysis for Uncommon, I also include the log of the number of firms in the industry (NumFirms).
Table 4.10 shows descriptive statistics for these variables. The mean (median) of Fog for the
MD&A is 21.447 (21.389). This is somewhat higher than that reported in Li (2008), which may
be attributable to the issue with the Fathom package mentioned above. The mean (median) of
Length is 5,837 (4,971), higher than but in line with Li’s results. Uncommon and Change are also
in line with prior analyses. Table 4.11 shows univariate correlations between these variables. The
univariate results are consistent with predictions: Uncommon is significantly positively correlated

41

with Fog and Length. The relationship between uncommon information and the Fog index appears
to be due to uncommon information containing longer sentences rather than longer words. In
addition, Change is negatively correlated with Fog , W PS and Length but positively correlated
with %Complex.
Table 4.10: Descriptive Statistics for Readability Regressions
Statistic
Fog
%Complex
W PS
Length
Uncommon
Change
Size
BT M
Special
ReturnVol
EarnVol
NBUSSEG
NGEOSEG
NumItems
DE
SEO
MA
Age
NumFirms

N

Mean

St. Dev.

Min

Median

Max

7,842
7,842
7,842
7,842
7,842
7,842
7,842
7,842
7,842
7,842
7,842
7,842
7,842
7,842
7,842
7,842
7,842
7,842
7,842

21.447
27.867
25.751
5,837.828
−0.783
0.157
1,300.269
0.644
−0.017
0.139
0.074
4.973
5.732
294.547
0.561
0.863
0.328
15.742
101.958

1.263
1.645
2.634
3,719.285
0.142
0.143
3,184.431
0.510
0.113
0.079
0.067
3.814
5.284
34.738
0.496
0.344
0.469
10.808
65.428

18.078
24.309
20.743
259
−0.998
0.001
1.769
0.002
−5.056
0.029
0.002
1
1
196
0
0
0
1
20

21.389
27.830
25.569
4,971
−0.796
0.110
281.058
0.516
0.000
0.121
0.053
3
3
294
1
1
0
12
87

25.799
31.549
33.497
44,672
−0.288
0.956
34,884.640
3.848
1.032
0.521
0.349
30
65
391
1
1
1
66
326

All variables defined in Table A.1

42

Table 4.11: Correlations between Readability Variables
1
1. Fog
2. %Complex
3. W PS
4. Length
5. Uncommon
6. Change
7. Size
8. BT M
9. Special
10. ReturnVol
11. EarnVol
12. NBUSSEG
13. NGEOSEG
14. NumItems
15. DE
16. SEO
17. MA
18. Age
19. NumFirms

0.54
0.84
0.24
0.08
-0.06
0.10
-0.08
-0.06
-0.01
0.03
0.01
0.02
0.10
0.08
0.03
0.04
-0.09
0.12

2
0.55
0.04
-0.05
0.02
0.06
0.03
-0.01
0.01
-0.06
-0.08
-0.02
-0.07
-0.06
0.00
-0.03
0.06
-0.03
0.03

3
0.85
0.04
0.32
0.08
-0.11
0.11
-0.09
-0.07
0.03
0.08
0.02
0.08
0.16
0.09
0.07
0.01
-0.10
0.12

4
0.22
-0.05
0.29
0.15
-0.27
0.39
-0.11
-0.15
-0.08
-0.01
0.38
0.39
0.55
0.06
0.15
0.15
-0.04
0.15

5
0.07
0.02
0.08
0.12
0.03
0.01
-0.08
-0.07
0.14
0.17
0.01
0.10
0.07
0.05
0.03
0.02
-0.07
0.65

6
-0.06
0.03
-0.09
-0.21
0.02
-0.18
0.02
-0.07
0.16
0.13
-0.09
-0.04
-0.24
0.00
-0.04
0.04
0.05
0.03

7
0.03
0.03
0.02
0.22
0.05
-0.07
-0.45
-0.03
-0.34
-0.22
0.21
0.24
0.32
0.04
0.22
0.24
0.09
0.04

8
-0.06
-0.02
-0.07
-0.10
-0.05
0.03
-0.20
0.00
-0.03
-0.18
0.04
-0.11
-0.18
-0.08
-0.24
-0.07
0.07
-0.13

9
-0.02
0.01
-0.03
-0.04
-0.05
-0.10
0.01
-0.01
-0.13
-0.11
-0.03
-0.13
-0.10
-0.07
-0.03
-0.11
0.07
-0.05

43

10
-0.01
-0.05
0.02
-0.06
0.12
0.16
-0.17
0.06
-0.12
0.54
-0.11
0.08
-0.17
0.17
0.05
-0.06
-0.23
0.15

11
0.06
-0.05
0.09
0.01
0.15
0.11
-0.09
-0.12
-0.10
0.43
-0.14
0.12
0.00
0.19
0.10
-0.05
-0.22
0.19

12
0.02
0.00
0.02
0.33
0.01
-0.08
0.20
0.02
0.01
-0.12
-0.15
0.41
0.36
-0.06
0.03
0.17
0.15
-0.01

13
0.01
-0.07
0.05
0.26
0.07
-0.05
0.19
-0.09
-0.03
0.07
0.09
0.23
0.50
0.09
0.12
0.16
0.02
0.11

14
0.10
-0.05
0.15
0.46
0.05
-0.23
0.17
-0.20
-0.01
-0.16
-0.01
0.28
0.39
0.06
0.17
0.13
0.06
0.08

15
0.07
0.00
0.09
0.05
0.04
0.00
0.05
-0.04
-0.03
0.14
0.16
-0.06
0.08
0.06
0.02
0.08
-0.17
0.02

