Research " Earnings Warnings: Market Reaction and Management Motivation " ppt

Specifically, this study first investigates whether self-selection bias exists in a firm’s warning choice and if so, whether the warning effect i.e., a differential market reaction assoc

Earnings Warnings:

Market Reaction and Management Motivation

by Somchai Supattarakul, B.B.A., M.B.A., M.P.A

Dissertation

Presented to the Faculty of the Graduate School of

The University of Texas at Austin

in Partial Fulfillment

of the Requirements for the Degree of

Doctoral of Philosophy

The University of Texas at Austin

May, 2003

UMI Number: 3116199

UMI Microform 3116199 Copyright 2004 by ProQuest Information and Learning Company All rights reserved This microform edition is protected against unauthorized copying under Title 17, United States Code

ProQuest Information and Learning Company

300 North Zeeb Road

PO Box 1346 Ann Arbor, MI 48106-1346

Dedication

To my parents

Acknowledgements

I would like to express my most sincere gratitude to my committee members: Rowland Atiase (Chair), Robert Freeman, Tom Shively, Laura Starks, and Senyo Tse Rowland has been an invaluable mentor to me, and has been generous with his time and advice throughout my doctoral studies In addition, I wish to acknowledge with sincere thanks to many helpful suggestions of Peggy Weber Also, I would like to thank First Call Corporation, and Steve Sommers in particular, for the corporate earnings guidance data I appreciate the financial support of Ministry of University Affairs (Thailand) and Faculty of Commerce and Accountancy, Thammasat University (Thailand)

I would like to thank my colleagues and friends at Thammasat University (Thailand) for their support and encouragement Finally, my deepest thanks go to

my parents, my sister and brothers, for their unconditional love and support They have always been my source of strength Without them this would have not been possible

Earnings Warnings:

Market Reaction and Management Motivation

Publication No. _

Somchai Supattarakul, Ph.D

The University of Texas at Austin, 2003

Supervisor: Rowland K Atiase

This dissertation provides empirical evidence on the market reaction to earnings warnings as well as management’s motivation to issue earnings warnings Specifically, this study first investigates whether self-selection bias exists in a firm’s warning choice and if so, whether the warning effect (i.e., a differential market reaction associated with earnings news between warning and no-warning scenarios) is positive (negative) for good (bad) news warnings after controlling for potential self-selection bias I find that self-selection does exist in

a firm’s warning choice and it creates a downward bias in the warning effect I also find that the warning effect after controlling for self-section bias, on average,

is positive (negative) for good (bad) news warnings suggesting that empirical evidence in Kasznik and Lev [1995] and Atiase, Supattarakul, Tse [2003] is robust after controlling for self-selection bias More importantly, this study

investigates whether and how the warning effect affects a firm’s warning choice (i.e., to warn or not to warn) I find that a firm’s tendency to warn is positively associated with the warning effect after controlling for other management motives

to issue earnings warnings, i.e., litigation concerns, reputation concerns, and information asymmetry consequence concerns, suggesting that the warning effect itself provides management with an economic motivation to issue earnings warnings

Table of Contents

List of Tables x

List of Figure xii

Chapter 1: Introduction 1

Chapter 2: Prior Studies and Hypotheses 10

2.1 Differential market reaction (The warning effect) 10

2.2 Self-selection in a firm’s warning choice 12

2.3 Management motives to issue earnings warnings 15

Chapter 3: Model Specification and Estimation Procedures 20

3.1 Model specification 20

3.2 Limited dependent variable or self-selectivity problem 22

3.3 Lee [1978]’s approach 23

3.4 Methodological problems in Shu [2001] 27

Chapter 4: Sample Design and Variable Definitions 30

4.1 Sample selection criteria 30

4.2 Sample Description 31

4.3 Variable Definitions 32

4.3.1 Market reaction associated with earnings news (MRW and MRN) 32

4.3.2 Determinants of market reaction associated with earnings news (Z*) 33

4.3.3 Warning choice (WARN) 37

4.3.4 Management motives to issue earnings warnings (S*) 37

4.4 Sample descriptive statistics 41

Chapter 5: Empirical Tests and Results 43

5.1 Self-selection and its impacts on the warning effect 43

5.2 A firm’s tendency to warn and the warning effect 57

Chapter 6: Summary and Conclusions 62

6.1 Summary 62

6.2 Contributions and future research 63

Figure and Tables 66

References 97

Vita… 101

List of Tables

Table 1 – Reconciliation of Sample Data 68 Table 2 – Distribution of Earnings Warnings by Year and Quarter 69 Table 3 – Sample Descriptive Statistics 70 Table 4 – Correlations of Market Reaction Associated with Earnings News

(MR), Unexpected Earnings (UE) and Analyst Forecast

Revisions (AFR) 72 Table 5 – Results of Probit Maximum Likelihood Estimation of Warning

Choice Model to Obtain Parameters to Calculate “Inverse Mills Ratio” 73 Table 6 – Results of OLS Estimation of Market Reaction Model under

Warning Scenario – Controlling for Self-selection Bias 76 Table 7 – Results of OLS Estimation of Market Reaction Model under

