Disclosure Quality and Information Asymmetry

Transcription

1 Disclosure Quality and Information Asymmetry Stephen Brown # Stephen A. Hillegeist December 2003 Abstract: We examine the association between firms disclosure quality and information asymmetry using a three-stage least squares estimation procedure that takes into account the endogeneity between these two variables. Our measure of information asymmetry is the Probability of Informed Trade (Easley, Kiefer and O'Hara [1997]), which measures the probability of information-based trade. Our primary result is that firms overall disclosure quality is negatively associated with information asymmetry. This negative association continues to hold for each of the three major components of disclosure quality, although the results are strongest for the quality of investor relations activities. Further analyses indicate that the underlying source of this negative association is a negative association between disclosure quality and the relative amount of trading by privately informed investors. Since prior work indicates that the cost of equity capital is increasing with information asymmetry, our results suggest that firms with higher disclosure quality have lower costs of capital. Key Words: Disclosure Quality; Information Asymmetry; Market Microstructure This paper has benefited from the comments and suggestions of Eli Bartov, Tarun Chordia, Paul Fisher, Simon Gervais, Ole-Kristian Hope, Ravi Jagannathan, Joseph Paperman, Gideon Sarr, Yong-Chul Shin, Sri Sridar, Beverly Walther, Greg Waymire, and seminar participants at University of Chicago, Georgia State University, University of Illinois at Chicago, University of Michigan, University of Minnesota, New York University, and Northwestern University. The authors especially wish to thank Mark Finn for his efforts on an earlier version of this paper. The second author gratefully acknowledges the financial support of the Accounting Research Center at Northwestern University. We also wish to thank Christine Botosan, Marlene Plumlee, and Mark Soszek for supplying the AIMR scores used in this study. We appreciate IBES for providing the analyst forecast data. # Department of Accounting, Goizueta Business School, Emory University Department of Accounting Information and Management, Kellogg School of Management, Northwestern University. Corresponding author: Kellogg School of Management, 2001 Sheridan Rd., Room 6223, Evanston, IL 60208; Ph.: ; s-hillegeist@kellogg.northwestern.edu.

2 1. Introduction In this paper, we contribute to the growing literature on the capital market effects of voluntary disclosure choices. Consistent with economic theory (Diamond [1985], Verrecchia [2001]), we predict a negative association between a firm s voluntary disclosure quality and the level of information asymmetry between investors. Disclosure quality is directly related to asymmetry because it decreases the amount of private information relative to public information; it is indirectly related to asymmetry because it reduces incentives to search for private information. After controlling for the endogenous relation between disclosure quality and asymmetry, we find that asymmetry is negatively associated with the quality of a firm s disclosures. Understanding this association is important because information asymmetry is negatively related to the cost of equity capital (COC). This negative relation arises because privatelyinformed traders profit at the expense of uninformed investors. Rational uninformed investors anticipate these future losses to informed traders and price protect themselves by demanding an ex ante risk premium based on the expected level of information asymmetry. The risk is not diversifiable since uninformed investors are always at a disadvantage compared to informed investors (O'Hara [2003]). Firms can affect their COC through their choice of disclosure quality, as disclosure essentially transforms private information into public information. Empirical evidence in Easley, Hvidkjaer and O'Hara [2002] indicates that less asymmetry is associated with a lower cost of equity capital. 1 We employ a recently-developed model the EKO model from the market microstructure literature (Easley, Kiefer and O'Hara [1997]) to estimate the firm-specific level of information asymmetry. In the EKO model, large trade order imbalances result from the trading 1 Additionally, there is a widespread belief among accounting regulators, standards setters, and accounting practitioner organizations (AICPA [1994], FASB [2001], FASC [1998], Levitt [1998]) that high quality disclosures reduce a firm s cost of capital. 1

3 activities of privately-informed traders. We use the daily order flow over an annual period to estimate the Probability of Informed Trade, PIN. 2 The PIN is a firm-specific estimate of the probability that a particular trade order originates from a privately-informed investor; hence, it directly captures the extent of information asymmetry among investors in the secondary market. An important advantage of this approach is that it enables us to analyze the sources of the underlying relation between disclosure quality and information asymmetry. As Section 2 discusses, higher disclosure quality can either reduce the relative amount of informed trading and/or decrease the frequency with which certain investors obtain private information. Such analyses are not possible using spread-based proxies of information asymmetry. Consistent with many previous studies, including Botosan and Plumlee [2002] and Lang and Lundholm [1996], we use analysts evaluations of disclosure compiled by the Association for Investment Management and Research (AIMR) as our proxy for disclosure quality. While they are noisy measures of disclosure quality, the AIMR scores offer several advantages over alternative proxies. First, they are based on a comprehensive evaluation of the firm s disclosure activities. Thus, our study generalizes and complements other studies that focus on a specific type of disclosure. 3 Second, they allow us to gain insights into the relative importance of different types of disclosures (i.e., those related to the annual report, quarterly reports, and investor relations activities). Finally, they allow us to examine the effects of disclosure quality on a relatively large and representative population of firms. 4 2 While new to the accounting literature, the PIN methodology has been successfully used in the finance literature to investigate issues such as how informed trading differs across stock exchanges (Easley, Kiefer and O'Hara [1996]) and securities (Easley, O'Hara and Srinivas [1998]), and the effect of stock splits on information asymmetry (Easley, O'Hara and Saar [2001]). 3 For example, Coller and Yohn [1997] and Marquardt and Wiedman [1998] examine management forecasts. 4 In contrast, Lang and Lundholm [2000], Marquardt and Wiedman [1998], Schrand and Verrecchia [2002], and Sengupta [1998] examine the effects of disclosure in the context of raising capital using relatively small samples. 2