16
0.04
-0.03
0.06
0.12
0.03
-0.01
0.06
-0.29
0.01
0.04
0.10
0.02
0.10
0.18
0.02

17
0.04
0.06
0.01
0.13
0.02
0.03
0.14
-0.08
-0.02
-0.08
-0.06
0.16
0.13
0.13
0.08
0.09

18
-0.08
-0.02
-0.09
0.04
-0.07
0.02
0.17
0.03
0.03
-0.22
-0.22
0.22
0.05
0.09
-0.16
-0.07
0.04

19
0.11
0.05
0.10
0.16
0.58
0.02
0.03
-0.10
-0.04
0.18
0.21
-0.02
0.11
0.07
0.03
0.08
0.02
-0.17

0.09
-0.05
0.01
0.08
-0.01
-0.15
All variables defined in Table A.1
Pearson correlations are above the diagonal, Spearman below.
Correlations in bold are significant at below the 5% level.

Table 4.12 shows the results of regressing the readability variables on Uncommon, Change
and previously identified determinants. In the multivariate case, Uncommon and Change are no
longer associated with Fog. However, Uncommon is significantly positively associated with W PS
and Length, consistent with Hypothesis 4. This result is also consistent with recent work that
suggests that the percentage of complex words is a misspecified measure of readability. Loughran
& McDonald (2011a) find that most complex words in financial reporting are very familiar and
readable to the average user, while words per sentence displays convergent validity with other
readability measures. Overall, my results are consistent with uncommon information being longer
and less readable than common disclosure. Since uncommon information is also informative,
these results suggest that information content, as well as managers’ incentives to hide information
Li (2008), can decrease readability.
Interestingly, Change is significantly associated with shorter MD&As and shorter sentences
within the MD&A. This result is consistent both with managers taking greater care in discussing
changes in the firm and with managers providing less informative (more common) disclosure when
a change in the firm occurs.

4.3

Summary of Results

Overall, my results are consistent with my hypotheses. I find that uncommon information both
reflects and provides firm-specific information to investors. However, the provision of that information imposes additional processing costs on investors, decreasing the readability of and delaying
the market response to the disclosure.

44

Table 4.12: Regression of Readability Variables on Uncommon
Dependent variable:
Fog
Uncommon
Change
Size
BT M
Special
ReturnVol
EarnVol
NBUSSEG
NGEOSEG
NumItems
DE
SEO
MA
Age
NumFirms
Year fixed effects
Industry fixed effects
Observations
Adjusted R2

0.278
(0.174)
−0.208
(0.185)
0.056∗∗∗
(0.017)
0.034
(0.032)
−0.137
(0.140)
0.651∗∗∗
(0.252)
0.995∗∗
(0.417)
−0.028
(0.037)
0.006
(0.031)
0.250
(0.432)
0.175∗∗∗
(0.051)
−0.064
(0.053)
0.072∗
(0.038)
−0.090∗∗∗
(0.033)
19.498∗∗∗
(2.303)

%Complex
−0.018
(0.208)
0.297
(0.292)
0.006
(0.019)
−0.018
(0.039)
0.137
(0.127)
−0.774∗∗
(0.381)
−0.983∗∗
(0.477)
−0.024
(0.063)
−0.033
(0.047)
1.664∗∗∗
(0.536)
0.024
(0.062)
−0.176∗∗
(0.076)
0.140∗∗∗
(0.046)
−0.035
(0.046)
19.441∗∗∗
(2.936)

W PS

Length

0.028∗∗
(0.011)
−0.032∗∗∗
(0.008)
0.005∗∗∗
(0.001)
0.004
(0.003)
−0.019
(0.013)
0.095∗∗∗
(0.018)
0.135∗∗∗
(0.030)
−0.002
(0.003)
0.002
(0.003)
−0.040
(0.032)
0.016∗∗∗
(0.004)
0.001
(0.004)
0.002
(0.003)
−0.008∗∗∗
(0.003)
3.380∗∗∗
(0.171)

0.332∗∗∗
(0.111)
−0.361∗∗∗
(0.062)
0.090∗∗∗
(0.006)
0.082∗∗∗
(0.013)
−0.285∗∗∗
(0.062)
0.874∗∗∗
(0.122)
0.531∗∗∗
(0.146)
0.110∗∗∗
(0.013)
0.065∗∗∗
(0.013)
0.888∗∗∗
(0.185)
0.040∗∗
(0.020)
0.038∗∗
(0.019)
0.054∗∗∗
(0.010)
−0.068∗∗∗
(0.015)
2.597∗∗
(1.053)

Yes
Yes

Yes
Yes

Yes
Yes

Yes
Yes

7,842
0.094

7,842
0.107

7,842
0.103

7,842
0.475

∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01

Note:

All variables defined in Table A.1
Standard errors in parentheses are clustered at the firm and year level.

45

CHAPTER 5
FUTURE WORK AND CONCLUSION

In this study, I provide evidence that uncommon information affects investors’ valuation models
but imposes significant processing costs on investors. In doing so, I contribute to the literature in
several ways. First, I describe one mechanism by which firm-specific components of firm value
are communicated between managers and investors. Second, I tie information processing costs directly to the information content of disclosure, as opposed to prior studies that focus on managerial
manipulation of word and sentence structure. Third, I provide evidence that uncommon information, and thus the firm-specific component of firm value, is incorporated slowly into market prices.
Finally, I introduce Latent Semantic Analysis to the accounting literature.
There are several limitations to this study. First, I assume that Latent Semantic Analysis, a
statistical procedure from computational linguistics, is able to represent text in a similar way as
it is processed by investors. There may be other methods that would provide a better measure of
uncommon information. Second, I only examine similarities across firms in the same industry year.
This is a naive, and possibly noisy, measure of uncommon information. To the extent that there
is a time-series component to uncommon information or a firm’s peers are in a different industry
classification, my measure may be misspecified. Third, my data collection procedure for firms’
MD&As is imperfect in that it omits a significant percentage of firm-years. Since my analysis in
Table 3.1 suggests that my sample differs significantly from Compustat, increasing my sample size
may improve tests of my hypotheses. Fourth, many of my variables, such as market beta or return
synchronicity, are measured over a wide time period. As a result, there may be events of which
I am unaware that bias my results. Fifth, since uncommon information is new to the literature,
correlated omitted variable bias is a concern. Sixth, my tests depend on how the market interacts
with uncommon information and thus may not capture the effect on individuals. Examining the
effect of uncommon information on individuals, such as analysts, could be a fruitful area for future