No-Warning Scenario – Controlling for Self-selection Bias 79 Table 8 – Distribution of the Warning Effect after Controlling for

Self-selection Bias (∆MˆR) 82 Table 9 – Results of OLS Estimation of Market Reaction Model under

Warning Scenario – without Controlling for Self-selection Bias 84 Table 10 – Results of OLS Estimation of Market Reaction Model under

No-Warning Scenario – without Controlling for Self-selection

Bias 87 Table 11 – Distribution of the Warning Effect without Controlling for

Self-selection Bias (∆Mˆ ′R ) 90

Table 12 – Distribution of Self-selection Bias (SSˆB) in the Warning Effect 92 Table 13 – Results of Probit Maximum Likelihood Estimation of Warning

Choice Model 94

List of Figure

Figure 1 – Timeline of Events 67

Chapter 1: Introduction

This dissertation provides empirical evidence on the market reaction to earnings warnings and management’s motivation to issue earnings warnings Earnings warnings are any earnings-related management voluntary disclosures made prior to the earnings announcement date.1 Firms use earnings warnings to provide timely information to their shareholders and investors as well as financial analysts regarding their expected current period performance prior to the earnings announcement date (Ip [1997], McLean [2001] and Stone [2002]) Kasznik and Lev [1995; hereafter KL] and Atiase, Supattarakul, and Tse [2003; hereafter AST] find a differential market reaction to earnings news between warning and no-

warning scenarios (i.e., “the warning effect”).2 Specifically, AST document that the warning effect is positive for good news warnings while KL and AST find that the warning effect is negative for bad news warnings.3 Both KL and AST document that the majority of good news firms do not warn despite a positive warning effect, but that a significant number of bad news firms do warn despite a negative warning effect.4 These counter-intuitive findings raise a number of

1 Consistent with prior research, I use the terms an earnings warning and earnings guidance

interchangeably for both good and bad news Earnings warnings are also referred to as earnings preannouncements (Soffer, Thiagarajan, and Walther [2000])

2 Market reaction to earnings news under a warning scenario is the reaction to the news in both the earnings warnings and earnings announcements combined while market reaction to earnings news under a no-warning scenario is measured as the returns over a comparable window to that of a warning scenario

3 Good (bad) news firms are those with positive (negative) total earnings news revealed through a warning (if any) and an earnings announcement

4 KL document that 90% of firms with relatively large good news do not warn while 21% of firms with relatively large bad news do warn Similarly, AST document that 97% of all good news firms do not warn while 13% of all bad news firms do warn Moreover, the business press reports

questions: First, is the warning effect indeed positive (negative) for good (bad) news warnings? Second, does the warning effect motivate managers to issue earnings warnings, and if so, how? This study empirically investigates these questions

Healy and Palepu [2001] and Core [2001] point out that a firm’s warning choice is likely to be endogenous Thus, empirical evidence on the warning effect estimated by OLS regression (as in KL and AST) may be sensitive to self-selection bias due to biased estimates (Maddala [1991]).5 To address this concern, I investigate (1) whether self-selection bias exists in a firm’s warning choice, (2) whether, after controlling for self-selection bias (if any), the warning effect is indeed positive for good news warnings, as documented in AST, and (3) whether, after controlling for self-selection bias (if any), the warning effect remains negative for bad news warnings, as documented in KL and AST

KL and Shu [2001] implicitly assume that the warning effect is the ultimate motive for management to issue earnings warnings Prior research suggests has documented several differential firm-characteristics between warning and no-warning firms I classify them into three management motives (other than the warning effect) to warn: (1) a litigation-concern motive (i.e., management issue warnings to reduce or even avoid shareholder litigation risk), (2) a reputation-concern motive (i.e., management warn to establish or maintain good relationship or reputation with financial analysts and investors), and (3) an

that bad news warnings are more common than good news warnings (Ip [1997], Hovanesian

[2000], Stone [2002], and Wahlegren [2002])

5 A problem of self-selection bias arises whenever there is non-random sampling caused by

individual choices (Maddala [1991])

information asymmetry consequence-concern motive (i.e., management warn to mitigate consequences of information asymmetry, such as a higher cost of capital) Despite the implicit presumption in KL and Shu [2001] that the warning effect motivates managers to issue earnings warnings, this motive has not been explicitly examined in the warnings literature This study is the first to explicitly consider the warning effect as a motive for management to issue earnings warnings I examine how the warning effect affects a firm’s warning choice after controlling for three other management motives affect a firm’s warning choice Specifically, this study investigates whether and how the warning effect is associated with a firm’s tendency to warn after controlling for the effects of the litigation-concern, reputation-concern, and information asymmetry consequence-concern motives on a firm’s tendency to warn

Because the warning effect is both a motive and a result of earnings warnings, I use a simultaneous equations model with qualitative and limited dependent variables as introduced by Lee [1978] This approach allows me to examine (1) the presence of self-selection bias in a firm’s warning choice, (2) the warning effect after controlling for self-selection bias, and more importantly (3) the relations between a firm’s tendency to warn and the warning effect after controlling for three other management motives to warn