4 In addition to disclosure s effect on information asymmetry, the concurrent level of information asymmetry is likely to influence a firm s choice of disclosure quality. If the current level of information asymmetry is high, for example, the firm may choose a higher level of disclosure quality, expecting it to reduce the level of asymmetry. A failure to incorporate this potential endogeneity into our research design can result in biased coefficients and misleading inferences (Maddala [1983]). In order to address the endogeneity problem, we employ a threestage least squares (3SLS) estimation procedure that models disclosure quality and information asymmetry simultaneously. Our primary tests indicate that the overall quality of the firm s disclosures (Total) is significantly and negatively related to the level of information asymmetry after controlling for the endogenous relation between disclosure quality and asymmetry. Separately examining the three component subscores indicates that this overall negative relation extends to each component (Annual, Quarterly, and IR); the relation is strongest for the quality of the investor relations activities. These findings continue to hold in alternative regression specifications. In addition, we examine the underlying sources of the negative relation between disclosure quality and information asymmetry. We find that higher quality is negatively associated with the relative amount of trading by privately-informed investors. This finding is consistent with the theoretical results in Fishman and Hagerty [1989] and Merton [1987], which indicate that more informative disclosures increase the amount of uninformed trading. Combined with the results in Easley, Hvidkjaer and O'Hara [2002], our findings provide some empirical support for regulators beliefs about the capital market benefits of high quality disclosures. Our use of a direct measure of information asymmetry contrasts with indirect spreadbased proxies of asymmetry used in prior studies, such as Welker [1995] and Leuz and 3

5 Verrecchia [2000]. The spread between the quoted bid and ask prices compensates the market maker for expected losses to privately-informed traders, in addition to covering her inventory holding and order processing costs. Relying on the EKO methodology avoids the numerous econometric problems and interpretation difficulties that occur when using spread-based measures of information asymmetry (Callahan, Lee and Yohn [1997], Neal and Wheatley [1998], and O'Hara [1995]). For example, market makers protect themselves from information asymmetry by simultaneously manipulating both the quoted bid and ask prices along with the quoted depths associated with those prices. Unless depths are included in research design, spread-based analyses are incomplete and difficult to interpret (Lee, Mucklow and Ready [1994]). 5 In the next section, we develop our hypothesis about the association between disclosure quality and information asymmetry. In Section 3, we develop our empirical proxy of information risk based on the EKO model, and we discuss our proxy for disclosure quality. We also discuss some of the restrictive assumptions that underlie the PIN measure and their likely impact on our analyses. In Section 4, we describe how our research methodology takes the simultaneous relation between disclosure quality and information asymmetry into account. Section 5 describes the data sources, variable construction, and provides descriptive statistics. Section 6 presents the results of our empirical tests, and Section 7 summarizes and concludes the paper. 2. Disclosure Quality and Information Asymmetry Private information represents information about firm value that has not (yet) been incorporated into the firm s stock price. Hasbrouck's [1991] characterization of private information as 5 Furthermore, since spreads and depths are determined simultaneously, this endogeneity should be incorporated into the research design. 4

6 essentially prior knowledge of public information indicates that timing differences account for much of the distinction between private and public information. Information asymmetry in the stock market occurs when one or more investors possess private information about the firm s value while other investors are uninformed. This dichotomy of information among investors is consistent with Admati and Pfleiderer [1988], Diamond and Verrecchia [1991], Easley and O'Hara [1992], Glosten and Milgrom [1985], Kim and Verrecchia [2001], Kyle [1985], and McNichols and Trueman [1994], among others. 6 The level of information asymmetry can be characterized by the risk of trading with a privately-informed investor. A firm s choice of disclosure quality--which we define as the precision, timeliness, and quantity of information provided--affects this information risk by altering the distribution of public and private information among investors. Disclosure quality is related to information risk because it affects the incentives to search for private information. Higher disclosure quality reduces the incentives to search for private information by reducing the expected benefits from obtaining private information. Diamond [1985], Hakansson [1977], and Verrecchia [1982] examine settings where public and private information are substitutes for each other. In these settings, increased public disclosures by firms generally reduce the incentives to collect costly private information. These findings suggest that higher disclosure quality leads to less private information being produced. Higher disclosure quality also leads to less private information search activities, since it raises the costs of these activities. As timeliness is an important attribute of disclosure quality, firms with higher disclosure quality are more likely to release material information promptly and to provide forward-looking information. When search activities take time to complete, fewer 6 This characterization of information asymmetry is, however, distinct from the settings analyzed in Kim and Verrecchia [1991] and Verrecchia [1983], where all investors receive private information, but each private signal contains an idiosyncratic noise component. 5