46

work.
There are many remaining research opportunities regarding uncommon information. First, a
significant unanswered question in this study is whether there is an opportunistic component to
uncommon information. Prior research has found that managers manipulate perceptions of firm
value in a variety of ways, including methods that delay the incorporation of information into
stock price, and are particularly apt to do so when it is difficult for investors to verify provided
information (Rogers & Stocken, 2005). While my results regarding firm-specific information suggest that at least some portion of uncommon information is informative, uncommon information is
hard to verify and thus represents an opportunity for managerial opportunism. Future research can
examine the sensitivity of uncommon information to management incentives.
Second, investors vary in their information sets and their sensitivity to information processing
costs (Bamber, 1987; Holthausen & Verrecchia, 1990; Morse et al., 1991; Bushee & Noe, 2000).
As such, they should respond differently to the information value and processing costs associated
with uncommon information. Future research can examine whether investors who are more sophisticated, have greater firm-specific knowledge or lower information processing costs respond
more quickly and more accurately to uncommon information than other investors.
Third, uncommon information provides a rich setting to examine investor opinion formation.
Future research can explore how uncommon information interacts with other characteristics of the
firm’s disclosures and information environment. For instance, while this study examines changes in
the MD&A separately from uncommon information, investors should react differently to changes
in the firm that increase or decrease uncommon information. Understanding that response can
help researchers to understand how investors respond to changes in a firm’s economic environment
more broadly.
Fourth and finally, this study uses Latent Semantic Analysis (LSA) to identify sources of variation in the text of financial reports. Whereas prior research uses simple word counts to examine
qualitative disclosures, LSA allows the researcher to observe the underlying meaning of text. This
ability is potentially very powerful for identifying how managers talk about firm value. Future

47

studies can apply LSA to a variety of settings and research questions.

48

APPENDICES

49

A

Variables
Table A.1: Variable List

Name

Description

Dependent Variables:

CAR−1,1

The absolute cumulative abnormal return during the 10-K
filing window (the day before to the day after). Abnormal
returns are calculated using a market model estimated for
the 255 days ending 63 days prior to the 10-K filing date.

CAR2,60

The absolute cumulative abnormal return during the postfiling window (2 to 60 days after the filing).

%Complex

The percentage of total words in the MD&A that are three
syllables or more in length.

Fog

Words per sentence (W PS) plus percent complex words
(%Complex), multiplied by 0.4.

Length

The log of the total number of words in the firm’s MD&A.

PCAR−1,1

The absolute cumulative abnormal return during the 3-day
window 60 days prior to the earnings announcement (the
day before to the day after). Abnormal returns are calculated using a market model estimated for the 255 days ending 63 days prior to the selected date.

PCAR2,60

The absolute cumulative abnormal return during the 58 days
leading up to the earnings announcement window (2 to 60
days prior to the announcement).
Continued on next page

50

Table A.1 (cont’d)
Name
ReturnSync

Description
R2
, where
1 − R2
R2 is the the R-squared from a regression of weekly firm reFollowing Piotroski & Roulstone (2004), log

turns on concurrent and lagged value-weighted market and
industry returns over the fiscal year.
∆ReturnSync

ReturnSync minus its own value in the previous year.

Uncommon

The mean cosine dissimilarity between a firm’s MD&A and
the MD&As of the four closest firms in its industry year.
Cosine dissimilarity is measured in a vector space model
after performing Latent Semantic Analysis on all reports in
the industry year and selecting the top N dimensions, where
N is equal to 10% of the total number of firms in that industry year.

∆Uncommon

Uncommon minus its own value in the previous year.

W PS

The average number of words per sentence in the MD&A.

Independent Variables:

Age

The number of years since the firm first appeared in CRSP.

AnaFollow

The number of analysts in the IBES database issuing at least
one forecast for the firm during the fiscal year.

Beta

Market beta calculated over the 60 months prior to the 10-K
filing.

BT M

The book value per share divided by the price per share,
calculated at the end of the fiscal year.
Continued on next page

51

Table A.1 (cont’d)
Name
Change

Description
The MD&A modification score from Brown & Tucker
(2011), obtained from Brown (2013).

DE

An indicator variable equal to one if the firm is incorporated
in Delaware.

RevConc

The log of the Herfindahl index, calculated by firm as
the proportion of sales per segment, squared and summed
across the segments.

EA−1,1

The absolute cumulative abnormal return in the earnings
announcement window (the day before to the day after).
Abnormal returns are calculated using a market model estimated for the 255 days ending 63 days prior to the earnings
announcement date.

EarnSync

R2
, where
1 − R2
R2 is the the R-squared from a regression of quarterly return
Following Piotroski & Roulstone (2004), log

on assets on concurrent and lagged value-weighted industry
return on assets over the last 12 quarters.
EarnVol

The standard deviation of earnings over the last five years.