Lee’s [1978] approach is briefly described as follows First, three models are established: (1) a warning choice model, (2) a market reaction model under a warning scenario, and (3) a market reaction model under a no-warning scenario

A firm’s warning choice is likely to be determined by the warning effect as well

as three other management motives to issue earnings warnings (i.e., concern, reputation-concern, and information-asymmetry-concern motives) Therefore, all four motives need to be specified as independent variables in the warning choice model The dependent variable for this model is the choice made

litigation-by management to warn or not to warn The warning effect is a difference between the market reaction when a firm warns and the market reaction when it

does not warn, ceteris paribus However, the no-warning market reaction cannot

be observed for a firm that chooses to warn Similarly, the warning market reaction cannot be observed for a firm that does not warn Lee [1978] therefore suggests that both market reaction models be substituted directly into the warning choice model as the warning effect The resulting warning choice model is estimated using Probit Maximum Likelihood Estimation and the resulting

estimated coefficients are used to calculate self-selectivity variables (Inverse Mills

Ratio), which are in turn included as a dependent variable in both market reaction models to control for potential self-selection bias The resulting market reaction models are estimated with Ordinary Least Square (OLS) regression OLS regression should yield unbiased estimators once self-selection bias is controlled for, allowing empirical assessment of the warning effect.6 Finally, the original warning choice model is estimated with the unbiased warning effect using Probit Maximum Likelihood Estimation and thus associations between a firm’s tendency

to warn and the warning effect after controlling for three other management motives are investigated

My sample consists of all firms included in the Institutional Broker

Estimate System (I/B/E/S) database from 1998 to 2000 I obtain reported

quarterly earnings and financial analysts’ quarterly earnings estimates from

I/B/E/S Earnings warnings are obtained from the First Call Historical database

Quarterly earnings announcement dates and selected quarterly accounting

information are obtained from the Compustat database while daily security returns are obtained from the Center for Research in Security Prices (CRSP) daily stock database Security Data Company (SDC) is the source for debt and equity

issuance data My final sample consists of 23,018 firm-quarters (4,482 firms) of which, 13,818 firm-quarters are good news and 9,200 are bad news.7 Consistent with prior research, I find that firms are more apt to warn about bad news than good news Specifically, 1,258 (9.10%) firm-quarters with good news have earnings warnings while 2,045 (22.23%) firm-quarters with bad news have earnings warnings

As expected, I find evidence of self-selection bias in a firm’s warning choice Moreover, after controlling for the bias, I find that the warning effect, on average, is positive for good news warnings and negative for bad news warnings This suggests that even though self-selection bias exists in a firm’s warning choice, the empirical findings in KL and AST do not appear to be materially

altered by it I also find that self-selection in a firm’s warning choice appears to create a downward bias in the warning effect

More importantly, I find that a firm’s tendency to warn is significantly positively associated with the warning effect after controlling for three other management motives, suggesting that the warning effect itself provides management with an economic incentive to issue earnings warnings For other management motives, I document that, consistent with prior research, the litigation-concern motive and the reputation-concern motives determine a firm’s warning choice That is, my empirical evidence suggests that firms that are more vulnerable to shareholder litigation are more likely to warn than other firms and that firms that are more concerned with their reputation with financial analysts and investors are more likely to warn than other firms The association between a firm’s tendency to warn and the information asymmetry consequence-concern motives, however, is insignificant

In summary, my findings suggest that although the warning effect provides management with an economic motivation to issue earnings warnings, it

is only one of several management motives Therefore, the notion that a firm will warn only if the warning effect is positive and will not warn otherwise, as implied

by KL and Shu [2001], is not completely accurate A more appropriate depiction

is that that a firm prefers a more positive warning effect, all else being equal

My findings contribute to the literature on management earnings warnings

in two ways First, I provide empirical evidence suggesting that self-selection does exist in a firm’s warning choice and it creates a downward bias in the

warning effect and that after controlling for the bias, the warning effect, on average, remains positive (negative) for good (bad) news warnings Thus, contrary to Shu [2001] findings, I find that the results in KL and AST are not altered by the self-selection bias Shu [2001] re-examines KL’s results on bad news warnings of firms with bad news of at least 1% of a stock price and finds that the warning effect is positive Her results, however, could be spurious due to logical inconsistency resulting in model misspecification, and methodological problems avoided in my design.8 Furthermore, Shu [2001] only examines the warning effect for bad news warnings of firms with relatively large bad news while I examine the warning effect for both good news and bad news warnings of firms with all magnitudes of news

Second, the main contribution of this study is that it is the first to investigate an association between a firm’s tendency to warn and the warning effect Prior research has not specified a warning choice model with the warning effect as a determinant of the warning choice Overall, my findings suggest that the stock market responds to earnings warnings and that managers consider a capital-market incentive (i.e., the warning effect) when they are making a warning decision, among other things