7 opportunities arise to discover and trade on private information about those firms with high disclosure quality. This reduction occurs because less time is available between the information s emergence and when potentially informed traders expect the firm to disclose the information publicly. Less search for private information occurs since the net costs of discovering private information increase (Grossman and Stiglitz [1980]). Consequently, higher disclosure quality reduces the frequency with which certain investors obtain private information (which we call private information events) by reducing the intensity of search activities. Fewer private information events, ceteris paribus, reduce the risk of trading with a privately-informed investor, thereby reducing the level of information asymmetry. In addition to its impact on information search incentives, higher disclosure quality also reduces information asymmetry by altering the trading behavior of uninformed investors. Merton [1987] and Fishman and Hagerty [1989] analyze settings in which firms more informative disclosures reduce the costs associated with processing and assimilating public information. As a result, greater disclosure induces more investing by uninformed liquidity traders. Diamond and Verrecchia [1991] find that the amount of uninformed trading by large investors can increase as the firm discloses more information. 7 Assuming that the costs of processing public information decrease with the quality of the information, we expect that disclosure quality is positively associated with the amount of uninformed trading. Consistent with this assumption, accounting regulators and practitioners argue that more and better disclosure by firms makes the capital markets more attractive to uninformed investors by leveling the playing field (AICPA [1994], FASB [2001], Levitt [1998]). More uninformed trading decreases the risk of trading against a privately-informed investor, 7 Consistent with these models, Leuz and Verrecchia [2000] find a significant increase in trading volume for German firms committing to higher disclosure levels. Since bid-ask spreads also decrease, their evidence suggests that uninformed investors generate much of the increased trading. 6

8 ceteris paribus. However, previous research indicates that increases in uninformed trading are associated with more informed trading. Kyle [1985] demonstrates that if investors are risk neutral and are not capital constrained, then the amount of informed trading varies proportionately with the expected amount of uninformed, liquidity-based trading. Informed traders act strategically to maximize their trading profits by taking advantage of the noise provided by the trading activity of uninformed investors. The end result is that the relative amount of informed trading remains unchanged when the expected amount of uninformed trading changes. In practice, however, it is likely that informed traders are risk averse and/or capital constrained, and thus, there is a less than fully proportional change in the amount of informed trading. Accordingly, we expect that higher disclosure quality is associated with relatively less informed trading and, therefore, less information asymmetry. The above discussion suggests that the level of disclosure quality should be negatively associated with the level of information asymmetry for two reasons. First, higher disclosure quality reduces the incentives to search for private information. Second, higher disclosure quality increases the relative amount of trading by uninformed investors. Accordingly, we test the following hypothesis, stated in alternate form: H1: A negative association exists between the levels of disclosure quality and information asymmetry. 3. Variable Measurement In this section, we describe our empirical proxies for the theoretical constructs we use to test hypothesis H1. First, we describe the EKO model and explain how we use it to estimate the probability of informed trade or PIN, which is our proxy for information asymmetry. Second, 7

9 we discuss our proxy for disclosure quality, which is based on analysts evaluations of firms disclosures Proxy for Information Asymmetry Information asymmetry manifests itself when investors trade on their private information, which suggests that asymmetry is observable in the form of relatively large imbalances between buy and sell orders. This observability forms the intuition behind the recent EKO microstructure model of information risk (Easley, Kiefer and O'Hara [1997]). The model allows us to estimate the probability of information-based trading (PIN) for a given stock based on the actual order flow. In the model, the market maker observes the order flow, updates her beliefs about the probability of information-based trading, and sets trading prices. Over time, the process of trading and learning from trading results in the convergence of prices to their full information levels as the private information is revealed through the imbalance between buys and sells in the order flow. The EKO model focuses on the role of private information and analyzes how it becomes impounded into prices through the trading process. The results in French and Roll [1986] and Barclay, Litzenberger and Warner [1990] indicate that this focus on private information is warranted, since they find evidence suggesting that return volatility is primarily caused by the incorporation of information into prices through the trading activities of privately-informed investors. Additionally, Hasbrouck [1991] shows that the trading process itself is a significant source of information to the market. Together, these studies suggest that private information is the dominant cause of trade imbalances and price movements in the market. The basic structure of the EKO model is shown in the game tree diagram in Figure 1. The model depicts trading as a game repeated daily between the market maker and traders. 8