FileLate

An indicator variable equal to one if the 10-K was filed 90
or more days after the end of the fiscal period.

IndConc

The log of the Herfindahl index, calculated by industry
as the proportion of industry sales per firm, squared and
summed across the industry.

Instown

The percentage of a firm’s equity held by institutional owners, as reported in the Thomson-Reuters database .
Continued on next page

52

Table A.1 (cont’d)
Name
Leverage

Description
Total debt divided by total assets, calculated at the end of
the fiscal year.

MA

An indicator variable equal to one if the firm had a merger
or acquisition during the year.

NBUSSEG

The log of the firm’s number of operating segments.

NGEOSEG

The log of the firm’s number of geographic segments.

NumItems

The log of the number of non-missing line items in Compustat.

NumFirms

The number of firms in the industry-year.

Price

The log of the stock price at the end of the fiscal year.

PVolume−1,1

The log of the proportion of total shares traded over the 3day window 60 days prior to the earnings announcement

PVolume2,60

The log of the proportion of total shares traded over the 58
days leading up to the earnings announcement window (2
to 60 days prior to the announcement).

Reg

An indicator variable set equal to one of the firm is in a
regulated industry (two-digit SIC code is 49 or 62).

ReturnVol

The standard deviation of monthly returns over the fiscal
year.

SEO

An indicator variable equal to one if the firm has had a seasoned equity offering during the year.

Size

The log of the market value of equity at the end of the fiscal
year.

Special

The amount of special items scaled by total assets.
Continued on next page

53

Table A.1 (cont’d)
Name
UE

Description
IBES reported actual earnings for the year minus the mean
estimate from the last consensus period prior to the earnings
announcement, scaled by the price at the end of the fiscal
year.

Volume−1,1

The log of the proportion of total shares traded over the
three days surrounding the 10-K filing date.

Volume2,60

The log of the proportion of total shares traded over the 58
days surrounding the 10-K filing date.

VolumeEA−1,1

The log of the proportion of total shares traded over the
three days surrounding the earnings announcement date.

54

B

Excerpts from Most and Least Common MD&As

NCI Information Systems, Inc.
NCI has the most common MD&A in the Software Industry in 2007. In the business description
portion of their MD&A, they say:
We are a provider of information technology (IT) and professional services and solutions to federal government agencies. Our technology and industry expertise enables
us to provide a full spectrum of services and solutions that assist our clients in achieving their program goals. We deliver a wide range of complex services and solutions
by leveraging our skills across seven core service offerings:
• network engineering;
• information assurance;
• systems engineering and integration;
• enterprise systems management;
• engineering and logistics;
• medical transformation/health IT; and
• distance learning and training.
We generate substantially all of our revenue from federal government contracts. We
report operating results and financial data as one operating segment. Revenue from our
contracts and task orders is generally linked to trends in federal government spending
by defense, intelligence and federal civilian agencies.
We believe that our contract base is well diversified. As of December 31, 2007, we
had approximately 200 active contracts and 600 task orders. As of December 31,
2007, our total contract backlog was approximately $756 million of which approximately $189 million was funded. We define backlog as our estimate of the remaining
55

future revenue from existing signed contracts over the remaining base contract performance period and from the option periods of those contracts, assuming the exercise
of all related options. Our backlog does not include any estimate of future potential
delivery orders that might be awarded under our GWAC or other multiple award contract vehicles. We define funded backlog as the portion of backlog for which funding
currently is appropriated and obligated to us under a contract or other authorization for
payment signed by an authorized purchasing agency, less the amount of revenue we
have previously recognized. Our funded backlog does not represent the full potential
value of our contracts, as Congress often appropriates funds for a particular program
or agency on a quarterly or yearly basis, even though the contract may provide for
the provision of services over a number of years. We define unfunded backlog as the
total backlog less the funded backlog. Unfunded backlog includes values for contract
options that have been priced but not yet funded.

Borland Software Corporation
Borland has the least common MD&A in the Software Industry in 2007. In the business description
portion of their MD&A, they say:
Borland is a leading vendor of Open Application Lifecycle Management solutions, or
ALM, which represents the segment of the ALM market in which vendors’ solutions
are flexible enough to support a customer’s specific processes, tools and platforms.
Open ALM is a new, customer-centric approach to helping IT organizations transform
software delivery into a managed, efficient and predictable business process. We offer
a combination of software products as well as consulting and education services to help
our customers better manage the growing complexity of software development. Our
goal is to provide customers with a foundation which will allow them to consistently
deliver software on-time, on-budget and with increased business value.

56

Borland’s solutions address five critical ALM processes: Project & Portfolio Management, Requirements Definition & Management, Lifecycle Quality Management,
Model Driven Development and Software Change Management. Each solution can
play an important role in helping enterprises manage the complexity of software development and delivery, by providing business, development and operational teams
with increased visibility and control over all phases of the software delivery lifecycle.
We believe this is especially crucial for large enterprises working within heterogeneous
and distributed environments.
We have been evolving our business and strategy in recent years in response to the
many changes occurring in the software industry and, specifically in our market. In a
March 2005 study, IDC forecasted the ALM market to grow to $3.3 billion in 2009,
achieving a 9.2% compound annual growth rate between 2004 and 2009. In order to
capitalize on the ALM market growth, we have made changes to our overall product
portfolio, our worldwide services organization, our R&D investments, as well as our
global sales and marketing models to reflect our Open ALM vision and product strategy. As part of this transformation, we have shifted our focus from selling individual
stand-alone products to selling more multi-product, enterprise-class solutions.
Effective January 1, 2007, consistent with how we manage our business, we changed
from reporting one segment to reporting two segments: Enterprise and CodeGear.
A summary of the types of products and services provided by the Enterprise and
CodeGear segments is provided below.
Enterprise. Our Enterprise segment focuses on Open Application Lifecycle Management solutions, or ALM, which includes a combination of software products as well as
consulting and education services to help our customers better manage their software
development projects. Our ALM portfolio includes products and services for project
and portfolio management, requirements definition and management, lifecycle quality management, software configuration and change management and modeling. The
57

Enterprise segment also includes our Deployment Product Group, or DPG, products.
CodeGear. Our CodeGear segment focuses on developing tools for individual developers and currently offers a number of Integrated Developer Environment, or IDE,
and database products for Java, .NET and Windows development. CodeGear products
include Delphi, Delphi for PHP, C++Builder, C#Builder, JBuilder, Turbotm and Interbase. CodeGear also provides worldwide developer support and education services.