8 Shu [2001] implicitly assumes that the warning effect is the sole determinant of a firm’s warning choice She models her warning choice model without the warning effect as an independent variable, instead includes certain firm-specific characteristics, which prior research has found to be other determinants of a firm’s warning choice This logical inconsistency clearly results in the model misspecification Furthermore, Maddala [1983; 1991], among others, explicitly suggests

that (1) self-selectivity variables should be used to estimate unbiased OLS estimates, and never be

used to estimate dependent variables; and (2) the warning effect should be estimated based on

estimated values of both market reaction associated with earnings news under warning and

no-warning scenarios Shu [2001] fails to follow these guidelines

Potential avenues for future research may include an investigation of management motives to issue different types of earnings warnings, namely, point, range, open-ended, and qualitative warnings as well as a use of self-selection analysis in other settings Specifically, an investigation of whether and how the warning effect determines a firm’s choice of warning type would give insights regarding management voluntary disclosure choice In addition, management may self-selectively choose one accounting choice over others and management’s decision may affect how the stock market reacts to his accounting choice decision

The remainder of this dissertation is organized as follows Chapter 2 discusses prior studies on the warning effect, self-selection bias in a firm’s warning choice, and management motives to issue earnings warnings Chapter 2 also develops hypotheses regarding the existence of self-selection bias in a firm’s warning choice and the association between a firm’s tendency to warn and the warning effect Chapter 3 specifies the market reaction and warning choice models, and describes estimation procedures that address self-selection bias Chapter 4 addresses the sample selection criteria, defines all empirical proxies employed in the study and discusses the basic characteristics of the sample firms Chapter 5 presents empirical tests and discusses the results of these tests related to the presence of self-selection bias in a firm’s warning choice and its impacts on the warning effect as well as an association between a firm’s tendency to warn and the warning effect after controlling for three other management motives to

issue earnings warnings Chapter 6 reviews the contribution of the study, proposes possible avenues for future research and concludes the dissertation

Chapter 2: Prior Studies and Hypotheses

This chapter discusses prior studies on differential market reactions induced by earnings warnings (i.e., the warning effect), management motives for issuing earnings warnings and issues related to self-selection bias It also develops hypotheses regarding the existence of the self-selection bias in a firm’s warning choice and the association between a firm’s tendency to warn and the warning effect

2.1 D IFFERENTIAL MARKET REACTION (T HE WARNING EFFECT )

Kasznik and Lev [1995; hereafter KL] examine whether there is a differential market reaction associated with earnings news between firms that warn and firms that do not warn (i.e., a differential market reaction induced by

earnings warnings or “the warning effect”) Market reaction to earnings news for

warning firms is the reaction to the news in both the earnings warnings and the earnings announcements (i.e., a combination of cumulative abnormal returns around the warning date and the earnings announcement date) Market reaction to earnings news for no-warning firms is measured as the cumulative abnormal returns over a comparable window to that of warning firms Their sample is limited to firms with large earnings news (i.e., the absolute value of earnings news

of at least 1% of the beginning-of-quarter stock price) from 1988 to 1990 They find that, all else being equal, market reaction associated with earnings news of bad news firms that warn is more negative than that of bad news firms that do not

warn (i.e., a negative warning effect for bad news warnings) KL specifically state that “We consider this finding [a negative warning effect] counter-intuitive because a warning generally provides partial information about the subsequent earnings surprise, and is therefore expected to be rewarded by investors” (pp 128, KL) However, they find insignificant results for good news firms

Atiase, Supattarakul, and Tse [2003, hereafter AST] extend KL by examining the warning effect for firms with earnings news covering a broad range

of magnitude over the period 1995 to 1999 They find the market reaction associated with earnings news of good news firms that warn is more positive than that of good news firms that do not warn (i.e., a positive warning effect for good news warnings).9 In addition, consistent with empirical results in KL, they find the market reaction associated with earnings news of bad news firms that warn is more negative than that of bad news firms that do not warn (i.e., a negative warning effect for bad news warnings)

Libby and Tan [1999] provide a possible explanation for a negative warning effect for bad news warnings They conjecture that a negative warning effect for bad news warnings arises from financial analysts’ (or investors’) sequential information processing Their experimental evidence suggests that a bad news warning itself does not give rise to a negative warning effect, but rather financial analysts’ sequential information processing induces the negative warning effect Specifically, they find that financial analysts’ forecasts of future

9 This seems to be consistent with King, Pownell, and Waymire’s [1990] conjecture that

management voluntary disclosures reduce investors’ need to privately acquire information, thus reducing capital-market transaction costs (i.e., the transaction cost saving argument) This

transaction cost savings may induce investors’ positive response to earnings warnings

earnings are lower when they receive a bad news warning first and then later receive an earning announcement than when they receive no warnings whatsoever In addition, they find that financial analysts’ forecasts of future

earnings are higher when they receive a bad news warning concurrently with an

earnings announcement than when they receive no warnings whatsoever This suggests that the bad news warning itself has a positive warning effect, but the fact that financial analysts have to revise their forecasts twice (i.e., once after receiving a bad news warning and again after receiving an earnings announcement) induces a negative warning effect Experimental results in Libby and Tan [1999] corroborate empirical results in KL and AST that the warning effect is negative for bad news warnings