10 At the beginning of the trading day, nature determines whether one or more informed investors observe new private information about a firm. Private information events occur with probability α. The private information contains bad ( good ) news with probability δ (1 - δ). Bad (good) news indicates that the asset is currently overvalued (undervalued), and hence, the profit maximizing trade is to sell (buy) the stock. The asset is correctly valued on no news days, which occur with probability (1 - α). Trade orders arrive sequentially according to independent Poisson processes. Orders from informed traders arrive at the daily arrival rate µ on information events days, while orders from uninformed buyers (sellers) arrive at the daily rate ε b (ε s ) each trading day. The market maker sets prices to buy or sell one unit of stock at each point in time, and she executes orders as they randomly arrive. Although unaware whether a private information event has occurred on any given day, the market maker knows the probability of and the expected order process associated with each event and uses the actual order flow to update her beliefs throughout the trading day. Thus, prices evolve in response to the observed order flow on a particular day. While the occurrence of private information events and the identity of traders are unobservable, their effects on the order flow are observable. The structure provided by the EKO model allows us to work backwards and use observable data on the daily number of buy orders (B) and sell orders (S) to make inferences about unobservable private information events and the division of trade between informed and uninformed traders. The likelihood function induced by the EKO model for a trading day, conditional on the parameter vector θ = (α, δ, ε b, ε s, µ), is determined by a mixture model in which the weights on the three components (no news, good news, and bad news) reflect the probabilities of their occurrence in the data. On a no-news day, the model predicts a roughly equal number of buyer- and seller-initiated trade orders, all of 9

11 which come from uninformed traders. On a good- (bad-) news day, a relatively large order imbalance emerges, with buyer-initiated (seller-initiated) trades predominating. L( B, S θ ) = (1 α) e εb B ε b e B! + α(1 δ ) e ε s ( µ + εb S ε s + αδe S! ) ( µ + ε b ) B! ε B B ε b b e B! S ε ε s s e S! ( µ + ε s ) ( µ + ε s ) S! S (1) Inspection of equation (1) reveals that the daily number of buys and sells is a sufficient statistic for the data. The parameter vector θ can be estimated via maximum likelihood with increasing precision as the daily order flow is observed over an increasing number of days. The empirical implementation of the EKO model implicitly assumes that a firm s underlying information process and hence the PIN parameters are stationary over the one-year estimation period. Evidence in Easley, Hvidkjaer and O'Hara [2002] indicates that this assumption is reasonable. Once the model s parameters (α, δ, ε b, ε s, µ) have been estimated, the PIN is calculated as follows: PIN = αµ αµ + ε + ε (2) b s Equation (2) shows that information risk increases with more frequent information events (α) and more informed trading (µ), and decreases with the amount of trading by uninformed investors (ε b and ε s ). PIN represents the percentage of trades that are expected to be information based each day, since αµ is the expected number of orders from privately-informed investors and αµ + ε b + ε s is the expected total number of orders. Thus, the ratio of the two is the ex ante probability that the first trade of the day is information based. For example, consider a stock for which on 60% of days there are 50 buys and 50 sells, on 20% of the days there are 80 buys and 50 sells; and on 20% of the days, there are 50 buys and 80 sells. The EKO model parameters 10

12 would be identified as ε b = ε s = 50, µ = 30, α =.40, and δ =.50. The corresponding PIN is 10.7%. One concern with using the PIN as a proxy for the level of information asymmetry is that by assuming all orders are the same size (or at least have the same information content), the EKO model ignores an important source of information since traders potentially reveal information by their choice of trade order size. However, privately-informed investors will disguise their presence by mimicking the trade sizes of uninformed traders (Barclay and Warner [1993], Chakravarty [2001]); otherwise, the market maker could infer their identity from their order size, which would significantly reduce the expected profitability of trading on their private information. Jones, Kaul and Lipson [1994] find that the relation between the volume of trade and the volatility of stock price changes virtually disappears when the number of trade orders is controlled for. These results suggest that the loss of information due to ignoring trade size will be small. To the extent that order size provides additional information about the identity of investors, it will add noise to our estimates. Two other important restrictions in the EKO model are: (1) each trading day is assumed to be independent, and (2) privately-informed buy and sell orders are not allowed to occur on the same day. As Easley, Kiefer and O'Hara [1997] discuss, the main impact of dependence across days is to change the interpretation of the α and δ parameters to unconditional probabilities. For example, αδ would represent the unconditional probability of a bad news day and correspond to the fraction of abnormally large, sell-led order imbalances over the sample period. Since our empirical tests and interpretations correspond to this interpretation, the effects of dependence in our study should be small. Thus, as long as information event days are classified correctly, 11