58

BIBLIOGRAPHY

59

BIBLIOGRAPHY

Aboody, David, Reuven Lehavy & Brett Trueman. 2010. Limited attention and the earnings announcement returns of past stock market winners. Review of Accounting Studies 15(2). 317–344.
doi:10.1007/s11142-009-9104-9.
Arya, Anil & Brian Mittendorf. 2005. Using disclosure to influence herd behavior and alter competition. Journal of Accounting and Economics 40(1–3). 231–246. doi:10.1016/j.jacceco.2005.
07.001.
Baddeley, Alan D., Neil Thomson & Mary Buchanan. 1975. Word length and the structure of
short-term memory. Journal of Verbal Learning and Verbal Behavior 14(6). 575–589. doi:
10.1016/S0022-5371(75)80045-4.
Ball, Christopher, Gerard Hoberg & Vojislav Maksimovic. 2013. Disclosure informativeness and
the tradeoff hypothesis: A text-based analysis. SSRN Scholarly Paper ID 2260371 Social Science Research Network Rochester, NY.
Ball, Ray & Philip Brown. 1968. An empirical evaluation of accounting income numbers. Journal
of Accounting Research 6(2). 159–178. doi:10.2307/2490232.
Ball, Ray & Lakshmanan Shivakumar. 2008. How much new information is there in earnings?
Journal of Accounting Research 46(5). 975–1016. doi:10.1111/j.1475-679X.2008.00299.x.
Bamber, Linda Smith. 1987. Unexpected earnings, firm size, and trading volume around quarterly
earnings announcements. The Accounting Review 62(3). 510–532. doi:10.2307/247574.
Bamber, Linda Smith, Orie E. Barron & Douglas E. Stevens. 2011. Trading volume around earnings announcements and other financial reports: Theory, research design, empirical evidence,
and directions for future research. Contemporary Accounting Research 28(2). 431–471. doi:
10.1111/j.1911-3846.2010.01061.x.
Barron, Orie E., Charles O. Kile & Terrence B. O’Keefe. 1999. MD&A quality as measured by
the SEC and analysts’ earnings forecasts. Contemporary Accounting Research 16(1). 75–109.
doi:10.1111/j.1911-3846.1999.tb00575.x.
Barry, Christopher B. & Stephen J. Brown. 1985. Differential information and security market
equilibrium. Journal of Financial & Quantitative Analysis 20(4). 407–422.
Becker, Curtis A. 1976. Allocation of attention during visual word recognition. Journal of Experimental Psychology: Human Perception and Performance 2(4). 556–566. doi:http://dx.doi.org.
proxy1.cl.msu.edu/10.1037/0096-1523.2.4.556.
Becker, Curtis A. 1980. Semantic context effects in visual word recognition: An analysis of
semantic strategies. Memory & Cognition 8(6). 493–512. doi:10.3758/BF03213769.
Bellman, Richard Ernest. 2003. Dynamic programming. Courier Dover Publications.
60

Bhojraj, Sanjeev & Charles M. C. Lee. 2002. Who is my peer? a valuation-based approach
to the selection of comparable firms. Journal of Accounting Research 40(2). 407–439. doi:
10.1111/1475-679X.00054.
Blei, David M., Andrew Y. Ng & Michael I. Jordan. 2003. Latent dirichlet allocation. J. Mach.
Learn. Res. 3. 993–1022.
Bloomfield, Robert. 2008. Discussion of "Annual report readability, current earnings, and earnings
persistence". Journal of Accounting and Economics 45(2-3). 248–252. doi:10.1016/j.jacceco.
2008.04.002.
Bloomfield, Robert J. 2002. The “Incomplete Revelation Hypothesis" and financial reporting.
Accounting Horizons 16(3). 233–243.
Bloomfield, Robert J. 2012. A pragmatic approach to more efficient corporate disclosure. Accounting Horizons 26(2). 357–370.
Bloomfield, Robert J., Frank D. Hodge, Patrick E. Hopkins & Kristina M. Rennekamp. 2010. Does
enhanced disaggregation and cohesive classification of financial information help credit analysts
identify firms’ operating structures? SSRN eLibrary .
Branch, William A. & George W. Evans. 2006. Intrinsic heterogeneity in expectation formation.
Journal of Economic Theory 127(1). 264–295. doi:10.1016/j.jet.2004.11.005.
Branch, William A. & George W. Evans. 2010. Asset return dynamics and learning. Review of
Financial Studies 23(4). 1651–1680. doi:10.1093/rfs/hhp112.
Brown, Stephen V. 2013. Data - Stephen V. Brown.
Brown, Stephen V. & Jennifer Wu Tucker. 2011. Large-sample evidence on firms’ year-overyear MD&A modifications. Journal of Accounting Research 49(2). 309–346. doi:10.1111/j.
1475-679X.2010.00396.x.
Bushee, Brian J. & Christopher F. Noe. 2000. Corporate disclosure practices, institutional investors, and stock return volatility. Journal of Accounting Research 38. 171–202. doi:
10.2307/2672914.
Clarkson, Pete, Jose Guedes & Rex Thompson. 1996. On the diversification, observability, and
measurement of estimation risk. Journal of Financial & Quantitative Analysis 31(1). 69–84.
Clarkson, Peter M., Jennifer L. Kao & Gordon D. Richardson. 1999. Evidence that management
discussion and analysis (MD&A) is a part of a firm’s overall disclosure package. Contemporary
Accounting Research 16(1). 111–134. doi:10.1111/j.1911-3846.1999.tb00576.x.
Coles, Jeffrey L., Uri Loewenstein & Jose Suay. 1995. On equilibrium pricing under parameter
uncertainty. The Journal of Financial and Quantitative Analysis 30(3). 347–364. doi:10.2307/
2331345.
Core, John E., Wayne R. Guay & Rodrigo Verdi. 2008. Is accruals quality a priced risk factor?
Journal of Accounting and Economics 46(1). 2–22. doi:10.1016/j.jacceco.2007.08.001.
61