2.2 S ELF - SELECTION IN A FIRM ’ S WARNING CHOICE

Healy and Palepu [2001] and Core [2001] point out in their review papers that a firm’s warning choice is likely to be endogenous.10 Specifically, a firm chooses to warn or not to warn based on certain exogenous factors, which are likely to represent management motives to issue earnings warnings, and thus characteristics of warning firms and no-warning firms are likely to be systematically different (e.g., Cox [1985], Waymire [1985], KL, Shu [2001], Chen [2002], and AST) As a consequence, the warning effect estimated by OLS regression (as in KL and AST) may be sensitive to potential self-selection bias due to biased estimators (Maddala [1983; 1991]) Market reaction associated with

10 A firm does not randomly choose to warn or not to warn or a firm self-selectively chooses to warn or not to warn

earnings news is observable only in the state chosen by the firm (warning or warning) while market reaction associated with earnings news in the alternative condition (no-warning or warning) is unobservable For example, if firm A decides to warn, the market reaction associated with its earnings news in a warning scenario is observable, but the market reaction associated with its earnings news in a no-warning scenario is unobservable This creates a limited dependent variable or self-selectivity problem In the presence of a limited dependent variable or self-selectivity problem, OLS regression cannot be used to obtain unbiased estimated coefficients (Maddala [1983; 1991])

no-This study examines how earnings warnings affect the market reaction associated with earnings news, after controlling for potential self-selection bias I investigate (1) whether self-selection bias exists in a firm’s warning choice, (2) whether, after controlling for potential self-selection bias, the warning effect remains positive for good news warnings, as documented in AST, and (3) whether, after controlling for potential self-selection bias, the warning effect remains negative for bad news warnings, as found in KL and AST

Shu [2001] re-examines KL’s results by using Heckman two-stage regression to control for potential self-selection bias in an attempt to explain KL’s counter-intuitive findings that the warning effect is negative for bad news warnings of firms with relatively large bad news She finds that after controlling for the self-selection bias, the warning effect is positive for warning firms, but the warning effect would have been negative for no-warning firms had they decided

to warn She concludes that the firms in her sample, on average, make rational

warning choices Her findings, however, could be spurious due to a problem of logical inconsistency resulting in model misspecification, and some methodological problems as briefly discussed below

Shu’s conclusion rests on the unstated premise that a firm will warn only

if the warning effect is positive and will not warn otherwise, i.e., the warning effect is the sole management motive to issue earnings warnings However, the warning choice model she employs fails to include the warning effect as an independent variable; instead she includes certain firm-specific characteristics that proxy for management motives to issue earnings warnings As a result, her model

is misspecified

Furthermore, according to Maddala [1983; 1991], among others, selectivity variables should only be used to obtain unbiased OLS estimates (in the

self-market reaction models), and never be used to estimate dependent variables (i.e.,

market reactions associated with earnings news); and differential market reactions

in this case, i.e., the warning effect, should be calculated based on estimated

market reactions associated with earnings news under warning and no-warning scenarios Shu [2001] fails to follow these guidelines

I use a simultaneous equations model with qualitative and limited dependent variables as introduced by Lee [1978] to more properly address the self-selection issue Lee’s [1978] approach avoids the logical inconsistency and methodological problems found in Shu [2001] Chapter 3 describes this approach

in detail

As discussed earlier, prior research has documented that firm characteristics of warning and no-warning firms are systematically different and thus it is likely that the firm characteristics are related to the firm’s decision to warn or not to warn As a result, I hypothesize that self-selection bias exists in a firm’s warning choice However, it is not clear whether or not extant empirical findings that the warning effect is positive for good news warnings (AST) and negative for bad news warnings (KL and AST) are influenced by the potential self-selection bias Therefore, I do not make any predictions regarding the sensitivity of the empirical results in KL and AST to the possible self-selection bias

2.3 M ANAGEMENT MOTIVES TO ISSUE EARNINGS WARNINGS

Prior research has documented seven differential firm-characteristics between warnings and no-warning firms: (1) membership in high-litigation risk industries, (2) the magnitude of earnings news, (3) market capitalization, (4) past warning pattern, (5) analyst following, (6) membership in regulated industries, and (7) future external finance offering I classify these firm-characteristics into three management motives to issue earnings warnings: (1) a litigation-concern motive (items 1-3), (2) a reputation-concern motive (items 4-6), and (3) an information asymmetry consequence-concern motive (item 7) Therefore, these three motives need to be controlled for in the warning choice model

Skinner [1994] suggests that management may issue earnings warnings to reduce or even avoid shareholder litigation risk.11 Skinner [1997] provides empirical evidence suggesting that earnings warnings can reduce settlement amounts in shareholder lawsuits consistent with the litigation-concern motive

Baginski, Hassell, and Kimbrough [2002] also find that a firm’s shareholder litigation environment has a significant impact on a firm’s warning decision to warn or not to warn Similarly, KL and Shu [2001] provide empirical evidence suggesting that firms in industries that tend to be vulnerable to shareholder litigation (mostly high-tech industries) are more likely to warn than firms in other industries In addition, they find that firms with relatively large earnings news, especially bad news, and those with large market capitalization, who are likely to be targets of shareholder lawsuits, are more likely to warn than other firms All these studies support the litigation-concern motive