13 dependence should not affect the estimation of the PIN parameters (Easley, Kiefer and O'Hara [1997]). Violation of the second restriction may not be as benign. To the extent that informed investors initiate buy and sell orders on the same day, as in Kim and Verrecchia [1991], the estimation and interpretation of the PIN parameters will be affected. Specifically, we expect that balanced informed trading will tend to bias our estimates of ε upwards and µ downwards, which in turn will bias PIN downwards. However, since these trades offset one another, the market maker need not concern herself with balanced informed trades when setting prices. Furthermore, as long as the measurement errors in ε and µ are not correlated with our disclosure quality variables, then these errors should not result in misleading inferences. Nevertheless, our results should be interpreted with these caveats in mind Proxy for Disclosure Quality Disclosure quality reflects the overall informativeness of a firm s disclosures and depends on the amount, timeliness, and precision of the disclosed information. We use the Association of Investment Management and Research (AIMR) disclosure scores as our empirical proxy for a firm s disclosure quality. 8 According to the AIMR, the scores are intended to evaluate a firm s effectiveness in communicating with investors and the extent to which a firm s aggregate disclosures ensure that investors have the information necessary to make informed judgments. Based on a review of the specific reasons cited for large disclosure score increases, Healy, Hutton and Palepu [1999] find that firms exhibit a great deal of discretion when determining the 8 Prior studies that have used the AIMR disclosure scores include Botosan and Plumlee [2002], Gelb and Zarowin [2002], Healy, Hutton and Palepu [1999], Lang and Lundholm [1993], Lang and Lundholm [1996], Lundholm and Myers [2002], Sengupta [1998], and Welker [1995]. 12

14 amount of information to disclose, the level of detail to provide, and the timeliness with which to convey information, for both mandatory reports and purely voluntary disclosures. Each year during our sample period, the AIMR forms industry-based committees composed of leading analysts to undertake a comprehensive evaluation of disclosure quality for a subset of firms in a select number of industries. To allow efficient and convenient comparisons among industries, ( AIMR Report, p. 1) the committees use a common checklist to guide their evaluations, although they can modify or augment this checklist as they see fit. The industries selected for evaluation, the firms within an industry, the analyst committee composition, and the checklists change from year to year. In most cases, the end result of the evaluation process is a numerical score representing the overall quality of the firm s disclosures throughout the year. We refer to this result as the Total score. Additionally, most committees also report separate subscores that reflect the three major categories of disclosure: (1) the annual report and other required published information (Annual), (2) the quarterly report and nonrequired published information (Quarterly), and (3) investor relations and related activities (IR). 9 Consistent with most prior research, we convert each score into a percentile score based on the maximum possible score for each disclosure category. While the scores for a single industry-year are directly comparable, the scores across industries in a given year or across years for a given industry may not be comparable since each analyst committee might be using a different rating scales and criteria. In order to avoid industry- confounding effects, we make the conservative assumption that all industry-year differences are induced by different evaluation criteria. Accordingly, we subtract the industry-year mean percentile disclosure score from the unadjusted percentile score, and use this value in our empirical tests. To make the analysis 9 Detailed discussions of the AIMR rating process and the disclosure scores can be found in Lang and Lundholm [1993] and Healy, Hutton and Palepu [1999]. 13

15 consistent, we also industry-adjust all of the variables in our empirical analyses. While some studies have converted the industry-adjusted scores into ranks (Lang and Lundholm [1993], Gelb and Zarowin [2002]), we choose not to do so because converting cardinal scores into ranks results in a substantial decrease in power. 10 Additionally, converting the industry-adjusted scores into ranks can artificially increase or decrease the variation in the scores, which can bias our results unpredictably. 4. Methodology We are interested in analyzing the association between disclosure quality and information asymmetry. The potentially endogenous relation between these two attributes makes this task more difficult. In addition to disclosure quality s effect on information asymmetry, a manager s choice of disclosure quality is likely to be influenced by the concurrent level of information asymmetry. This potential endogeneity will cause the association between disclosure quality and asymmetry to be less negative since firms with higher asymmetry are more likely to choose higher disclosure quality because the expected benefits are higher. Failure to incorporate this endogeneity into our research design could result in misleading inferences. 11 Accordingly, we estimate the association between disclosure quality and information asymmetry using a threestage, instrumental variables procedure (Maddala [1983]), whereby the AIMR disclosure quality scores, Score, are modeled as a function of PIN and one set of exogenous variables and PIN is modeled as a function of Score and another set of exogenous variables. The disclosure quality and information asymmetry models are as follows, where firm and year subscripts are understood: 10 The industry committee reports clearly indicate that the disclosure scores are meant to be cardinal representations. 11 While most studies examining the effects of disclosure quality have not addressed this potential endogeneity, notable exceptions include Leuz and Verrecchia [2000], Marquardt and Wiedman [1998], and Welker [1995]. 14

16 Score = β + β PIN + β Size + β Return + β Surprise + β Correlation + β Capital β InstOwn + β Analysts + β Owners + β EarnVol + ε (3) PIN = γ + γ Score + γ Size + γ InstOwn + γ Analysts + γ Dispersion γ Leverage + γ EarnVol + ψ 6 7 (4) Depending on the specification, Score refers to either the Total, Annual, Quarterly, or IR score. As we discuss in Section 3.2, all variables are industry-adjusted by subtracting the industry-year mean. We discuss these models in more detail below. The analysis proceeds as follows: In the first stage, Score is modeled as a function of all the exogenous variables in both equations (3) and (4), and the endogenous variable PIN is excluded from the regression. Predicted values for Score, Score, are calculated using the estimated coefficients from this regression and the firm specific values for the exogenous variables. Likewise, the predicted values for PIN, PIN, are the fitted values from a regression of PIN on the all of the exogenous variables in equations (3) and (4), but excluding Score. In the second stage of the analysis, we estimate consistent estimates of the β coefficients in equation (3) by regressing Score on PIN and the exogenous variables in the disclosure model. Similarly, we obtain consistent estimates of the γ coefficients in equation (4) by regression PIN on the exogenous variables in the information asymmetry model. Score and 4.1. Disclosure Quality Model As discussed above, we expect that managers take the level of information asymmetry into account when they choose the quality of their disclosures. The previous literature identifies a number of other factors associated with firms disclosure quality choices. Lang and Lundholm [1993] examine the cross-sectional determinants of the AIMR disclosure quality scores. Based on their findings, we include the following variables in the disclosure quality model: (1) Size - 15