Davis, Angela K. & Isho Tama-Sweet. 2012. Managers’ use of language across alternative disclosure outlets: Earnings press releases versus MD&A. Contemporary Accounting Research 29(3).
804–837. doi:10.1111/j.1911-3846.2011.01125.x.
DeFranco, Gus, S.P. Kothari & Rodrigo S. Verdi. 2011. The benefits of financial statement comparability. Journal of Accounting Research 49(4). 895–931. doi:10.1111/j.1475-679X.2011.
00415.x.
Dellavigna, Stefano & Joshua M. Pollet. 2009. Investor inattention and friday earnings announcements. The Journal of Finance 64(2). 709–749. doi:10.1111/j.1540-6261.2009.01447.x.
Drake, Michael S., Darren T. Roulstone & Jacob R. Thornock. 2012. The informativeness of stale
financial disclosures. SSRN Scholarly Paper ID 2083812 Social Science Research Network
Rochester, NY.
Durnev, Artyom, Randall Morck, Bernard Yeung & Paul Zarowin. 2003. Does greater firm-specific
return variation mean more or less informed stock pricing? Journal of Accounting Research
41(5). 797–836.
Financial Accounting Standards Board. 2010. Conceptual framework for financial reporting. Statement of Financial Accounting Concepts No. 8.
Forster, Kenneth I. & Susan M. Chambers. 1973. Lexical access and naming time. Journal of
Verbal Learning and Verbal Behavior 12(6). 627–635. doi:10.1016/S0022-5371(73)80042-8.
Forster, K.I. 1981. Frequency blocking and lexical access: One mental lexicon or two? Journal of
Verbal Learning and Verbal Behavior 20(2). 190–203. doi:10.1016/S0022-5371(81)90373-X.
Francis, Jennifer, Ryan LaFond, Per Olsson & Katherine Schipper. 2005. The market pricing of
accruals quality. Journal of Accounting and Economics 39(2). 295–327. doi:10.1016/j.jacceco.
2004.06.003.
Francis, Jennifer, Dhananjay Nanda & Per Olsson. 2008. Voluntary disclosure, earnings quality,
and cost of capital. Journal of Accounting Research 46(1). 53–99. doi:10.1111/j.1475-679X.
2008.00267.x.
Glanzer, Murray & S.L. Ehrenreich. 1979. Structure and search of the internal lexicon. Journal of
Verbal Learning and Verbal Behavior 18(4). 381–398. doi:10.1016/S0022-5371(79)90210-X.
Gong, Guojin, Laura Yue Li & Ling Zhou. 2013. Earnings non-synchronicity and voluntary disclosure. Contemporary Accounting Research 1560–1589. doi:10.1111/1911-3846.12007.
Goodman, George O., James L. McClelland & Raymond W. Gibbs. 1981. The role of syntactic
context in word recognition. Memory & Cognition 9(6). 580–586. doi:10.3758/BF03202352.
Grossman, Sanford J. & Joseph E. Stiglitz. 1980. On the impossibility of informationally efficient
markets. The American Economic Review 70(3). 393–408.