Skinner [1994] also conjectures that management may issue earnings warnings to establish or maintain a good reputation with financial analysts and investors Miller and Piotroski [2000] and Chen [2002] provide empirical evidence, consistent with this reputation-concern motive Specifically, Miller and Piotroski [2000] find that firms that have issued earnings warnings in prior periods are likely to do so in the current period, suggesting that these firms intend

to establish or maintain a reputation as warning firms

11 Vicker [1999] states in her article in Business Week that if companies fail to meet analysts’

estimates (i.e., have bad news on the earnings announcement date), they risk shareholder lawsuits

In fact, she reports that shareholders recently filed suit against Compaq Company Corp charging that the company did not timely inform them about its disappointing performance

Chen [2002] finds that analyst following is positively associated with a firm’s tendency to warn Firms with a large analyst following need to maintain a good reputation with their analysts Issuing earnings warnings is one way to maintain a good relationship with financial analysts.12

In addition, KL find that firms in regulated industries (i.e., utility, communication, and financial firms) are less likely to warn than other firms KL speculate that this may be because firms in the regulated industries are required to disclose financial information in more detail and thus the incremental benefit of earnings warnings to financial analysts and investors is likely to be minimal, compared to the benefit for firms in unregulated industries

Lang and Lundholm [1993] conjecture that high quality disclosures may reduce information asymmetry and increase firm value at the time of debt or equity issuance Botosan [1997] documents that disclosure quality is adversely associated with cost of equity capital and similarly, Sengupta [1998] finds an adverse relationship between disclosure quality and cost of debt Ruland, Tung, and George [1990], Frankel, McNichols, and Wilson [1995] and Miller and Piotroski [2000] provide empirical results that firms planning to issue public offerings (either debt or equity) are more likely to provide voluntary disclosures

to the stock market Shu [2001] also finds a positive association between a firm’s tendency to warn and a firm’s tendency to issue debt or equity in the market Taken as a whole, these studies indicate that earnings warnings may be a mechanism used to reduce potential consequences of information asymmetry

12 Skinner [1994] argues that financial analysts may impose costs on firms whose managers are less than candid about a potential earnings surprise For example, analysts may choose not to follow firms that continue not to warn analysts about their earnings surprise

(e.g., a high cost of capital) This conforms to the information asymmetry consequence-concern motive to issue earnings warnings

KL and Shu [2001] tacitly imply that the warning effect is management’s sole motive to issue earnings warnings, i.e., a firm will warn only if the warning effect is positive and will not otherwise For example, KL specifically state that their finding that the warning effect is negative for bad news warnings is

“counter-intuitive.” Shu [2001] likewise concludes that bad news firms in her sample, on average, make rational warning choices based on her problematic findings that the warning effect is positive for warning firms but the warning effect would have been negative for no-warning firms if they had decided to warn

It is unlikely that the warning effect is the sole management motive to issue earnings warnings since prior research shows litigation concerns, reputation concerns, and information asymmetry consequence concerns all appear to motivate management to issue earnings warnings Despite the implicit presumption in KL and Shu [2001] that the warning effect motivates managers to issue earnings warnings, this motive has not been explicitly examined in the warnings literature This study is the first to explicitly consider the warning effect itself as a management motive for issuing earnings warnings and examine how the warning effect affect a firm’s warning choice How the warning effect affects a firm’s warning decision (to warn or not to warn) is a fundamental question that has not been addressed in the earnings warnings literature This study therefore investigates whether and how the warning effect affects a firm’s tendency to warn after controlling for litigation concerns, reputation concerns, and information

asymmetry consequence concerns Since intuitively management prefers a more positive market reaction, all else being equal, I hypothesize that a firm’s propensity to warn is positively associated with the warning effect

Because the warning effect is both a motive and a result of earnings warnings, I use a simultaneous equations model with qualitative and limited dependent variables introduced by Lee [1978] This approach allows me to measure the warning effect after controlling for self-selection bias and to explicitly examine the associations between a firm’s tendency to warn and the warning effect after controlling for three other management motives

Chapter 3: Model Specification and Estimation Procedures

This chapter describes the models and estimation procedures used in the study I specify market reaction models for warning and no-warning scenarios and a warning choice model I also employ estimation procedures to deal with possible self-selection bias This study implements a simultaneous equations model with qualitative and limited dependent variables as introduced by Lee

[1978] Lee’s [1978] approach allows me to estimate the unbiased warning effect

as well as to explicitly examine the association between a firm’s tendency to warn and the warning effect, controlling for other management motives to issue earnings warnings