17 the natural log of the firm s market value of equity measured at the end of the firm s fiscal period; (2) Return - the market-adjusted stock return of the firm s equity measured over the fiscal period; (3) Surprise - the difference between the firm s actual per share earnings and the consensus analyst forecast (scaled by price) eight months prior to fiscal year end and winsorized at the 1% level; (4) Correlation - the correlation between annual stock returns and annual earnings measured over the ten years prior to the current fiscal period; and (5) Capital an indicator variable that equals one if the firm issues public debt or equity during the current and following two-year period, and zero otherwise. Based on the results in Lang and Lundholm [1993], we expect the coefficients on Size, Return, and Capital to be positive and the coefficients on Surprise and Correlation to be negative. We include four additional variables that we expect to affect disclosure quality. The first three are designed to capture differences in shareholders demands for disclosure quality. We anticipate that managers take the demand for disclosure quality by shareholders into account when they make their disclosure quality choices. Botosan and Harris [2000] and Bushee, Matsumoto and Miller [2003] find evidence indicating that firms respond to investor demands for increased disclosure. Accordingly, we expect the following three variables to have positive coefficients in the disclosure quality model: (6) InstOwn - the percentage of shares owned by institutional shareholders; (7) Analysts - the number of analysts covering the firm; and (8) Owners - the natural log of the number of shareholders. We include the standard deviation of earnings (scaled by assets) measured over the previous ten years, EarnVol, as an additional explanatory variable. Zhang [2001] shows that the equilibrium level of disclosure quality (measured by the precision of the disclosure) increases with the volatility of earnings. Firms increase their disclosure quality to (partially) counteract the higher level of private information- 16

18 based trading that is induced by high earnings volatility. Thus, we expect EarnVol to have a positive coefficient in the disclosure quality model Information Asymmetry Model As Section 2 discusses, we expect the level of disclosure quality to be negatively related to information asymmetry. In addition, we expect several other factors associated with the firm s information environment to be associated with the level of information asymmetry. The first variable we include is Size. Previous research indicates that stock prices incorporate information about large firms earlier than information about small firms. Based on the results in Atiase [1985], Bamber [1987], and Diamond and Verrecchia [1991], we expect a negative association between Size and PIN. Ayers and Freeman [2001] and Jiambalvo, Rajgopal, and Venkatachalam [2002] find evidence that current returns reflect future earnings to a greater extent when institutional ownership (InstOwn) is higher. These findings suggest that sophisticated investors are more actively trading on private information relating to future earnings, and current prices thus reflect future earnings information to a greater extent. If the trading activities of sophisticated investors are more frequently based on private information, then we expect PIN to be positively associated with InstOwn. While Ayers and Freeman [2001] find that analyst following plays a similar role to institutional ownership, Jiambalvo, Rajgopal and Venkatachalam [2002] find that the analyst following is negatively associated with the extent that prices lead future earnings. The analysis in Easley, O Hare and Paperman [1998] suggests that the role of analysts is more complex since the number of analysts cannot simply be used as a proxy for informed trade. They find that analyst following is positively associated with both the amount of informed and uninformed trading; with the net effect that information asymmetry is negatively associated with analyst 17

19 following. Therefore, we expect a negative coefficient on Analysts. The next variable included in the information asymmetry model is the dispersion of analyst forecasts, Dispersion, which is measured as ln((standard deviation of forecast earnings per share in the 4th month of fiscal period/stock price) ). We expect that when more uncertainty exists regarding expected earnings, more potential private information can be discovered and traded upon. In this case, private information search incentives are increasing with the amount of earnings uncertainty. Therefore, we expect a positive association between Dispersion and PIN. The amount of private information search activities will be positively associated with the expected benefits of obtaining the information. Boot and Thakor [1993] demonstrate that the incentives for private information acquisition are increasing with a firm s debt-to-assets ratio (Leverage). This result occurs because, for a given amount of private information about the value of a firm s assets, the expected profits from trading on that information in the equity market increase with the firm s leverage, ceteris paribus. Additionally, Zhang [2001] demonstrates that the endogenously-determined level of private information production increases with the volatility of earnings. When earnings volatility is higher, greater benefits arise from obtaining private information about future earnings. Assuming that the amount of informationbased trade increases with the amount of private information search activities, we expect Leverage and EarnVol to be positively associated with PIN. 5. Sample Description Our sample is based on firms that were evaluated in the 1986 through 1996 editions of the Annual Review of Corporate Reporting Practices by the AIMR. Our final sample consists of 2,432 firm-year observations representing 444 individual firms that have Total scores alone and 18