62

Halgren, Eric, Rupali P. Dhond, Natalie Christensen, Cyma Van Petten, Ksenija Marinkovic, Jeffrey D. Lewine & Anders M. Dale. 2002. N400-like magnetoencephalography responses modulated by semantic context, word frequency, and lexical class in sentences. NeuroImage 17(3).
1101–1116. doi:10.1006/nimg.2002.1268.
Hall, John F. 1954. Learning as a function of word-frequency. The American Journal of Psychology
67(1). 138–140. doi:10.2307/1418080.
Handa, Puneet & Scott C. Linn. 1993. Arbitrage pricing with estimation risk. Journal of Financial
& Quantitative Analysis 28(1). 81–100.
Healy, Paul M & Krishna G Palepu. 2001. Information asymmetry, corporate disclosure, and the
capital markets: A review of the empirical disclosure literature. Journal of Accounting and
Economics 31(1–3). 405–440. doi:10.1016/S0165-4101(01)00018-0.
Heneman, Robert L. 1986. The relationship between supervisory ratings and results-oriented
measures of performance: A meta-analysis. Personnel Psychology 39(4). 811–826. doi:
10.1111/j.1744-6570.1986.tb00596.x.
Hirshleifer, David, Sonya Seongyeon Lim & Siew Hong Teoh. 2009. Driven to distraction: Extraneous events and underreaction to earnings news. The Journal of Finance 64(5). 2289–2325.
doi:10.1111/j.1540-6261.2009.01501.x.
Hodder, Leslie, Patrick E. Hopkins & David A. Wood. 2008. The effects of financial statement and
informational complexity on analysts, cash flow forecasts. Accounting Review 83(4). 915–956.
Holthausen, Robert W. & Robert E. Verrecchia. 1990. The effect of informedness and consensus
on price and volume behavior. The Accounting Review 65(1). 191–208. doi:10.2307/247883.
Hong, Harrison, Jeremy C. Stein & Jialin Yu. 2007. Simple forecasts and paradigm shifts. The
Journal of Finance 62(3). 1207–1242. doi:10.1111/j.1540-6261.2007.01234.x.
Hopkins, Patrick E. 1996. The effect of financial statement classification of hybrid financial instruments on financial analysts’ stock price judgments. Journal of Accounting Research 34. 33–50.
doi:10.2307/2491424.
Huang, Allen, Reuven Lehavy, Amy Y. Zang & Rong Zheng. 2014. A thematic analysis of analyst
information discovery and information interpretation roles. SSRN Scholarly Paper ID 2409482
Social Science Research Network Rochester, NY.
Hulme, Charles, Steven Roodenrys, Richard Schweickert, D. A, Sarah Martin & George Stuart.
1997. Word-frequency effects on short-term memory tasks: Evidence for a redintegration process in immediate serial recall. Journal of Experimental Psychology: Learning, Memory, and
Cognition 23(5). 1217–1232. doi:10.1037/0278-7393.23.5.1217.
Karelaia, Natalia & Robin M. Hogarth. 2008. "Determinants of linear judgment: A meta-analysis
of lens model studies": Correction. Psychological Bulletin 134(5). 741. doi:http://dx.doi.org.
proxy1.cl.msu.edu/10.1037/a0013550.

63

Khan, Arshad M. & V. Manopichetwattana. 1989. Innovative and noninnovative small firms: Types
and characteristics. Management Science 35(5). 597–606.
Kim, Oliver & Robert E. Verrecchia. 1991. Trading volume and price reactions to public announcements. Journal of Accounting Research 29(2). 302–321. doi:10.2307/2491051.
Kim, Oliver & Robert E. Verrecchia. 1994. Market liquidity and volume around earnings
announcements. Journal of Accounting and Economics 17(1–2). 41–67. doi:10.1016/
0165-4101(94)90004-3.
Kintsch, Walter. 1988. The role of knowledge in discourse comprehension: A constructionintegration model. Psychological Review 95(2). 163–182. doi:http://dx.doi.org.proxy1.cl.msu.
edu/10.1037/0033-295X.95.2.163.
Lambert, Richard, Christian Leuz & Robert E. Verrecchia. 2007. Accounting information, disclosure, and the cost of capital. Journal of Accounting Research 45(2). 385–420. doi:
10.1111/j.1475-679X.2007.00238.x.
Landauer, Thomas K. & Susan T. Dumais. 1997. A solution to Plato’s problem: The latent semantic analysis theory of acquisition, induction, and representation of knowledge. Psychological
Review 104(2). 211–240. doi:10.1037/0033-295X.104.2.211.
Landauer, Thomas K, Peter W. Foltz & Darrell Laham. 1998. An introduction to latent semantic
analysis. Discourse Processes 25(2-3). 259–284. doi:10.1080/01638539809545028.
Lee, Yen-Jung. 2012. The effect of quarterly report readability on information efficiency of stock
prices. Contemporary Accounting Research 29(4). 1137–1170. doi:10.1111/j.1911-3846.2011.
01152.x.
Lee, Yen Jung, Kathy R. Petroni & Min Shen. 2006. Cherry picking, disclosure quality, and
comprehensive income reporting choices: The case of property-liability insurers. Contemporary
Accounting Research 23(3). 655–700. doi:10.1506/5QB8-PBQY-Y86L-DRYL.
Lehavy, Reuven, Feng Li & Kenneth Merkley. 2011. The effect of annual report readability on
analyst following and the properties of their earnings forecasts. Accounting Review 86(3). 1087–
1115.
Lewellen, Jonathan & Jay Shanken. 2002. Learning, asset-pricing tests, and market efficiency. The
Journal of Finance 57(3). 1113–1145. doi:10.1111/1540-6261.00456.
Li, Feng. 2006. Do stock market investors understand the risk sentiment of corporate annual
reports? SSRN eLibrary .
Li, Feng. 2008. Annual report readability, current earnings, and earnings persistence. Journal of
Accounting and Economics 45(2–3). 221–247. doi:10.1016/j.jacceco.2008.02.003.
Li, Feng. 2010. The information content of forward-looking statements in corporate filings: A
naive bayesian machine learning approach. Journal of Accounting Research 48(5). 1049–1102.
doi:10.1111/j.1475-679X.2010.00382.x.