~ε where

;εZββ

W

W i

W i

* i

*W 1

~ε where

;εZββ

N

N i

N i

* i

*N 1

* 2 i

* 1

0

0 WARNif

1 WARN

i

* i i

In Eqs (1) and (2), MRWi and MRNi denote firm i’s market reactions

associated with earnings news under warning and no-warning scenarios,

respectively, and denotes a vector of firm i’s firm-specific characteristics that determine firm i’s market reaction to earnings news denote firm i’s

random residuals under warning and no-warning scenarios and are assumed to be

and , respectively

* iZ

σN(0,

N i

In Eq (3), ∆MRi =MRWi −MRNi, is firm i’s warning effect (i.e., the

differential market reaction to earnings news between warning and no-warning

scenarios) and S represents a vector of firm i’s firm-specific characteristics that

proxy for other management motives to issue earnings warnings: the concern motive, the reputation-concern motive, and the information asymmetry consequence-concern motive

litigation-* i

* iWARN

WARN

is a latent variable that represents firm i’s unobserved tendency

to warn; i denotes firm i’s observable warning choice, where WARNi =1

if firm i chooses to warn and WARNi =0 if firm i chooses not to warn The

This study is the first to specify a warning choice model using the warning effect as one of determinants of warning choice

0WARN*

Since vectors in Eqs (1) and (2) and S in Eq (3) contain common

* iZ

Zi

* i]

X[

i = i denotes a vector of

variables common to the warning choice and market reaction models for firm i

Thus, Eqs (1), (2), and (3), respectively, are re-written as follows:

W i i

W 2 i

W 1

W 0

N i i

N 2 i

N 1

N 0

i i 3 i 2 i i

1 0

3.2 L IMITED DEPENDENT VARIABLE OR SELF - SELECTIVITY PROBLEM

If, for any particular firm i, both MRWi and MRNi were observable, firm

i’s warning effect ( ) could be easily measured as , and the warning choice model (Eq (6)) could be estimated using Probit Maximum

Likelihood Estimation For any particular firm i, either MRW

i

∆MR

N 1

observable depending upon firm i’s warning choice (WARNi), but not both For

example, if firm i decides to warn (WARNi = 1), MRWi is observable, but MRNi

is not Similarly, if firm i decides not to warn (WARNi = 0), MRNi is observable, but MRWi is not Even if only one state is observable, if the waning choice is

random, OLS regression gives unbiased estimates coefficients

easily obtained As a result, is measurable and the warning choice model (Eq (6)) can again be estimated using Probit Maximum Likelihood Estimation

1

N 0

Mˆ

∆

It follows from the warning choice model (Eq (6)) that firm i

self-selectively chooses to warn or not to warn That is, firm i chooses to warn or not

to warn based on the warning effect ( ) and other management motives to issue earnings warnings ( ) This creates a limited dependent variable or self-selectivity problem In this situation, the error terms, , in Eqs (4)

and (5) are conditioned on firm i’s warning choice (WARN

i

∆MRi

X

N i

XβZββ

i i

W 2 i

W 1

W 0

where E(εW WARNi 1) 0

;0)WARN(ε

XβZββ

i i

N 2 i

N 1

N 0

00)WARNE(ε

Since E(εW WARNi 1) 0

i = ≠ and E(εN WARNi 0) 0

)σN(0,

~

W

W i

, the OLS regression assumption that and is clearly violated

As a result, using OLS regression to estimate Eqs (4) and (5) gives biased estimated coefficients and the warning effect estimated based on these biased coefficients is biased as well

)σN(0,

~

W

W i

If E(εW WARNi 1)

i = and εNWARNi 0)

incorporated into Eqs (7) and (8), the conditional error terms will have a mean of zero, yielding unbiased OLS estimates

3.3 L EE [1978]’ S APPROACH

Generally, to correct for self-selectivity (i.e., to obtain E(εW WARNi 1)

and E(εNWARNi 0)

i = in Eqs (7) and (8)), Heckman two-stage regression (as described in Maddala [1983; 1991]) is used.13 If the warning choice is modeled without the warning effect (∆MR), the warning choice model can be estimated using Probit Maximum Likelihood Estimation Next, the self-selectivity variables for warning and no-warning firms are calculated based on estimated coefficients from the warning choice model The self-selectivity variables are then incorporated into the corresponding market reaction models to control for potential self-selection bias and OLS regression is used to estimate the resulting market reaction models, yielding unbiased estimated coefficients Finally,

13 Shu [2001] uses Heckman two-stage regression

unbiased and are obtained and an unbiased estimate of is

iWRˆ

M

RˆM

iNRˆM

iNRˆM

−

M

iR

Mˆ

∆

iW

RMRN

i

0δ

=(+W 1β

MRW

-∆MRiWARN

W

0 ,β

W 0

1(βδ+

W 2

1(β

N 0

W

2 ,ββ ,

The warning choice model (Eq (6)) in this study is specified with the warning effect (∆ ) and other management motives (S ) Since

, Eqs (7) and (8) need to be estimated before estimating

Eq (6) Equations (7) and (8), however, cannot be estimated properly without Eq (6) due to the self-selectivity problem Therefore, Heckman two-stage regression

is not appropriate Lee [1978] introduces a simultaneous equations model with qualitative and limited dependent variables that can be used in this situation