20 1,951 firm-year observations representing 350 individual firms that also have Annual, Quarterly, and IR scores. Our sample comprises 41 industries. 12 For each firm-year observation, we collect trade data from the ISSM Transactions File database and the Trades and Quotes (TAQ) database over the twelve-month period beginning eight months prior to the firm s fiscal year end. This time period likely corresponds to the AIMR evaluation period. 13 We match fiscal years to report years as follows: disclosure scores in the 1996 AIMR report correspond to fiscal-year ends that fall between 4/1/95 and 3/31/96. The other reports are matched similarly. For each firm, we gather data on every trade during the twelve-month sample period and classify it as either a buyer- or seller-initiated trade using the standard Lee-Ready algorithm (Lee and Ready [1991]). The algorithm classifies a trade that takes place above (below) the midpoint of the current quoted spread as a buy (sell). For trades taking place at the midpoint, we use a tick test based on the most recent transaction price to classify the trade. 14 When a sample firm has two or more classes of common stock, we only use the class of shares that has the highest trading volume during the sample period. We require firms to have at least one hundred days on which shares were actually traded in order to reduce estimation errors. Given the number of daily buys and sells for each trading day during our sample period, we use equation (1) to compute the maximum likelihood estimates for the PIN parameters (θ = {α, δ, ε b, ε s, µ}) for each firm-year in our sample. Given θ, we calculate the PIN for each firm using equation (2). In our sample, the number of trading days ranges from 107 to 254, with a mean (median) of 249 (252). 12 For approximately five industry-years, the committees reported only ranks - not Scores. We therefore estimate the value of Score based on the reported ranks and the range of scores in other years. 13 This information is based on conversations that Mark Soczek had with Patricia McQueen, Vice President of Advocacy Programs, AIMR, who headed the evaluation program in Following standard practice, we use a five-second lag on reported quote times to adjust for differences in reporting times between quotes and trades. Additionally, large trades are often broken down and matched against multiple investors. Reporting conventions will often classify such a transaction as multiple trades. Following Hasbrouck [1988], we classify all trades occurring within five seconds of each other as a single trade. 19

21 Data for the control variables come from a variety of sources. Accounting data is obtained from COMPUSTAT and market prices and return data come from CRSP. Institutional ownership data is derived from the CDA/Spectrum 13F Institutional Holdings database, and SDC Platinum is the source of our data on capital issuance. Analyst forecast data is obtained from IBES. Table 1 provides descriptive statistics for our final sample prior to industry adjusting. The mean (median) PIN is 17.6 (17.4), which indicates roughly a 17% - 18% chance that the opening trade on any given day is based on private information. The mean and median values of α demonstrate that private information events occur on almost half of trading days. The average value of ln(µ/ε) corresponds to a value of µ/ε of 0.45, where ε = ε b + ε s. This value indicates that informed trades are less than half the level of uninformed trades and represent about 30% of total trades on information-event days. Thus, our sample is characterized by firms with frequent information events and high levels of uninformed trading. The AIMR scores presented in Table 1 represent the reported score as a percentage of the maximum possible score. The mean Total score is 73.1, and the average subscore ranges from 72.3 to Considerable variation emerges within each disclosure category as the standard deviation ranges from 13.0 for the Annual score to 16.4 for the IR score. Table 1 also indicates that the firms rated by the AIMR tend to be large firms with large analyst followings (median = 19) and institutions generally holding over half of all shares outstanding. Ownership in these firms also tends to be widespread, with the median firm having over 18,000 owners. Table 2 presents the Spearman correlations for our sample. (All figures in Tables 2 to 4 are based on the industry-adjusted figures.) While ln(µ/ε) has a strong positive correlation with PIN (0.84) as expected, no significant correlation exists between PIN and α (0.00). This 20

22 unexpected result likely arises from the strong negative correlation between ln(µ/ε) and α (- 0.48). Untabulated results show that after controlling for the correlation between ln(µ/ε) and PIN, the partial correlation between PIN and α is significantly positive. As expected, the correlations between PIN and the disclosure score variables are all significantly negative, although somewhat moderate in magnitude (ranging from to -0.14). We find that the correlations between the four disclosure scores and Size, Return, Correlation, Capital, InstOwn, Analysts, and Owners all have the expected signs. While the correlations between the Scores and Surprise and EarnVol have unexpected signs, the magnitudes of the correlations are generally small, with the eight correlations ranging between 0.06 and 0.14 in absolute magnitude. The correlations between PIN and the exogenous variables in the information asymmetry model have the expected signs, with the exceptions of InstOwn (- 0.10) and Leverage (-0.04). 6. Analysis and Results In this section, we report the results of analyses that examine the relation between the quality of a firm s disclosures and the level of information asymmetry. In addition to examining the relation between information asymmetry and the Total disclosure quality score, we examine the relation separately for each of its three components: Annual, Quarterly, and IR. We extend our analysis by examining the relation between the disclosure scores and the underlying parameters on which PIN is based: α and ln(µ/ε). 6.1 Association between Disclosure Quality and Information Asymmetry Our primary analysis involves estimating the disclosure quality model (equation (3)) and information asymmetry model (equation (4)) simultaneously. Untabulated Hausman [1978] tests 21