64

Libby, Theresa, Steven E. Salterio & Alan Webb. 2004. The balanced scorecard: The effects of
assurance and process accountability on managerial judgment. The Accounting Review 79(4).
1075–1094.
Lipe, Marlys Gascho & Steven E. Salterio. 2000. The balanced scorecard: Judgmental effects of
common and unique performance measures. The Accounting Review 75(3). 283–298.
Loughran, Tim & Bill McDonald. 2011a. Measuring readability in financial disclosures. SSRN
Scholarly Paper ID 1920411 Social Science Research Network Rochester, NY.
Loughran, Tim & Bill McDonald. 2011b. When is a liability not a liability? textual analysis,
dictionaries, and 10-ks. The Journal of Finance 66(1). 35–65. doi:10.1111/j.1540-6261.2010.
01625.x.
Loughran, Tim & Bill McDonald. 2013. IPO first-day returns, offer price revisions, volatility, and
form s-1 language. Journal of Financial Economics 109(2). 307–326. doi:10.1016/j.jfineco.
2013.02.017.
Miller, Brian P. 2010. The effects of reporting complexity on small and large investor trading.
Accounting Review 85(6). 2107–2143. doi:10.2308/accr.00000001.
Morse, Dale, Jens Stephan & Earl K. Stice. 1991. Earnings announcements and the convergence
(or divergence) of beliefs. The Accounting Review 66(2). 376–388.
Ohlson, James A. 1995. Earnings, book values, and dividends in equity valuation. Contemporary
Accounting Research 11(2). 661–687.
Payne, John W., James R. Bettman & Eric J. Johnson. 1993. The adaptive decision maker. Cambridge University Press.
Peng, Lin & Wei Xiong. 2006. Investor attention, overconfidence and category learning. Journal
of Financial Economics 80(3). 563–602. doi:10.1016/j.jfineco.2005.05.003.
Pierce, John R. 1980. An introduction to information theory: Symbols, signals and noise. New
York, NY: Dover Publications, Inc. 2nd edn.
Piotroski, Joseph D. & Darren T. Roulstone. 2004. The influence of analysts, institutional investors,
and insiders on the incorporation of market, industry, and firm-specific information into stock
prices. Accounting Review 79(4). 1119–1151.
Postman, Leo. 1970. Effects of word frequency on acquisition and retention under conditions
of free-recall learning. Quarterly Journal of Experimental Psychology 22(2). 185–195. doi:
10.1080/00335557043000113.
Rayner, Keith & Susan A. Duffy. 1986. Lexical complexity and fixation times in reading: Effects of
word frequency, verb complexity, and lexical ambiguity. Memory & Cognition 14(3). 191–201.
doi:10.3758/BF03197692.
Rennekamp, Kristina. 2012a. Processing fluency and investors’ reactions to disclosure readability.
Journal of Accounting Research 50(5). 1319–1354. doi:10.1111/j.1475-679X.2012.00460.x.
65

Rennekamp, Kristina M. 2012b. The influence of performance and reporting goals on managers’
choice of reporting complexity in disclosures. SSRN eLibrary .
Richardson, Scott A., Richard G. Sloan, Mark T. Soliman & Irem Tuna. 2005. Accrual reliability,
earnings persistence and stock prices. Journal of Accounting and Economics 39(3). 437–485.
doi:10.1016/j.jacceco.2005.04.005.
Rogers, Jonathan L. & Phillip C. Stocken. 2005. Credibility of management forecasts. Accounting
Review 80(4). 1233–1260.
Scarborough, Don L., Charles Cortese & Hollis S. Scarborough. 1977. Frequency and repetition
effects in lexical memory. Journal of Experimental Psychology: Human Perception and Performance 3(1). 1–17. doi:10.1037/0096-1523.3.1.1.
Schvaneveldt, Roger W., David E. Meyer & Curtis A. Becker. 1976. Lexical ambiguity, semantic
context, and visual word recognition. Journal of Experimental Psychology: Human Perception
and Performance 2(2). 243–256. doi:10.1037/0096-1523.2.2.243.
Securities and Exchange Commission. 2003. Interpretation: Commission guidance regarding management’s discussion and analysis of financial condition and results of operations.
Shannon, Claude & Warren Weaver. 1949. The mathematical theory of communication. University
of Illinois Press 19(7). 1.
Shields, Michael D. 1980. Some effects on information load on search patterns used to analyze performance reports. Accounting, Organizations and Society 5(4). 429–442. doi:
10.1016/0361-3682(80)90041-0.
Simon, Herbert A. 1955. A behavioral model of rational choice. Quarterly Journal of Economics
69(1). 99–118.
Sloan, Richard G. 1996. Do stock prices fully reflect information in accruals and cash flows about
future earnings? The Accounting Review 71(3). 289–315. doi:10.2307/248290.
Slovic, Paul & Douglas MacPhillamy. 1974. Dimensional commensurability and cue utilization
in comparative judgment. Organizational Behavior and Human Performance 11(2). 172–194.
doi:10.1016/0030-5073(74)90013-0.
Sowa, John F. 1984. Conceptual structures: information processing in mind and machine The
Systems programming series. Reading, Mass: Addison-Wesley.
Tan, Hun-Tong, Elaine Ying Wang & Bo Zhou. 2013. When the use of positive language backfires:
The joint effect of tone, readability, and investor sophistication on earnings judgments. Journal
of Accounting Research 52(1). 273–302. doi:10.1111/1475-679X.12039.
Tetlock, Paul C. 2011. All the news that’s fit to reprint: Do investors react to stale information?
Review of Financial Studies 24(5). 1481–1512. doi:10.1093/rfs/hhq141.

66

Tweedy, James R., Robert H. Lapinski & Roger W. Schvaneveldt. 1977. Semantic-context effects
on word recognition: Influence of varying the proportion of items presented in an appropriate
context. Memory & Cognition 5(1). 84–89. doi:10.3758/BF03209197.
Verrecchia, Robert E. 1983. Discretionary disclosure. Journal of Accounting and Economics 5.
179–194. doi:10.1016/0165-4101(83)90011-3.
Xie, Hong. 2001. The mispricing of abnormal accruals. Accounting Review 76(3). 357.
Zhang, Shi & Arthur B. Markman. 2001. Processing product unique features: Alignability and
involvement in preference construction. Journal of Consumer Psychology 11(1). 13–27. doi:
10.1207/S15327663JCP1101_2.
Zipf, George Kingsley. 1949. Human behavior and the principle of least effort: an introduction to
human ecology. Cambridge, Mass: Addison-Wesley Press.

67