* i

∆MR=

Lee’s [1978] procedures require using Eqs (4) and (5) to substitute MRWi

and MRNi as in Eq (6), resulting in the following equation

i

N 1

W 1 1

N i

W i 1

N

2) δ )X δ S ε

Since are not identified in Eq (9), this model

cannot be used to obtain estimates of these coefficients Letting

and , Eq (9) may be re-written as

N 2

N

1 ,andββ

,

)ε(εδ)β

1

N 0

1

W 1

2

W 2 1

i i 3 i 2 i 1 0

3 2

1

0, γ , γ ,and γ

γ are identified in Eq (10) and can be estimated using

Probit Maximum Likelihood Estimation

Assume that in Eqs (4), (5), and (10), respectively, are trivariate normally distributed, with mean vector zero and covariance matrix

i

N i

W

i , ε ,andεε

Σ , where

N

Wε WN

2 W N

Note that var( )εi =1 is assumed since the γ ’s in Eq (10) are estimable only up to a scale factor (Maddala [1983])14 Note also that the conditional distribution of , given , is normal, with mean and variance

and the conditional distribution of ε , is normal, with mean and variance (Maddala [1983]) Consistent with the Heckman two-stage approach,

W iε

2 Nσ

iε

Nεεσi

N

i , given ε2

N

σ −

1)WARNi =E(εW

i i = can be measured as follows:

1)WARN

=

)

SiγXγZγγεE(ε

+

++

+

=

)SγXγZγΦ(γ

)SγXγZγφ(γ-σ

i 3 i 2 i 1 0

i 3 i 2 i 1 0

0)WARN

=

)

SiγXγZγγεE(ε

+

++

+

=

)SγXγZγΦ(γ-1

)SγXγZγφ(γσ

i 3 i 2 i 1 0

i 3 i 2 i 1 0

As noted earlier, can be estimated from Eq (10) Letting

reaction model under a warning scenario (Eq (7)) is re-written as

3 2

1

0 ,γˆ ,γˆ ,andγˆγˆ

i

3S

γˆ πW =σWεi

2 i 1

0

e)Φ(ψ

)φ(ψπ

XβZββ

i i

i W

i

W 2 i

W 1

W 0

+

01)WARN

)φ(ψπ

XβZββ

i i

i N

i

N 2 i

N 1

N 0

+

00)WARN

2

W 1

W 0

From Eqs (7), (8), (14) and (15), it follows that

)φ(ψπ

)1WARN

E(ε

i

i W

)φ(ψπ

)0WARNε

i

i N

i

N i

non-to be significantly different from zero

N i

W

ε

After obtaining unbiased β from Eqs (14) and

(15), the warning effect ( ∆ ) is estimated as follows:

N 2

N 1

N 0

W 2

W 1

W

0 ,βˆ ,βˆ ,βˆ ,βˆ ,andβˆˆ

MR

15 Specifically, non-zero and imply that OLS estimated coefficients in Eqs (7) and (8) are biased due to omitted variables – self-selectivity variables (i.e., additional terms in squared blankets in Eqs (14) and (15)) These self-selectivity variables are generally referred to as the

“Inverse Mills Ratio.”

W

Mˆ

∆ =MRˆWi−MRˆNi

βˆZβˆ

N 0

2 i

N 1

N 0 i

W

i

N 2

W 2 i

In order to investigate whether and how the warning effect affects a firm’s tendency to warn after controlling for other possible management motives, I estimate Eq (6), the warning choice model, using Probit Maximum Likelihood Estimation as follows:

i i 3 i 2 i 1 0

3.4 M ETHODOLOGICAL PROBLEMS IN S HU [2001]

Both in Eq (16) are calculated without self-selectivity

variables According to Maddala [1983; 1991], self-selectivity variables should only be used to estimate unbiased OLS estimated coefficients of independent variables (i.e., and β ) and should not be used to estimate dependent variables

i

WRˆM

N Rˆ

W Rˆ M

(i.e., ) Moreover, ∆ is calculated based on both estimated

and (Lee [1978] and Maddala [1983; 1991]) Observed MRW

MˆRˆ

Shu [2001], in attempting to control for the self-selection bias, violates both of these guidelines Shu [2001] estimates M for warning firms and

for no-warning firms by including the self-selectivity variables along with other independent variables but fails to estimates for warning firms and for no-warning firms As a result, she calculates the warning effect for warning firms with observed MRW and estimated M (i.e.,

if firm i warns), and calculates the warning effect for

no-warning firms with estimated and observed MRN (i.e., if firm i does not warn) Therefore, her results

could be spurious due to these methodological mistakes

NRˆ

WRˆM

W

NRˆi

iMRN

More importantly, her model is misspecified, reflecting a logical inconsistency Specifically, she assumes that a firm will warn only if the warning effect is positive which is equivalent to assuming that the warning effect is the sole motive underlying the warning decision However, she specifies her warning choice model without including the warning effect as an independent variable, instead including certain firm-specific characteristics (i.e., tentative proxies for management motives other than the warning effect)

In summary, Shu’s [2001] empirical findings that, on average, the warning effect is positive for warning firms and it would have been negative for no-warning firms had they warned, could be spurious due to the methodological

problems and the logical inconsistency that compromises her model specification Her conclusion that bad news firms in her sample make rational warning decisions may be unfounded