23 for endogeneity reject the null hypothesis of no simultaneity for all the models on which we report. We present the results of this analysis in Table 3. The upper half of Column 1 presents the results for the information asymmetry model where PIN is the dependent variable and Total is the measure of disclosure quality. The Total coefficient is significantly negative (t = -2.8). This result supports our primary hypothesis that a firm s overall disclosure quality is negatively associated with the level of information asymmetry among investors. The magnitude of the coefficient, -0.12, indicates that an increase of 13% (the standard deviation of Total) in the total disclosure score is associated with a 1.6% decline in PIN. Based on the results in Easley, Hvidkjaer and O Hara (2002), a 1.6% reduction in PIN is associated with a reduction in the cost of capital of 41 basis points. We view this magnitude as a moderate and economically plausible effect of disclosure quality on the cost of capital. Since the median firm in our sample has an equity capitalization of $2.2 billion, this reduction in cost of capital translates into an annual savings of $9 million. Columns 2 4 present the results where Annual, Quarterly, and IR are the respective measures of disclosure quality. Each of the three coefficients is significantly negative, as expected (t-statistics vary from -2.3 to -4.0). The IR coefficient is the most significantly negative and is also the most robust to the alternate specifications discussed in Section 6.2. The IR results are consistent with the claims in Mahoney [1991] and Marcus and Wallace [1991] that reduced uncertainty and lower information asymmetry among stock participants are among the benefits of high quality IR activities. However, the results also suggest that Annual and Quarterly also contribute to the overall effect of disclosure quality. The magnitudes of their coefficients are in fact larger than those on IR, and a direct comparison of the magnitudes is valid since the standard deviations of the industry-adjusted Scores are similar. 22

24 One reason that IR has a higher level of significance may be that many regulatory requirements (including those of the SEC and FASB) govern the form and content of quarterly and annual reports, setting a lower bound on the quality of these disclosures. In contrast, investor relations activity is less regulated, and firms have a greater ability to distinguish themselves in this regard. Consequently, the variation in the underlying quality of the annual and quarterly reports for firms complying with the regulations is likely less than the underlying variation in the quality of the firm s IR activities. For this reason, we expect the extent of measurement errors in the IR scores to be relatively smaller than the errors in the Annual and Quarterly scores. Accordingly, the power of tests based on IR will be higher than for tests based on Annual and Quarterly. The coefficients on Size and Analysts in the information asymmetry model are significant and have the predicted signs for each of the four Score measures. However, the results for the other control variables do not coincide with our expectations. While InstOwn is positive and moderately significant in the Total regression (t = 1.8) as predicted, it is insignificantly negative in each of the three subscore regressions. Likewise contrary to our expectations, the coefficients on Dispersion and Leverage are generally significantly negative in each of the four regressions, and EarnVol is negative (though insignificantly so) in each model. In untabulated tests, we use several different combinations and specifications of these variables to investigate why the contrary results may have arisen. In all specifications, the contradictions remain offering an opportunity for future research. In the lower half of Table 3, we present the results from estimating the disclosure quality model. We find consistent evidence that the quality of disclosures chosen by managers is increasing in the level of information asymmetry. The coefficient on PIN is significantly 23

25 positive for all measures of disclosure quality. The positive coefficients on PIN in the Score regressions contrast strongly with the negative coefficients on Score in the PIN regressions and are consistent with the endogeneity of the two variables, as is strongly confirmed by untabulated Hausman tests. The coefficients on the control variables in the Score equations are generally consistent with our hypotheses. The coefficients on Size, Capital, InstOwn, Analysts, and Owners are significantly positive in general, while the Correlation coefficients are all negative and significantly so in the Total and IR specifications. As in Lang and Lundholm (1993), the coefficients on Surprise are not significantly negative and for Total are significantly positive. 15 Additionally, EarnVol is not significantly positive in any specification but is in fact significantly negative in the Annual equation. Having found evidence that disclosure quality is negatively associated with the level of information asymmetry, we now investigate the sources of this association. As we discuss in Section 2, we expect disclosure quality to be related to information asymmetry through two channels: a reduction in the frequency with which investors become aware of private information and a reduction in the relative trading intensity of informed investors. Therefore, we examine the relation between disclosure quality and the frequency of private information events, α, and the relative amount of uninformed trading, ln(µ/ε). The two component regressions are similar to equation (4), except that the dependent variable is based on the underlying PIN parameters rather than PIN itself. Panel A of Table 4 presents the results for the α-based regressions. The results show that the IR coefficient is significantly positive, while the Total, Annual, and Quarterly coefficients are 15 We also measure Surprise by reference to a random walk model for earnings. Surprise loses its significance in the Total equation, but otherwise the results are not materially different. 24